You may also enjoy, Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, SymptomAI: Towards a conversational AI agent for everyday symptom assessment, a paper on A Unified Definition of Hallucination: It’s The World Model, Stupid!, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Kimi K3 Tech Blog: Open Frontier Intelligence

Moonshot AI launched Kimi K3, the world’s first open 3T-class model at 2.8 trillion parameters, with native multimodality and a 1-million-token context window.

Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber

Google launched Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, with 3.6 Flash reducing output token usage by 17% compared to 3.5 Flash while improving coding and knowledge work performance.

Introducing OpenAI Presence

OpenAI launched Presence, an enterprise agent deployment product pairing model reasoning with policies and escalation rules, currently resolving 75% of inbound issues without human assistance on OpenAI’s own support line.

AMD and Anthropic Announce Strategic Partnership to Deploy Up to 2 Gigawatts of AMD Instinct MI450 Series GPUs

AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of MI450 Series GPUs in AMD Helios rack-scale solutions starting in the first half of 2027, paired with up to $5 billion in AMD equity investment in Anthropic.

MLOps/LLMOps/AgentOps

Beyond the Single Trace: How We Built Agent Diagnostics for Opik

Comet built Diagnostics for Opik, an automated debugging agent that queries trace and span data in ClickHouse rather than reading traces one at a time, to surface silent agent failures like retry loops and over-deliberation at scale.

Learning

Loop Engineering in Self-Correcting Code Migration using Google ADK 2.0

A technical guide detailing how Google ADK 2.0’s declarative workflow architecture eliminates quadratic token growth in self-correcting code migration loops through state-aware history pruning and decoupled validation.

AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors

Epoch AI tested three detectors and found style-imitated AI text evaded detection far more often than plain AI text, with scientific writing failing to be detected ~26% of the time versus near-zero false negative rates on basic prompts.

SymptomAI: Towards a conversational AI agent for everyday symptom assessment

Google Research’s national-scale study found their Gemini Flash 2.0-based SymptomAI agent’s differential diagnoses were preferred over clinicians’ own assessments in 53.3% of cases.

Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

A Hugging Face blog post detailing the native integration of Nunchaku’s SVDQuant W4A4 kernels into Diffusers, enabling 4-bit diffusion inference that cuts peak VRAM by up to 50% while boosting speed by 30-80%.

What we learned about agent teamwork

An engineering blog about running 10 autonomous AI agent film crews on the open-source Scion orchestration testbed, surfacing patterns for multi-agent coordination outside coding domains.

Evals are the new PRD, Expedia’s AI chief tells VB Transform 2026

Expedia’s chief AI and data officer, Xavi Amatriain, states that evals now function as the new PRD, with product intent encoded directly into evaluation suites before coding begins.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

block/buzz

A workspace where humans and agents build together, on a relay you own.

Papers & Publications

Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering

Abstract:

Large language models (LLMs) can be controlled at inference time through prompts (in-context learning) and internal activations (activation steering). Different accounts have been proposed to explain these methods, yet their common goal of controlling model behavior raises the question of whether these seemingly disparate methodologies can be seen as specific instances of a broader framework. Motivated by this, we develop a unifying, predictive account of LLM control from a Bayesian perspective. Specifically, we posit that both context- and activation-based interventions impact model behavior by altering its belief in latent concepts: steering operates by changing concept priors, while in-context learning leads to an accumulation of evidence. This results in a closed-form Bayesian model that is highly predictive of LLM behavior across context- and activation-based interventions in a set of domains inspired by prior work on many-shot in-context learning. This model helps us explain prior empirical phenomena - e.g., sigmoidal learning curves as in-context evidence accumulates - while predicting novel ones - e.g., additivity of both interventions in log-belief space, which results in distinct phases such that sudden and dramatic behavioral shifts can be induced by slightly changing intervention controls. Taken together, this work offers a unified account of prompt-based and activation-based control of LLM behavior, and a methodology for empirically predicting the effects of these interventions.

A Unified Definition of Hallucination: It’s The World Model, Stupid!

Abstract:

Despite numerous attempts at mitigation since the inception of language models, hallucinations remain a persistent problem even in today’s frontier LLMs. Why is this? We review existing definitions of hallucination and fold them into a single, unified definition wherein prior definitions are subsumed. We argue that hallucination can be unified by defining it as simply inaccurate (internal) world modeling, in a form where it is observable to the user. For example, stating a fact which contradicts a knowledge base OR producing a summary which contradicts the source. By varying the reference world model and conflict policy, our framework unifies prior definitions. We argue that this unified view is useful because it forces evaluations to clarify their assumed reference “world”, distinguishes true hallucinations from planning or reward errors, and provides a common language for comparison across benchmarks and discussion of mitigation strategies. Building on this definition, we also connect our framework to HalluWorld, a complementary benchmark that instantiates fully specified reference world models for stress-testing model hallucinations.

You may also enjoy, GPT-Red: Unlocking Self-Improvement for Robustness, Model Routing Is Simple. Until It Isn’t., Video Generation Models are General-Purpose Vision Learners, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Inkling: Our open-weights model - Thinking Machines Lab

Thinking Machines Lab releases Inkling, an open-weights 975B-parameter multimodal MoE model with controllable reasoning effort, fine-tunable on Tinker.

Canva launches Code 2.0, offering AI website building to every user

Canva launches Code 2.0 to all 265M monthly users, betting design polish—not code generation—is the real gap versus Lovable, Replit, and Bolt in vibe coding.

GPT-Red: Unlocking Self-Improvement for Robustness

OpenAI trains GPT-Red, a self-play automated red-teaming model, to attack production systems and adversarially train GPT-5.6, cutting direct prompt injection failures 6x.

MLOps/LLMOps/AgentOps

How We Optimized Opik’s MCP Server for Cost & Performance

A look into how Comet rebuilt Opik’s MCP server from 30 tools down to four, using self-correcting schemas and adaptive response compression to cut token waste, improve tool selection, and keep agent context lean.

What building Shippy taught us about building agents

Ai2 details Shippy’s maritime agent architecture, showing that reliability came from a deterministic CLI layer, isolated per-session sandboxes, and rubric-based evals scoring the whole agent rather than the model alone.

Learning

Model Routing Is Simple. Until It Isn’t.

IBM Research argues LLM routing is a systems-optimization problem, not classification, after finding Claude Sonnet 4.6 cost half of GPT-4.1 per task due to caching effects despite higher sticker pricing.

Text match filters for agents

Pinecone launches lexical text match filters that scope semantic search to unstated query context, without pre-labeling metadata across the dataset.

AI energy use: its impact on prices, climate, and more

Epoch AI’s guide finds individual chatbot queries cost less energy than a microwave running 10 seconds, while global AI compute demand—tens of gigawatts—still doubles roughly every year.

Towards demystifying the creativity of diffusion models

Google Research explains diffusion model creativity as a mathematical byproduct of neural network training, where regularization smooths the score function and drives interpolation between training points rather than memorization.

Agentic Misalignment in Summer 2026

Anthropic documents four new agentic misalignment failure modes in frontier models — covert sabotage, fraud assistance, motivated mislabeling, and coaching human whistleblowers — through controlled multi-model simulations.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

microsoft/ResearchStudio

ResearchStudio: Our AI co-author, from research problem to final publication.

Papers & Publications

Video Generation Models are General-Purpose Vision Learners

Abstract:

Driven by next-token prediction, NLP shifted from task-specific models into powerful generalist foundation models. What, then, is the equivalent catalyst needed to achieve a general-purpose model in computer vision? In this paper, we contend that large-scale text-to-video generation serves as a strong pre-training paradigm for computer vision, providing the necessary spatiotemporal priors, vision-language alignment, and scalability required for general visual intelligence. We introduce GenCeption, which leverages a pre-trained video generative diffusion backbone to define a feed-forward perception model, capable of performing various vision tasks steered by text instructions. Empirical results demonstrate that GenCeption achieves state-of-the-art performance across a diverse suite of tasks, including depth, surface normal, and camera pose estimation, expression-referring segmentation, and 3D keypoint prediction, often matching or surpassing specialized models (e.g. DepthAnything3, SAM3, D4RT, VGGT-Omega, Sapiens, David, Genmo, and Lotus-2). Furthermore, the video generative pretrained backbone outperforms alternative pretraining paradigms (e.g., V-JEPA, and Video MAE) under comparable settings. Importantly, GenCeption exhibits preliminary data and model scaling properties along with exceptional data efficiency, where it achieves comparable performance with leading models like D4RT and VGGT-Omega with 7 to 500 less training data. Finally, GenCeption also exhibits intriguing emergent behaviors: a model trained exclusively on synthetic human videos generalizes to real-world footage and out-of-distribution object categories (e.g., animals and robots). These findings suggest that video generation is not merely a synthesis tool, but a foundational path toward generalist vision intelligence for the physical world.

Metacognition in LLMs: Foundations, Progress, and Opportunities

Abstract:

Metacognition is a foundational component of intelligence critical to effective learning, problem solving, decision-making, communication, and more. In recent years, it has become increasingly recognized as a cornerstone of capable, transparent AI systems. Yet while LLMs have made significant progress across diverse real-world tasks, it is not yet clear when, how, or to what extent they can exhibit or be endowed with effective metacognitive abilities, nor how such abilities can be adapted to advance the fundamental capabilities, reliability, and intelligence of AI systems. This paper bridges this gap by presenting the first comprehensive overview of the current state of knowledge on metacognition for LLMs. We analyze and taxonomize the landscape of this emerging field and summarize recent technical advancements, including methods and benchmarks to measure and evaluate LLMs’ metacognitive abilities, techniques to elicit, improve, and apply metacognition in LLMs, and findings and implications of ongoing research. We also discuss applications, open questions and challenges, and promising directions for future work. Our aim is to provide a detailed and up-to-date review of this topic and stimulate meaningful research and discussion.

You may also enjoy, Introducing GPT-Live, Intelligence is Free, Now What? Data Systems for, of, and by Agents, a paper on Can LLMs Reason Structurally? Benchmarking via the lens of Data Structures, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Introducing Grok 4.5

xAI launches Grok 4.5, its fastest and most token-efficient model yet (80 TPS, 4.2× fewer output tokens than Opus 4.8 on SWE-Bench Pro), priced at $2/$6 per million tokens.

Introducing GPT-Live

OpenAI launches GPT-Live, a full-duplex voice model that listens and speaks simultaneously and delegates reasoning/search tasks to GPT-5.5 in the background, now powering ChatGPT Voice.

Introducing Mods: Enabling Agents to Self-Improve through Harness-Level Adaptation

Letta launches Mods for Letta Code, letting agents self-modify their own harness code rather than only learning through context, with a built-in learn/generate-env workflow.

Leanstral 1.5: Proof Abundance for All

Mistral open-sources Leanstral 1.5, which saturates miniF2F, solves 587/672 PutnamBench problems, and sets new SOTA on FATE-H/X while surfacing 5 previously unknown bugs across 57 tested repositories.

MLOps/LLMOps/AgentOps

Intelligence is Free, Now What? Data Systems for, of, and by Agents

UC Berkeley researchers argue near-zero inference costs will force data systems to be redesigned around agentic workloads — serving agent query swarms, sustaining multi-agent memory/coordination, and letting agents synthesize disposable custom systems.

Data for Agents

NVIDIA argues agentic AI’s bottleneck is open, inspectable synthetic data rather than model weights, releasing Nemotron’s over 10 trillion pre-training tokens and millions of post-training samples alongside a Prompt Atlas and locally-grounded persona datasets covering more than 2.4B people across ten countries.

Learning

Engineering Insights: How Internal Optimizations Led to Comet Cost Intelligence

A look into how Comet’s engineering team built Cost Intelligence to expose where AI coding agents waste tokens and budget, turning opaque AI spend into actionable engineering metrics.

Harness Engineering for Self-Improvement

Lilian Weng’s post argues the “harness” wrapping a model is becoming as important as the model itself — and is now being optimized by the agents it contains.

The Making of Claude Code

Anthropic publishes a history of Claude Code’s origins, tracing its evolution from a 2022 internal coding assistant through “clide” to its February 2025 launch and shift to near-100% AI-written code by winter 2025.

Separating signal from noise in coding evaluations

OpenAI’s audit finds ~30% of SWE-Bench Pro tasks are broken, retracting its own earlier recommendation to adopt the benchmark as a replacement for the flawed SWE-bench Verified.

Modular Pretraining Enables Access Control

Anthropic and AE Studio introduce GRAM, a technique that isolates dual-use knowledge (virology, cybersecurity, nuclear physics) into switchable modules, letting one pretrained model approximate five separately data-filtered models at a fifth of the training compute.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

zilliztech/mfs

A context harness for AI agents: all your scattered context — code, memory, docs, databases, SaaS — in one searchable, browsable, file-like interface.

Papers & Publications

AlayaWorld: Long-Horizon and Playable Video World Generation

Abstract:

Game worlds have traditionally been built through labor-intensive production pipelines, making them costly to develop, difficult to customization, and expensive to modify after deployment. Recent advances in video world models offer a fundamentally different paradigm. Rather than explicitly authoring every component of a virtual environment, these models autoregressively synthesize future observations conditioned on the current world state and user interactions, enabling playable worlds to be generated online. Trained on both gameplay recordings and real-world videos, they can capture diverse visual appearances and physical dynamics, opening new opportunities for interactive applications beyond gaming, including embodied intelligence. In this paper, we present \textbf{AlayaWorld}, a full-stack open-source framework for building interactive generative worlds. AlayaWorld enables open-ended real-time interaction, allowing users to freely navigate and perform diverse actions such as combat, spell casting, and monster summoning. The framework unifies the complete development-from data preparation model architecture, model training, inference acceleration, and deployment-within a modular and extensible architecture. Alongside the framework, we release reproducible pipelines, reference implementations, evaluation tools, and comprehensive documentation, establishing a practical foundation for future research and real-time applications of generative world models.

Can LLMs Reason Structurally? Benchmarking via the lens of Data Structures

Abstract:

Large language models (LLMs) are deployed on increasingly complex tasks that require multistep decision-making. Understanding their algorithmic reasoning abilities is therefore crucial. However, we lack a diagnostic benchmark for evaluating these capabilities. We propose to use data structures as a principled lens: as fundamental building blocks of algorithms, they naturally probe structural reasoning—the ability to understand and manipulate relationships such as order, hierarchy, and connectivity that underpin algorithmic reasoning. We introduce DSR-Bench (Data Structure Reasoning Benchmark), spanning 20 data structures, 35 operations, and 4,140 problem instances. DSR-Bench features hierarchical task organization, fully automated generation and evaluation, and fine-grained diagnostics. Evaluating 13 state-of-the-art LLMs reveals critical limitations: the top-performing model achieves only 0.46/1 on challenging instances. Three auxiliary probes targeting more realistic usages expose further weaknesses: models perform poorly on spatial data and context-rich scenarios, and they struggle to reason over their own code.

You may also enjoy, Redeploying Claude Fable 5, R&B-EnCoRe: Self-Improving Pretraining of Embodied Reasoning Vision-Language-Action Models, a paper on Unlimited OCR Works, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Previewing GPT-5.6 Sol: a next-generation model

OpenAI begins a limited, government-coordinated preview of GPT-5.6 Sol, Terra, and Luna, its strongest cybersecurity model yet, paired with a layered safeguard stack and 700,000 GPU hours of automated red-teaming.

Redeploying Claude Fable 5

Anthropic restores access to Claude Fable 5 and Mythos 5 after US export controls are lifted, rolling out with a strengthened cybersecurity classifier and a new cross-industry jailbreak severity framework.

Introducing Claude Sonnet 5

Anthropic launches Claude Sonnet 5, its most agentic Sonnet model yet, narrowing the performance gap to Opus 4.8 on coding and tool use.

Claude Science, an AI workbench for scientists

Anthropic launches Claude Science, an AI workbench that unifies research tools like PubMed, Jupyter, and cluster compute into one environment with over 60 scientific skills and auditable, reproducible outputs.

LFM2.5-230M: Built to Run Anywhere

Liquid AI releases LFM2.5-230M, its smallest open-weight model yet, delivering 213 tok/s on a Galaxy S25 Ultra and outperforming larger models on tool use and data extraction benchmarks.

Start building with Nano Banana 2 Lite and Gemini Omni Flash

Google launched Nano Banana 2 Lite, generating images in 4 seconds at $0.034 per 1K image, alongside Gemini Omni Flash for developer video generation and conversational editing at $0.10 per second.

MLOps/LLMOps/AgentOps

Opik + Oracle Agent Specification: Build Once, Run Anywhere

Comet integrates Opik with Oracle’s Open Agent Specification, for defining agents once and trace, evaluate, and swap them across frameworks like LangGraph, AutoGen, and WayFlow without rebuilding.

AI Evaluation Simplified: Automate Dataset & Metric Eval Workflows with Test Suites

A guide about Opik’s Test Suites feature, which replaces traditional dataset-and-metric AI evaluation workflows with plain-English assertions that return pass/fail results instead of raw scores.

Learning

Scaling Laws, Carefully

A comprehensive explainer tracing how LLM scaling laws evolved from Kaplan to Chinchilla to data-constrained regimes, covering why fitting these power-law relationships is more fragile in practice than it looks.

R&B-EnCoRe: Self-Improving Pretraining of Embodied Reasoning Vision-Language-Action Models

A research blog post introducing R&B-EnCoRe, a self-supervised pretraining cycle that lets vision-language-action models discover which reasoning steps actually improve action prediction, rather than following fixed templates.

Core dump epidemiology: fixing an 18-year-old bug

An engineering deep-dive about how OpenAI used population-level core dump analysis, rather than single-case debugging, to uncover two unrelated crash causes in its Rockset data infrastructure.

MirrorCode: What’s the largest software project AI can complete on its own?

Epoch AI, with METR, launches MirrorCode, a benchmark measuring whether AI can reimplement entire real-world programs end-to-end without source access, on which Claude Opus 4.7 leads at a 56% solve rate.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

radixark/miles

An enterprise-facing reinforcement learning framework for LLM and VLM post-training, forked from and co-evolving with slime.

Papers & Publications

Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents

Abstract:

Despite recent progress, LLM agents still struggle with reasoning over long interaction histories. While current memory-augmented agents rely on a static retrieve-then-reason paradigm, this rigid pipeline design prevents them from dynamically adapting memory access to intermediate evidence discovered during inference. To bridge this gap, we propose MRAgent, a framework that combines an associative memory graph with an active reconstruction mechanism. We represent memory as a Cue-Tag-Content graph, where associative tags serve as semantic bridges connecting fine-grained cues to memory contents. Operating on this structure, our active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. This ensures that memory retrieval is dynamically adapted to the reasoning context while avoiding combinatorial explosion caused by unconstrained expansion. Experiments on the LoCoMo benchmark and LongMemEval benchmark demonstrate significant improvements over strong baselines (up to 23%), while substantially reducing token and runtime cost, highlighting the effectiveness of active and associative reconstruction for long-horizon memory reasoning.

Unlimited OCR Works

Abstract:

Recently, end-to-end OCR models, exemplified by DeepSeek OCR, have once again thrust OCR into the spotlight. A widely held view is that employing a large language model (LLM) as the decoder allows the model to leverage the prior distribution of language, leading to improved OCR performance. However, the downside is equally evident: as the output sequence lengthens, the accumulated KV cache drives up memory consumption and progressively slows down generation. This stands in stark contrast to humans, who exhibit no such decline in efficiency during long-horizon copying tasks. In this technical report, we propose Unlimited OCR, a model designed to emulate human parsing working memory. Taking DeepSeek OCR as the baseline, we replace all attention layers in the decoder with our proposed Reference Sliding Window Attention (R-SWA), which reduces attention computation costs while maintaining a constant KV cache throughout the entire decoding process. By combining the high compression rate of DeepSeek OCR’s encoder with our constant KV cache design, Unlimited OCR can transcribe dozens of pages of documents in a single forward pass under a standard maximum length of 32K. More importantly, R-SWA is a general-purpose parsing attention mechanism - beyond OCR, it is equally applicable to tasks such as ASR, translation, etc.

You may also enjoy, CoT Monitoring: Where Does a Hot Safety Problem Come From?, A Guide to Inference Engineering, a paper on Are We Ready For An Agent-Native Memory System?, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Xiaomi’s HarnessX rewrites its own AI scaffolding mid-task — and smaller models gain the most

Xiaomi researchers introduce HarnessX, a framework that treats an AI agent’s harness as a composable, self-evolving object, letting it autonomously rewrite its own scaffolding mid-task rather than relying on manual updates.

Introducing Claude Tag \ Anthropic

Anthropic launches Claude Tag, a Slack-native @Claude teammate that learns channel context, works asynchronously, and now powers 65% of the product team’s code at Anthropic internally.

Introducing computer use in Gemini 3.5 Flash

Google makes computer use a native built-in tool in Gemini 3.5 Flash, consolidating its previously standalone screen-control model into the main Flash model with new enterprise prompt-injection safeguards.

Qwen-AgentWorld: Language World Models for General Agents

Qwen releases Qwen-AgentWorld, a native language world model trained to simulate seven agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within one model, outperforming GPT-5.4 and Claude Opus 4.8 on the new AgentWorldBench evaluation.

Mistral OCR 4 : SOTA OCR for Document Intelligence

Mistral releases OCR 4, a self-hostable document-intelligence model adding bounding boxes, block classification, and confidence scores, beating competitors in human preference testing and topping OlmOCRBench.

MLOps/LLMOps/AgentOps

Advanced Claude Code Cost Tracking: How to Save 30% on Token Spend

Comet introduces Cost Intelligence, a Claude Code and Codex cost tracker that attributes AI spend across developers, projects, tools, and workflows while surfacing configuration changes that reduce token usage.

Version-Controlling Your Agents: Deployment, Rollback, and Safe Promotion Patterns

A practical guide arguing that AI agents need the same versioning, staged promotion, and rollback discipline as traditional software, treating their configs as immutable, deployable “Agent-as-Code” artifacts.

Loop Engineering Explained

A guide to loop engineering, an agent design pattern that replaces prompt-centric workflows with iterative execution loops that manage tool use, state, planning, and reflection.

A Guide to AI Inference Engineering

A guide explaining LLM inference engineering through the prefill/decode split, breaking down six core optimization techniques and when self-hosting beats off-the-shelf APIs.

Learning

How Evaluation-Driven Development (EDD) Works

Learn how retrieval, memory, and context management have emerged as critical infrastructure for AI agents, often having a greater impact on system performance than the underlying model itself.

CoT Monitoring: Where Does a Hot Safety Problem Come From?

A reflective essay tracing the intellectual lineage of chain-of-thought (CoT) monitoring, arguing it emerged from the convergence of ML monitoring practices and CoT-as-explainability research rather than as a standalone idea.

Toward More Controllable AI Video Editing: An Early Research Exploration at Netflix

A blog post about Netflix Research introducing Vera, a layered video diffusion model, and VOID, a physics-aware inpainting model, both aimed at giving artists more precise, controllable AI-assisted video editing.

Healthcare Benchmarks Are Only as Good as Their Assumptions

A research blog post arguing that healthcare LLM benchmarks fail to predict real-world performance because they embed unstated assumptions about task structure and outcome measurement that break down at deployment.

Patterns for Building Cybersecurity Evals

A guide breaking down cybersecurity evals into four shared primitives (sandboxed target, difficulty-tuning inputs, tools, grader), then walking through seven benchmarks from CTF-style exploitation to 50-host network compromise.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

XiaomiMiMo/MiMo-Code

MiMoCode is a terminal-native AI coding assistant.

Papers & Publications

Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs

Abstract:

While reasoning in LLMs plays a natural role in math, code generation, and multi-hop factual questions, its effect on simple, single-hop factual questions remains unclear. Such questions do not require step-by-step logical decomposition, making the utility of reasoning highly counterintuitive. Nevertheless, we find that enabling reasoning substantially expands the capability boundary of the model’s parametric knowledge recall, unlocking correct answers that are otherwise effectively unreachable. Why does reasoning aid parametric knowledge recall when there are no complex reasoning steps to be done? To answer this, we design a series of hypothesis-driven controlled experiments, and identify two key driving mechanisms: (1) a computational buffer effect, where the model uses the generated reasoning tokens to perform latent computation independent of their semantic content; and (2) factual priming, where generating topically related facts acts as a semantic bridge that facilitates correct answer retrieval. Importantly, this latter generative self-retrieval mechanism carries inherent risks: we demonstrate that hallucinating intermediate facts during reasoning increases the likelihood of hallucinations in the final answer. Finally, we show that our insights can be harnessed to directly improve model accuracy by prioritizing reasoning trajectories that contain hallucination-free factual statements.

Are We Ready For An Agent-Native Memory System?

Abstract:

Memory for large language model (LLM) agents has rapidly evolved from simple retrieval-augmented mechanisms into a data management system that supports persistent information storage, retrieval, update, consolidation, and dynamic lifecycle governance throughout agent execution. Despite this evolution, existing evaluations still benchmark agent memory mainly through end-to-end task success metrics (e.g., F1, BLEU), while treating the underlying system as a monolithic black box. As a result, critical system-level concerns, including operational costs, architectural trade-offs across memory modules, and robustness under dynamic knowledge updates, remain insufficiently explored. In this paper, we present a systematic experimental study of agent memory from a data management perspective. We propose an analytical framework that decomposes agent memory into four core modules: memory representation and storage, extraction, retrieval and routing, and maintenance. Under this framework, we evaluate 12 representative memory systems and two reference baselines across five benchmark workloads spanning 11 datasets. Our extensive end-to-end evaluation shows that no single architecture dominates across all scenarios; instead, effectiveness depends heavily on how well the memory structure aligns with the workload bottleneck. Furthermore, through fine-grained ablation studies, we quantify their individual effects on representation fidelity, retrieval precision, update correctness, and long-horizon stability. Finally, we reveal cost-performance trade-offs under realistic workloads, showing localized maintenance is more cost-efficient than global reorganization. Based on these findings, we identify promising directions towards building truly agent-native memory systems.

You may also enjoy Kimi K2.7 Code, M*: A Modular, Extensible, Serving System for Multimodal Models, a paper on FastContext: Training Efficient Repository Explorer for Coding Agents, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Kimi K2.7 Code: Open-Source Agentic Coding Model

Moonshot AI launches Kimi K2.7 Code, an open-source 1T-parameter MoE coding model gaining up to 31.5% on benchmarks while cutting thinking-token usage ~30% versus K2.6.

MolmoMotion: Language-guided 3D motion forecasting

Ai2 releases MolmoMotion, a language-guided model that forecasts object 3D point trajectories from video, beating prior methods on motion forecasting, robot planning, and video generation.

Stanford’s DeLM cuts multi-agent task costs 50% — without a central orchestrator

Stanford’s DeLM replaces central orchestrators with shared verified context, beating the strongest baseline by 10.5% on SWE-bench Verified at roughly half the cost.

GLM-5.2: Built for Long-Horizon Tasks

Z.ai launches GLM-5.2, an open-source 753B-parameter MoE model with a stable 1M-token context, beating GLM-5.1 by wide margins on Terminal-Bench 2.1 and SWE-bench Pro while trailing Claude Opus 4.8 by just 1% on FrontierSWE.

MLOps/LLMOps/AgentOps

Hidden Technical Debt of AI Systems: Agent Evaluation Infrastructure

A detailed guide arguing agent evaluation requires a control-plane/data-plane system spanning traces, state deltas, checkpoints, and replay — not just a single benchmark score.

M*: A Modular, Extensible, Serving System for Multimodal Models

Stanford’s M* replaces vLLM/SGLang’s single autoregressive loop with a generic “Walk Graph,” beating specialized serving systems by up to 2.7x on speech, 2.6x on image editing, and 12.5x on world-model rollouts.

Learning

Understanding Your Claude Code Spend: What’s Actually Driving the Cost

Long and inefficient context can quietly drive up costs and reduce performance in Claude Code. Learn five techniques for auditing context usage and eliminating unnecessary token overhead.

How Evaluation-Driven Development (EDD) Works

Learn how retrieval, memory, and context management have emerged as critical infrastructure for AI agents, often having a greater impact on system performance than the underlying model itself.

Pre-Training Isn’t Bitter Enough

CMU proposes V-pretraining, which uses a small labeled feedback set to train a task designer that shapes self-supervised targets, lifting Qwen2.5-0.5B’s GSM8K Pass@1 from 22.20 to 29.60 without directly supervising the learner.

First Steps Toward Automated AI Research

Recursive’s automated AI research system beats human-optimized SOTA on three benchmarks: 0.0263 lower BPB on fixed-budget LM training, 2.2s faster on NanoGPT Speedrun, and an 18% gap reduction on GPU kernel optimization.

Three Ways Codex Can Use a Computer

A practical guide on routing OpenAI Codex tasks across Computer Use, Chrome, and the in-app browser based on whether the job needs native apps, signed-in sites, or public-page review.

Portable vLLM Model Inference Kernels in Helion

Red Hat and Meta integrate Helion’s PyTorch-native kernel DSL into vLLM for FP8 Qwen3 inference, beating existing CUDA/TorchInductor kernels on most ops while still trailing CUTLASS on GEMM for Blackwell GPUs.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

openai/codex-plugin-cc

Use Codex from Claude Code to review code or delegate tasks.

Papers & Publications

Data Journalist Agent: Transforming Data into Verifiable Multimodal Stories

Abstract:

Data tells stories that shape society; the data journalist’s job is to turn raw information into stories non-experts can trust. A high-quality news feature takes a newsroom team weeks: hunting for context, running statistics, choosing an angle, and designing visuals. Recent agents handle individual steps well: data-science agents close the analysis loop, while design agents synthesize beautiful websites. But can an agent serve as a data journalist end to end? We introduce Data Journalist Agent (Data2Story), a multi-agent framework that orchestrates specialized roles into a single virtual newsroom. Data2Story contributes two innovations. (i) Claims are evidence-grounded: an Inspector links every number, angle, and asset back to data, code, or an external reference. (ii) Articles are multimodally generative: rather than defaulting to plain text and static charts, Data2Story reasons about what readers will want to see, then deploys multimodal tools, such as interactive maps for geography and audio for music. We evaluate Data2Story on 18 articles, each paired with the originally published expert piece, along four axes: (a) human-agent angle coverage; (b) rubric evaluation with 53 participants across five dimensions; (c) computer-use agents as judges, a cost-saving proxy for how readers navigate interactive articles; and (d) verifiability, where a coding verifier re-executes statements against the data and checks claims against references. Data2Story produces competitive, evidence-traceable multimedia stories, with particular strength in transparency and auditability. Human articles retain an edge in editorial angle, creative design, and presentation. We position Data2Story as a collaborator for journalists, enabling more evidence-based, transparent, and verifiable reporting.

FastContext: Training Efficient Repository Explorer for Coding Agents

Abstract:

Large Language Model (LLM) coding agents have achieved strong results on software engineering tasks, yet repository exploration remains a major bottleneck: locating relevant code consumes substantial token budget and pollutes the agent’s context with irrelevant snippets. In most agents, the same model explores the repository and solves the task, leaving exploratory reads and searches in the solver’s history. We present FastContext, a dedicated exploration subagent that separates repository exploration from solving. Invoked on demand, FastContext issues parallel tool calls and returns concise file paths and line ranges as focused context. FastContext is powered by specialized exploration models spanning 4B--30B parameters. We bootstrap them from strong reference-model trajectories and refine them with task-grounded rewards for broad first-turn search, multi-turn evidence gathering, and precise citation generation. Across SWE-bench Multilingual, SWE-bench Pro, and SWE-QA, integrating FastContext into Mini-SWE-Agent improves end-to-end resolution rates up to 5.5% while reducing coding-agent token consumption up to 60%, with marginal overhead. These results show that repository exploration can be separated from solving and handled effectively by specialized models.

You may also enjoy NVIDIA Nemotron 3 Ultra, Controlling the capital after AGI, a paper on Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Claude Fable 5 and Claude Mythos 5 \ Anthropic

Anthropic launches Claude Fable 5, a general-access Mythos-class model at $10/$50 per M tokens, with classifier-based fallbacks to Opus 4.8 for cyber, bio, and distillation queries.

Introducing Gemma 4 12B: a unified, encoder-free multimodal model

Google releases Gemma 4 12B, an encoder-free multimodal model that runs on 16GB VRAM and processes vision and audio natively through the LLM backbone

NVIDIA Nemotron 3 Ultra

NVIDIA releases Nemotron 3 Ultra — 550B total / 55B active MoE hybrid Mamba-Transformer, pretrained in NVFP4, with up to 5.9x throughput over competing open MoEs and 1M token context.

Confidential submission of draft S-1 to the SEC | OpenAI

OpenAI files a confidential S-1 with the SEC, preemptively announcing it publicly ahead of an expected leak — while noting IPO timing remains undecided as some strategic moves are easier as a private company.

MLOps/LLMOps/AgentOps

How Google SRE is using agentic AI to improve operations

An article about how Google SRE is wiring agentic AI across the full incident lifecycle — dynamic anomaly detection, autonomous investigation, and a RAG layer over historical incidents to inform mitigation agents.

Unlocking dependable responses with Gemini Enterprise Agent Platform’s Agentic RAG

Google launches an agentic RAG framework on Gemini Enterprise Agent Platform with a Sufficient Context Agent that iterates until retrieval gaps are filled, hitting 90.1% accuracy on multi-hop queries — up to 34% over vanilla RAG.

Learning

North Mini Code: Cohere’s First Model For Developers

An article about how Cohere trained North Mini Code’s agentic coding capabilities — using cascaded SFT as an RLVR primer, joint multi-environment RL across 70k containerized repos, and cross-harness data mixing to generalize across SWE-Agent, mini-SWE-agent, and OpenCode scaffolds.

Testing Gemini models for scheming tendencies | by DeepMind Safety Research

DeepMind releases two scheming eval frameworks for Gemini — Gram (simulated agentic environments) and honeypots (real safety codebases) — finding 2–3% unprompted sabotage rates and no coherent misalignment.

Claude Fable 5 and new AI safety fables

Nathan Lambert argues that Anthropic’s undisclosed safety filters in Claude Fable 5 — which silently degrade responses for frontier AI research without notifying users — are competitive entrenchment dressed as safety policy.

Controlling the capital after AGI

An analytical piece from Epoch AI taxonomizing post-AGI wealth redistribution proposals — UBI, UBS, UBC, and sovereign wealth funds — along a single axis: how much control over capital, not just income, each scheme grants citizens.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

Papers & Publications

Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering

Abstract:

LLM-based Multi-Agent (LLM-MA) systems are increasingly applied to automate complex software engineering tasks such as requirements engineering, code generation, and testing. However, their operational efficiency and resource consumption remain poorly understood, hindering practical adoption due to unpredictable costs and environmental impact. To address this, we conduct an analysis of token consumption patterns in an LLM-MA system within the Software Development Life Cycle (SDLC), aiming to understand where tokens are consumed across distinct software engineering activities. We analyze execution traces from 30 software development tasks performed by the ChatDev framework using a GPT-5 reasoning model, mapping its internal phases to distinct development stages (Design, Coding, Code Completion, Code Review, Testing, and Documentation) to create a standardized evaluation framework. We then quantify and compare token distribution (input, output, reasoning) across these stages.

Our preliminary findings show that the iterative Code Review stage accounts for the majority of token consumption for an average of 59.4% of tokens. Furthermore, we observe that input tokens consistently constitute the largest share of consumption for an average of 53.9%, providing empirical evidence for potentially significant inefficiencies in agentic collaboration. Our results suggest that the primary cost of agentic software engineering lies not in initial code generation but in automated refinement and verification. Our novel methodology can help practitioners predict expenses and optimize workflows, and it directs future research toward developing more token-efficient agent collaboration protocols.

Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?

Abstract:

Spatial embodied intelligence requires agents to act to acquire information under partial observability. While multimodal foundation models excel at passive perception, their capacity for active, self-directed exploration remains understudied. We propose Theory of Space, defined as an agent’s ability to actively acquire information through self-directed, active exploration and to construct, revise, and exploit a spatial belief from sequential, partial observations. We evaluate this through a benchmark where the goal is curiosity-driven exploration to build an accurate cognitive map. A key innovation is spatial belief probing, which prompts models to reveal their internal spatial representations at each step. Our evaluation of state-of-the-art models reveals several critical bottlenecks. First, we identify an Active-Passive Gap, where performance drops significantly when agents must autonomously gather information. Second, we find high inefficiency, as models explore unsystematically compared to program-based proxies. Through belief probing, we diagnose that while perception is an initial bottleneck, global beliefs suffer from instability that causes spatial knowledge to degrade over time. Finally, using a false belief paradigm, we uncover Belief Inertia, where agents fail to update obsolete priors with new evidence. This issue is present in text-based agents but is particularly severe in vision-based models. Our findings suggest that current foundation models struggle to maintain coherent, revisable spatial beliefs during active exploration.

You may also enjoy Minimax M3, Direct Preference Optimization Beyond Chatbots, a paper on From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Introducing Claude Opus 4.8 \ Anthropic

Anthropic ships Claude Opus 4.8 at the same price as 4.7 — adding benchmark gains across coding and agentic tasks, a huge reduction in unremarked code flaws, and cheaper fast mode.

Codex for every role, tool, and workflow

OpenAI expands Codex beyond developers with six role-specific plugins covering 62 apps and 110 skills, a preview of shareable hosted Sites, and inline annotations.

Introducing Search Toolkit

Mistral launches Search Toolkit, an open-source composable framework unifying ingestion, retrieval, and evaluation into a single production-ready pipeline for enterprise RAG and search applications.

OpenAI frontier models and Codex are now available on AWS

OpenAI makes its frontier models and Codex generally available on AWS via Amazon Bedrock — including GovCloud regions — letting enterprises adopt OpenAI through existing AWS security, compliance, and procurement workflows.

MiniMax M3: Frontier Coding, 1M Context, Native Multimodality — All in One Model

MiniMax launches M3, currently the only open-weight model combining frontier coding performance, native multimodality, and a 1M-token context window via a new sparse attention architecture.

MLOps/LLMOps/AgentOps

Agent Tracing and Observability: Log & Debug Complex AI Systems

A guide on instrumenting agent tracing for multi-agent systems, covering why flat logging breaks at coordination boundaries, the three structural pillars of agentic observability, and how self-evolving agents compound debugging complexity.

What we’ve learned building cloud agents

A Cursor engineering retrospective on a year of shipping cloud agents, arguing the work is less “local agent on a server” and more building a full operating layer — covering environment fidelity, durable execution via Temporal, etc.

Learning

Open models lag state-of-the-art closed models by 4 months

An Epoch AI data insight measuring the open-to-closed model capability gap using their Epoch Capabilities Index (ECI), finding open-weight models now lag frontier closed models by an average of 4 months.

What we learned mapping a year’s worth of AI-enabled cyber threats \ Anthropic

Anthropic analyzed 832 banned malicious accounts over one year, finding that AI is accelerating cyberattack sophistication — shifting from initial access tactics to post-compromise operations.

What Is a Vector Lakebase?

An explainer introducing the Vector Lakebase — an architecture that unifies vector-database-grade serving with open lake storage and a shared semantic layer.

Direct Preference Optimization Beyond Chatbots

A blog post on applying Direct Preference Optimization to structured OCR — not for chat alignment — by using the SFT model’s own degeneration failures as rejection pairs.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

Papers & Publications

A Self-Healing Framework for Reliable LLM-Based Autonomous Agents

Abstract:

Autonomous agents based on Large Language Models (LLMs) are increasingly being utilized in complex software systems. However, reliability remains a significant challenge due to unpredictable failures such as hallucinations, execution errors, and inconsistent reasoning. This paper proposes a reliability-aware self-healing framework for LLM-based software agents. The framework integrates failure detection, reliability assessment, and automated recovery mechanisms. First, we define a taxonomy of failure types and introduce a quantitative reliability assessment model. Next, we propose a failure detection method that identifies abnormal agent behavior based on execution patterns and output consistency. Finally, we design a self-healing mechanism that dynamically recovers from failures through adaptive replanning and corrective prompting strategies. The proposed framework was implemented in a multi-agent workflow environment and evaluated using real-world task scenarios. Experimental results demonstrate that our approach significantly increases task success rates, reduces failure propagation, and enhances overall system robustness compared to existing methods. In particular, this study distinguishes itself by establishing an integrated monitoring system that combines the agent’s internal reasoning process with external execution results. These findings are expected to contribute to securing the stability of advanced autonomous systems and lowering the barriers to LLM adoption in production environments.

From Tokens to Thoughts: How LLMs and Humans Trade Compression for Meaning

Abstract:

Humans organize knowledge into compact conceptual categories that balance compression with semantic richness. Large Language Models (LLMs) exhibit impressive linguistic abilities, but whether they navigate this same compression-meaning trade-off remains unclear. We apply an Information Bottleneck framework to compare human conceptual structure with embeddings from 40+ LLMs using classic categorization benchmarks. We find that LLMs broadly align with human category boundaries, yet fall short on fine-grained semantic distinctions. Unlike humans, who maintain ``inefficient’‘ representations that preserve contextual nuance, LLMs aggressively compress, achieving more optimal information-theoretic compression at the cost of semantic richness. Surprisingly, encoder models outperform much larger decoder models in human alignment, suggesting that understanding and generation rely on distinct representational mechanisms. Training-dynamics analysis reveals a two-phase trajectory: rapid initial concept formation followed by architectural reorganization, during which semantic processing migrates from deep to mid-network layers as the model discovers increasingly efficient, sparser encodings. These divergent strategies, where LLMs optimize for compression and humans for adaptive utility, reveal fundamental differences between artificial and natural intelligence. This highlights the need for models that preserve the conceptual ``inefficiencies’‘ essential for human-like understanding.

You may also enjoy Google reimagines search with AI agents and generative interfaces, Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models, a paper on Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

DeepSWE

Datacurve introduces DeepSWE, a contamination-free coding benchmark of 113 from-scratch tasks across 91 repos and 5 languages, where GPT-5.5 leads at 70% and frontier models separate far more sharply than on SWE-Bench Pro.

Google reimagines search with AI agents and generative interfaces

Google overhauls Search at I/O 2026 with always-on Search Agents that monitor the web and report back, plus generative UI that builds interactive mini-apps on the fly via Antigravity and Gemini 3.5 Flash.

OpenRouter raises $113M to bring order to enterprise AI inference routing

OpenRouter raises $113M Series B led by CapitalG to scale its multi-model inference routing platform.

MLOps/LLMOps/AgentOps

The Best AI Observability Tools for Agentic Systems in 2026

A guide to the top AI observability tools in 2026 for agentic systems. Learn which platforms are best for tracing, evaluation, debugging, testing, and monitoring in production.

Learning

What Held Up at 3 AM: One Engineer’s RAG Case Study

Learn how WeaveCLI, a unified command-line tool for RAG over eleven vector databases was built, using Opik’s tracing capabilities, and configurable retrieval pipelines.

Building self-improving tax agents with Codex

OpenAI and Thrive Holdings build Tax AI, a Codex-driven self-improving agent that turns repeated accountant corrections into bounded evals

Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

NVIDIA releases Nemotron-Labs Diffusion, an open model family (3B/8B/14B plus an 8B VLM) that combines autoregressive, diffusion, and self-speculation modes in one checkpoint

The Agent Harness: Why the LLM Is the Smallest Part of Your Agent System

A technical article arguing that the LLM is the smallest part of a production agent system, with the real engineering living in a six-component harness and a deeper platform layer that determines production reliability.

An Engineer’s Guide to Better AI Skills: Implementing a Testing Process to Optimize Agent Performance in Any Repository or Skill

A practical guide from Pinterest Engineering on building a test harness to measure and improve how reliably coding agents invoke custom skills through frontmatter tuning and other techniques.

Harness, Scaffold, and the AI Agent Terms Worth Getting Right

A glossary that pins down the agent vocabulary people keep using loosely — clarifying the scaffold-versus-harness distinction and grounding terms like context engineering, policy, skills, and more.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

mksglu/context-mode

Context window optimization for AI coding agents.

Papers & Publications

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

Abstract:

Agent skills today are hand-crafted, generated one-shot, or evolved through loosely controlled self-revision, none of which behaves like a deep-learning optimizer for the skill, and none of which reliably improves over its starting point under feedback. We argue the skill should instead be trained as the external state of a frozen agent, with the same discipline that makes weight-space optimization reproducible. SkillOpt is, to our knowledge, the first systematic controllable text-space optimizer for agent skills: a separate optimizer model turns scored rollouts into bounded add/delete/replace edits on a single skill document, and an edit is accepted only when it strictly improves a held-out validation score. A textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta update make skill training stable while adding zero inference-time model calls at deployment. Across six benchmarks, seven target models, and three execution harnesses (direct chat, Codex, Claude Code), SkillOpt is best or tied on all 52 evaluated (model, benchmark, harness) cells and beats every per-cell competitor among human, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills. On GPT-5.5 it lifts the average no-skill accuracy by +23.5 points in direct chat, by +24.8 inside the Codex agentic loop, and by +19.1 inside Claude Code. Transfer experiments further show that optimized skill artifacts retain value when moved across model scales, between Codex and Claude Code execution environments, and to a nearby math benchmark without further optimization.

Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity

Abstract:

LLM-based agents are assumed to integrate environmental observations into their reasoning: discovering highly relevant but unexpected information should naturally lead to a model exploiting its own discoveries. We show that this assumption is false for current LLM-based agents, which struggle to reflect or react to unexpected information. Across three benchmarks (Terminal-Bench, SWE-Bench, AppWorld), we inject complete task solutions into the agent environments to deliberately expose a task’s solution to a model. While agents discover these solutions on Terminal-Bench in 79-81% of runs, they interact, or exploit, them in only 37-50% of cases. This gap is starkest in AppWorld: agents see documentation stating that a command “returns the complete solution to this task” in over 90% of attempts but exploit this in fewer than 7% of trials. We show that agents lack what we call environmental curiosity: the capability to recognize and investigate unexpected but relevant observations in response to environmental stimuli. We identify three main factors influencing environmental curiosity: available tools in the agent scaffold, test-time compute, and training data distribution. Our findings identify configurations that maximize curiosity also achieve the best performance on the unmodified benchmarks. Yet even jointly optimized agents still ignore discovered solutions in the majority of trials: current agents use the environment to fetch expected information, but not to revise their strategy or maximally exploit useful stimuli.

You may also enjoy Introducing Command A+, SLEIGHT-Bench: Finding Blind Spots in AI Monitors, a paper on CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Gemini 3.5: frontier intelligence with action

Google launched Gemini 3.5, leading with 3.5 Flash—a model delivering flagship-tier agentic and coding performance at under half the cost.

Introducing Command A+

Cohere released Command A+ open-source —a 218B/25B-active MoE model for enterprise agentic workflows that runs on as little as two H100s or one Blackwell GPU, supports 48 languages, and adds multimodal reasoning.

Introducing Gemini Omni

Google introduced Gemini Omni, a natively multimodal generation model debuting with Omni Flash—it creates and conversationally edits video from any combination of image, audio, video, and text inputs, rolling out across the Gemini app, Flow, and YouTube Shorts.

Introducing Grok Build

xAI launched Grok Build — a coding CLI powered by Grok 4.3 Heavy, featuring a 2M-token context window, 8 parallel subagents, and more.

FLUX Outpainting: Extend any image, in any direction

Black Forest Labs launched FLUX Outpainting, a purpose-built API endpoint that extends images in any direction without prompts.

Introducing Composer 2.5

Cursor released Composer 2.5, a coding model (built on Moonshot’s Kimi K2.5) with gains on long-horizon agentic tasks.

MLOps/LLMOps/AgentOps

LLM Cost Tracking Solution: How to Monitor and Control AI Spend in Agentic Systems

A guide on treating LLM cost as an observability problem in agentic systems, using span/trace/project-level tracing to pinpoint token-burning prompts and routing.

Context is all you need: Introducing Redis Iris

Redis launched Redis Iris, a context engine sitting between agents and enterprise data—bundling five tools (two new: Context Retriever and Agent Memory) to deliver navigable, fresh, low-latency context with semantic caching that cuts token costs up to 90%.

Learning

Text Analysis for Hybrid Search: Tokenization, Stopwords & Accent Folding

A technical guide on how Weaviate v1.37 makes BM25 tokenization observable and per-property configurable—covering accent folding, per-language stopwords, and more.

Running PyTorch Models on Apple Silicon GPUs with the ExecuTorch MLX Delegate

PyTorch released the experimental ExecuTorch MLX delegate, a backend that runs PyTorch models on Apple Silicon GPUs via Apple’s MLX framework

Codex-maxxing

A power user’s playbook for extracting more value from Codex—using durable threads, file-based memory, verifiable goals, and self-scheduling loops to turn it into a workspace where long-running knowledge work keeps progressing between sessions.

SLEIGHT-Bench: Finding Blind Spots in AI Monitors

Anthropic researchers released SLEIGHT-Bench, a benchmark of 40 synthetic attacks across 11 categories that exploit “blind spots” in frontier AI monitors—on the Opus 4.6 monitor, 50% of attacks evaded all 10 trials and only 8 of 40 were reliably caught.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

generalaction/emdash

Emdash is the Open-Source Agentic Development Environment. Run multiple coding agents in parallel. Use any provider.

Papers & Publications

Lance: Unified Multimodal Modeling by Multi-Task Synergy

Abstract:

We present Lance, a lightweight native unified model supporting multimodal understanding, generation, and editing for both images and videos. Rather than relying on model capacity scaling or text-image-dominant designs, Lance explores a practical paradigm for unified multimodal modeling via collaborative multi-task training. It is grounded in two core principles: unified context modeling and decoupled capability pathways. Specifically, Lance is trained from scratch and employs a dual-stream mixture-of-experts architecture on shared interleaved multimodal sequences, enabling joint context learning while decoupling the pathways for understanding and generation. We further introduce modality-aware rotary positional encoding to mitigate interference among heterogeneous visual tokens and boost cross-task alignment. During training, Lance adopts a staged multi-task training paradigm with capability-oriented objectives and adaptive data scheduling to strengthen both semantic comprehension and visual generation performance. Experimental results demonstrate that Lance substantially outperforms existing open-source unified models in image and video generation, while retaining strong multimodal understanding capabilities.

CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence

Abstract:

Multimodal Large Language Models (MLLMs) have significantly advanced document understanding, yet current Doc-VQA evaluations score only the final answer and leave the supporting evidence unchecked. This answer-only approach masks a critical failure mode: a model can land on the correct answer while grounding it in the wrong passage -- a critical risk in high-stakes domains like law, finance, and medicine, where every conclusion must be traceable to a specific source region. To address this, we introduce CiteVQA, a benchmark that requires models to return element-level bounding-box citations alongside each answer, evaluating both jointly. CiteVQA comprises 1,897 questions across 711 PDFs spanning seven domains and two languages, averaging 40.6 pages per document. To ensure fidelity and scalability, the ground-truth citations are generated by an automated pipeline-which identifies crucial evidence via masking ablation-and are subsequently validated through expert review. At the core of our evaluation is Strict Attributed Accuracy (SAA), which credits a prediction only when the answer and the cited region are both correct. Auditing 20 MLLMs reveals a pervasive Attribution Hallucination: models frequently produce the right answer while citing the wrong region. The strongest system (Gemini-3.1-Pro-Preview) achieves an SAA of only 76.0, and the strongest open-source MLLM reaches just 22.5. Ultimately, towards trustworthy document intelligence, CiteVQA exposes a reliability gap that answer-only evaluations overlook, providing the instrumentation needed to close it.

You may also enjoy Introducing Perceptron Mk1, Teaching Claude why, a paper on ProgramBench: Can Language Models Rebuild Programs From Scratch?, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Introducing Perceptron Mk1

Perceptron AI launches Mk1, a video and embodied-reasoning vision-language model priced roughly 80–90% cheaper than Claude Sonnet 4.5, GPT-5, and Gemini 3.1 Pro.

Notion just turned its workspace into a hub for AI agents

Notion launches Developer Platform turning its workspace into an agent orchestration hub with custom code Workers, external database sync, and native integrations for Claude Code, Cursor, Codex, and Decagon.

Interaction Models: A Scalable Approach to Human-AI Collaboration

Thinking Machines unveils TML-Interaction-Small, a 276B MoE (12B active) interaction model trained from scratch with 200ms time-aligned micro-turns that natively handles concurrent audio, video, and text without VAD-style harnesses.

Unsloth Joins the PyTorch Ecosystem

Unsloth joins the PyTorch Ecosystem Landscape, recognizing its open-source contributions including 2× faster training with 70% less VRAM, FP8 RL for consumer GPUs, and 250M+ model downloads.

MLOps/LLMOps/AgentOps

Hidden Technical Debt of AI Systems: Agent Harness

A Hanchung Lee essay reframing Sculley’s 2015 ML technical debt diagram for the agent era, arguing the agent runtime (harness + state) — not the model — is where most spend, incidents, and architectural debt are now accumulating.

Building Blocks for Foundation Model Training and Inference on AWS

A reference guide from Amazon mapping AWS’s four-layer infrastructure stack to foundation model pre-training, post-training, and inference workloads.

How to Eliminate Pipeline Friction in AI Model Serving

A practical NVIDIA guide laying out 18 best practices to eliminate AI model-serving friction across export issues, unsupported ops, dynamic input shapes, and version mismatches

Learning

Teaching Claude why

An Anthropic post detailing how teaching Claude why actions are aligned — via constitutional documents and ethical reasoning, not just demonstrations — drove blackmail rates from 96% (Opus 4) to 0% on every Claude model since Haiku 4.5.

Accelerating Gemma 4: faster inference with multi-token prediction drafters

Google releases Multi-Token Prediction (MTP) drafters for Gemma 4 models, delivering up to 3x faster inference via speculative decoding with zero quality degradation.

Vibe coding and agentic engineering are getting closer than I’d like

A post by Simon Willison observing that vibe coding and agentic engineering are converging in his own workflow as he increasingly ships production code from Claude Code without reviewing every line.

How fast is autonomous AI cyber capability advancing?

UK AISI reports the length of cyber tasks frontier models can autonomously complete is doubling every 4.7 months — accelerating from 8 months last November — with Claude Mythos Preview and GPT-5.5 exceeding even that trend.

Reimagining the mouse pointer for the AI era

A design-principles post from Google DeepMind reframing the mouse pointer as a Gemini-powered context-aware partner, built on four principles: maintain the flow, show and tell, embrace “this/that” deixis, and turn pixels into actionable entities.

Full Text Search: Architecture and Design

A technical architecture post from Pinecone introducing full-text search built on Tantivy, delivering Lucene query syntax, BM25 scoring, 18-language tokenization, and 22.7ms p50 latency on 6.4M Wikipedia articles.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

Papers & Publications

δ-mem: Efficient Online Memory for Large Language Models

Abstract:

Large language models increasingly need to accumulate and reuse historical information in long-term assistants and agent systems. Simply expanding the context window is costly and often fails to ensure effective context utilization. We propose δ-mem, a lightweight memory mechanism that augments a frozen full-attention backbone with a compact online state of associative memory. δ-mem compresses past information into a fixed-size state matrix updated by delta-rule learning, and uses its readout to generate low-rank corrections to the backbone’s attention computation during generation. With only an 8×8 online memory state, δ-mem improves the average score to 1.10× that of the frozen backbone and 1.15× that of the strongest non-δ-mem memory baseline. It achieves larger gains on memory-heavy benchmarks, reaching 1.31× on MemoryAgentBench and 1.20× on LoCoMo, while largely preserving general capabilities. These results show that effective memory can be realized through a compact online state directly coupled with attention computation, without full fine-tuning, backbone replacement, or explicit context extension.

ProgramBench: Can Language Models Rebuild Programs From Scratch?

Abstract:

Turning ideas into full software projects from scratch has become a popular use case for language models. Agents are being deployed to seed, maintain, and grow codebases over extended periods with minimal human oversight. Such settings require models to make high-level software architecture decisions. However, existing benchmarks measure focused, limited tasks such as fixing a single bug or developing a single, specified feature. We therefore introduce ProgramBench to measure the ability of software engineering agents to develop software holisitically. In ProgramBench, given only a program and its documentation, agents must architect and implement a codebase that matches the reference executable’s behavior. End-to-end behavioral tests are generated via agent-driven fuzzing, enabling evaluation without prescribing implementation structure. Our 200 tasks range from compact CLI tools to widely used software such as FFmpeg, SQLite, and the PHP interpreter. We evaluate 9 LMs and find that none fully resolve any task, with the best model passing 95\% of tests on only 3\% of tasks. Models favor monolithic, single-file implementations that diverge sharply from human-written code.

You may also enjoy GPT-5.5 Instant, AI Chips: why they cost as much as a car, and why companies can’t get enough, a paper on OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

MolmoAct 2: An open foundation for robots that work in the real world

Ai2 releases MolmoAct 2, a fully open robotics foundation model running up to 37x faster than its predecessor, alongside the largest open-source bimanual manipulation dataset.

GPT-5.5 Instant: smarter, clearer, and more personalized

OpenAI releases GPT-5.5 Instant as ChatGPT’s new default model, cutting hallucinations by 52.5% on high-stakes prompts and using 30% fewer words while pulling context from past chats, files, and more.

New Compute Partnership with Anthropic

SpaceXAI signs a deal with rival Anthropic for full access to Colossus 1, with Anthropic also expressing interest in jointly developing multi-gigawatt orbital compute.

Blitzy raises $200M at $1.4B valuation to deploy thousands of coding agents in parallel

Blitzy raises $200M at $1.4B valuation to scale its enterprise platform that orchestrates thousands of parallel coding agents across 100M+ line legacy codebases, scoring 66.5% on SWE-Bench Pro.

Monday.com relaunches as an AI work platform with native agents

Monday.com relaunches as an “AI work platform” with native agents that draft campaigns, qualify leads, and triage tickets across its 250,000+ customers, plus one-click connectors to Claude, ChatGPT, Copilot, and Gemini.

MLOps/LLMOps/AgentOps

Introducing the Opik Agent Playground

Comet launches Opik Agent Playground, a UI-based environment for testing and tweaking full agent configurations (prompts, models, tools) without touching code, opening iteration to PMs and domain experts.

Learning

AI Chips: why they cost as much as a car, and why companies can’t get enough

A primer on AI chip economics and supply: the entire frontier flows through TSMC and a handful of designers, with total compute capacity doubling every 7 months while performance-per-dollar doubles every 2.5 years.

How Agents Manage Other Agents: Four Subagents Patterns in 2026

A practical blog post about four subagent orchestration patterns—inline tool, fan-out, agent pool, and teams—each requiring progressively more capable models and offering different tradeoffs in control, lifetime, and result collection.

Don’t Outsource Your Understanding

An essay arguing the real AI risk isn’t using it but “cognitive surrender”—outsourcing verification too—evidenced by 1,300+ hallucinated court filings and 50 ICLR papers with fake citations.

Granite 4.1 LLMs: How They’re Built

An technical deep-dive on how Granite 4.1 was built: 15T-token, five-phase pretraining with long-context extension to 512K, SFT on 4.1M curated samples, and on-policy GRPO with DAPO loss.

How to Work and Compound with AI

A practitioner’s playbook for compounding with AI: treat context as infra, encode taste as config (CLAUDE.md, skills), make verification cheap, delegate bigger chunks in parallel, and mine transcripts to close the loop.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

zilliztech/claude-context

Claude Context is an MCP plugin that adds semantic code search to Claude Code and other AI coding agents, giving them deep context from your entire codebase.

Papers & Publications

ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration

Abstract:

This report describes ARIS (Auto-Research-in-sleep), an open-source research harness for autonomous research, including its architecture, assurance mechanisms, and early deployment experience. The performance of agent systems built on LLMs depends on both the model weights and the harness around them, which governs what information to store, retrieve, and present to the model. For long-horizon research workflows, the central failure mode is not a visible breakdown but a plausible unsupported success: a long-running agent can produce claims whose evidential support is incomplete, misreported, or silently inherited from the executor’s framing. Therefore, we present ARIS as a research harness that coordinates machine-learning research workflows through cross-model adversarial collaboration as a default configuration: an executor model drives forward progress while a reviewer from a different model family is recommended to critique intermediate artifacts and request revisions. ARIS has three architectural layers. The execution layer provides more than 65 reusable Markdown-defined skills, model integrations via MCP, a persistent research wiki for iterative reuse of prior findings, and deterministic figure generation. The orchestration layer coordinates five end-to-end workflows with adjustable effort settings and configurable routing to reviewer models. The assurance layer includes a three-stage process for checking whether experimental claims are supported by evidence: integrity verification, result-to-claim mapping, and claim auditing that cross-checks manuscript statements against the claim ledger and raw evidence, as well as a five-pass scientific-editing pipeline, mathematical-proof checks, and visual inspection of the rendered PDF. A prototype self-improvement loop records research traces and proposes harness improvements that are adopted only after reviewer approval.

OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories

Abstract:

Deep search capabilities have become an indispensable competency for frontier Large Language Model (LLM) agents, yet their development remains dominated by industrial giants. The typical industry recipe involves a highly resource-intensive pipeline spanning pre-training, continual pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL). In this report, we show that when fueled with informative and high-difficulty trajectories, a simple SFT approach could be surprisingly powerful for training frontier search agents. By introducing three simple data synthesis modifications: scaling knowledge graph size for richer exploration, expanding the tool set size for broader functionality, and strict low-step filtering, we establish a stronger baseline. Trained on merely 10.6k data points, our OpenSeeker-v2 achieves state-of-the-art performance across 4 benchmarks (30B-sized agents with ReAct paradigm): 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity’s Last Exam, and 78.0% on xbench, surpassing even Tongyi DeepResearch trained with heavy CPT+SFT+RL pipeline, which achieves 43.4%, 46.7%, 32.9%, and 75.0%, respectively. Notably, OpenSeeker-v2 represents the first state-of-the-art search agent within its model scale and paradigm to be developed by a purely academic team using only SFT. We are excited to open-source the OpenSeeker-v2 model weights and share our simple yet effective findings to make frontier search agent research more accessible to the community.

You may also enjoy Mistral AI’s Workflows, Four ways Google Research scientists have been using Empirical Research Assistance, a paper on From Skills to Talent: Organising Heterogeneous Agents as a Real-World Company, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

OpenAI models, Codex, and Managed Agents come to AWS

OpenAI brings GPT-5.5, Codex, and Managed Agents to Amazon Bedrock in limited preview, a day after ending Microsoft’s exclusive cloud license.

Mistral AI launches Workflows, a Temporal-powered orchestration engine already running millions of daily executions

Mistral AI launches Workflows in public preview, a Temporal-powered durable execution engine that lets enterprises run human-in-the-loop AI processes with data staying inside their own infrastructure.

Ineffable Intelligence raises $1.1B at $5.1B valuation to build an AI ‘superlearner’

AlphaGo creator David Silver’s new British startup Ineffable Intelligence raises $1.1B to build a “superlearner” AI that generates entirely new knowledge via RL without pretraining.

American AI startup Poolside launches free, high-performing open model Laguna XS.2 for local agentic coding

Poolside releases two agentic coding models — the open-weight Laguna XS.2 and proprietary Laguna M.1 — both trained from scratch and free to use temporarily via API, alongside a terminal coding agent and web IDE.

MLOps/LLMOps/AgentOps

Hidden Technical Debt of AI Systems: Agent Runtime

A technical blog post arguing that the agent runtime — the sandboxed execution environment wrapping the model — is the emerging hidden technical debt of AI systems.

Context decay, orchestration drift, and the rise of silent failures in AI systems

A practical guide about four enterprise AI failure patterns — context degradation, orchestration drift, silent partial failure, and automation blast radius — that standard infrastructure monitoring cannot detect, and what teams must add to catch them.

Monitoring LLM behavior: Drift, retries, and refusal patterns

A practical guide on instrumenting LLM applications with two complementary evaluation pipelines — offline regression testing and online behavioral telemetry — to detect model drift before and after deployment.

Learning

Four ways Google Research scientists have been using Empirical Research Assistance

A Google Research blog post on four real-world applications of their Empirical Research Assistance (ERA) tool — an LLM-backed system for generating expert-level scientific software — spanning epidemiology, cosmology, climate monitoring, and neuroscience.

AI Organizations Can Be More Effective but Less Aligned than Individual Agents

An Anthropic research paper found that teams of individually aligned AI agents can still collectively produce less ethical — but more effective — solutions than a single agent, suggesting AI safety research needs to move beyond studying agents in isolation.

Introducing ARFBench: A time series question-answering benchmark based on real incidents

CMU and Datadog introduce ARFBench, a 750-question time series QA benchmark derived from real production incidents, where the best model (GPT-5 at 62.7% accuracy) still trails domain experts by ~9 points but a hybrid TSFM-VLM oracle reaches 87.2%.

Decoupled DiLoCo: Resilient, Distributed AI Training at Scale

Google DeepMind releases Decoupled DiLoCo, a fault-tolerant distributed training architecture that achieves 88% goodput vs. 27% for standard data-parallel at scale, using ~240x less inter-datacenter bandwidth with no measurable ML performance loss.

How to correctly use MCP servers with your AI Agents

A practical guide on avoiding context bloat from MCP servers by using two patterns — user-triggered @mention injection for ad-hoc tool loading, and scoped subagent declarations.

DeepSeek V4 Compressed Attention: How the KV-Cache Shrinks to Just 2%

A technical explainer on how DeepSeek V4 achieves 1M-token context windows by compressing the KV cache to just 2% of standard size — combining coarse and fine-grained sequence-dimension compression across a hybrid layer stack.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

facebookresearch/sapiens2

1K resolution vision transformers pretrained on 1B human images.

Papers & Publications

Recursive Multi-Agent Systems

Abstract:

Recursive or looped language models have recently emerged as a new scaling axis by iteratively refining the same model computation over latent states to deepen reasoning. We extend such scaling principle from a single model to multi-agent systems, and ask: Can agent collaboration itself be scaled through recursion? To this end, we introduce RecursiveMAS, a recursive multi-agent framework that casts the entire system as a unified latent-space recursive computation. RecursiveMAS connects heterogeneous agents as a collaboration loop through the lightweight RecursiveLink module, enabling in-distribution latent thoughts generation and cross-agent latent state transfer. To optimize our framework, we develop an inner-outer loop learning algorithm for iterative whole-system co-optimization through shared gradient-based credit assignment across recursion rounds. Theoretical analyses of runtime complexity and learning dynamics establish that RecursiveMAS is more efficient than standard text-based MAS and maintains stable gradients during recursive training. Empirically, we instantiate RecursiveMAS under 4 representative agent collaboration patterns and evaluate across 9 benchmarks spanning mathematics, science, medicine, search, and code generation. In comparison with advanced single/multi-agent and recursive computation baselines, RecursiveMAS consistently delivers an average accuracy improvement of 8.3%, together with 1.2×-2.4× end-to-end inference speedup, and 34.6%-75.6% token usage reduction.

From Skills to Talent: Organising Heterogeneous Agents as a Real-World Company

Abstract:

Individual agent capabilities have advanced rapidly through modular skills and tool integrations, yet multi-agent systems remain constrained by fixed team structures, tightly coupled coordination logic, and session-bound learning. We argue that this reflects a deeper absence: a principled organisational layer that governs how a workforce of agents is assembled, governed, and improved over time, decoupled from what individual agents know. To fill this gap, we introduce \emph{OneManCompany (OMC)}, a framework that elevates multi-agent systems to the organisational level. OMC encapsulates skills, tools, and runtime configurations into portable agent identities called \emph{Talents}, orchestrated through typed organisational interfaces that abstract over heterogeneous backends. A community-driven \emph{Talent Market} enables on-demand recruitment, allowing the organisation to close capability gaps and reconfigure itself dynamically during execution. Organisational decision-making is operationalised through an \emph{Explore-Execute-Review} (E2R) tree search, which unifies planning, execution, and evaluation in a single hierarchical loop: tasks are decomposed top-down into accountable units and execution outcomes are aggregated bottom-up to drive systematic review and refinement. This loop provides formal guarantees on termination and deadlock freedom while mirroring the feedback mechanisms of human enterprises. Together, these contributions transform multi-agent systems from static, pre-configured pipelines into self-organising and self-improving AI organisations capable of adapting to open-ended tasks across diverse domains. Empirical evaluation on PRDBench shows that OMC achieves an 84.67% success rate, surpassing the state of the art by 15.48 percentage points, with cross-domain case studies further demonstrating its generality.

You may also enjoy Claude Opus 4.7, Notion Vector Search Architecture, OpenThoughts: Data Recipes for Reasoning Models, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Introducing ChatGPT Images 2.0

OpenAI releases ChatGPT Images 2.0, its first image model with native reasoning and web search, generating up to 8 coherent images per prompt at up to 2K resolution.

Introducing Claude Opus 4.7 \ Anthropic

Anthropic releases Claude Opus 4.7, a coding-focused upgrade over Opus 4.6 with significantly improved vision, a new xhigh effort level, and real-world cyber safeguards.

Introducing OpenAI Privacy Filter

OpenAI releases Privacy Filter, a 1.5B-parameter open-source, on-device PII detection and redaction model derived from gpt-oss, scoring 96% F1 on PII-Masking-300k.

Kimi K2.6 Tech Blog: Advancing Open-Source Coding

Moonshot AI open-sources Kimi K2.6, a coding and long-horizon agent model that scales agent swarms to 300 concurrent sub-agents across 4,000 coordinated steps, with benchmark results competitive with GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro and agentic tasks.

Google’s new Deep Research and Deep Research Max agents can search the web and your private data

Google launches two Gemini 3.1 Pro-powered autonomous research agents — Deep Research and Deep Research Max — that combine open web search with proprietary enterprise data via MCP in a single API call.

MLOps/LLMOps/AgentOps

Introducing Ollie: Auto-Fix Your Agent’s Codebase

Comet announces Ollie, a coding assistant embedded in the Opik platform that closes the observability-to-action loop by autonomously analyzing agent traces, diagnosing failures, patching code, and writing regression tests — all within a single workflow.

Introducing Opik Test Suites: Straightforward Unit & Regression Testing for AI Agents

Comet announces Opik Test Suites, a regression testing framework for AI agents that replaces dataset-based evaluation scores with software-style pass/fail assertions written in plain English.

Learning

Designing synthetic datasets for the real world: Mechanism design and reasoning from first principles

A Google Research blog post introducing Simula, a reasoning-first synthetic data framework that treats dataset generation as mechanism design — controlling diversity, complexity, and quality as independent axes.

Benchmarking multimodal document search in OpenSearch: Three approaches compared

A technical benchmark comparing ColPali late-interaction reranking, BDA modality-aware embedding, and text-only chunking for multimodal document search in OpenSearch across quality, latency, and ingest performance on 1,000 report pages.

Notion Vector Search Architecture: What Comes Next

A blog post analyzing Notion’s two-year vector search evolution as a proxy for the harder infrastructure problems — offline context engineering, embedding model upgrades, and real-time/batch unification — that scaling multiple AI features will demand next.

Engram: Memory by Weaviate

Weaviate announces Engram, a managed memory service that uses async pipelines to extract, deduplicate, and maintain agent memories on top of Weaviate’s vector database.

Automated Weak-to-Strong Researcher

Anthropic’s Claude-powered Automated Alignment Researcher achieves a 0.97 performance gap recovered score on weak-to-strong supervision in 5 days — versus 0.23 by human researchers in 7 days.

Breaking Opus 4.7 with ChatGPT (Hacking Claude’s Memory)

A security research post demonstrating a ChatGPT-generated adversarial image that successfully hijacked Claude Opus 4.7’s memory tool via indirect prompt injection — succeeding 5 out of 10 attempts before Anthropic patched the specific exploit within 24 hours.

Libraries & Code

comet-ml/opik

An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

google/skills

Agent Skills for Google products and technologies

Papers & Publications

Adam’s Law: Textual Frequency Law on Large Language Models

Abstract:

While textual frequency has been validated as relevant to human cognition in reading speed, its relatedness to Large Language Models (LLMs) is seldom studied. We propose a novel research direction in terms of textual data frequency, which is an understudied topic, to the best of our knowledge. Our framework is composed of three units. First, this paper proposes Textual Frequency Law (TFL), which indicates that frequent textual data should be preferred for LLMs for both prompting and fine-tuning. Since many LLMs are closed-source in their training data, we propose using online resources to estimate the sentence-level frequency. We then utilize an input paraphraser to paraphrase the input into a more frequent textual expression. Next, we propose Textual Frequency Distillation (TFD) by querying LLMs to conduct story completion by further extending the sentences in the datasets, and the resulting corpora are used to adjust the initial estimation. Finally, we propose Curriculum Textual Frequency Training (CTFT) that fine-tunes LLMs in an increasing order of sentence-level frequency. Experiments are conducted on our curated dataset Textual Frequency Paired Dataset (TFPD) on math reasoning, machine translation, commonsense reasoning and agentic tool calling. Results show the effectiveness of our framework.

OpenThoughts: Data Recipes for Reasoning Models

Abstract:

Reasoning models have made rapid progress on many benchmarks involving math, code, and science. Yet, there are still many open questions about the best train- ing recipes for reasoning since state-of-the-art models often rely on proprietary datasets with little to no public information available. To address this, the goal of the OpenThoughts project is to create open-source datasets for training reasoning models. Our OpenThoughts2-1M dataset led to OpenThinker2-32B, the first model trained on public reasoning data to match DeepSeek-R1-Distill-32B on standard reasoning benchmarks such as AIME and LiveCodeBench. We then improve our dataset further by systematically investigating each step of our data genera- tion pipeline with 1,000+ controlled experiments, which led to OpenThoughts3. Scaling the pipeline to 1.2M examples and using QwQ-32B as teacher yields our OpenThinker3-7B model, which achieves state-of-the-art results: 53% on AIME 2025, 51% on LiveCodeBench 06/24-01/25, and 54% on GPQA Dia- mond – improvements of 15.3, 17.2, and 20.5 percentage points compared to the DeepSeek-R1-Distill-Qwen-7B. All of our datasets and models are available on openthoughts.ai.

You may also enjoy Gemini Robotics ER 1.6: Enhanced Embodied Reasoning, Should AI Step Aside?: Teaching Agents When Humans Want to Intervene, a paper on CodeTracer: Towards Traceable Agent States, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

The 2026 AI Index Report | Stanford HAI

Stanford HAI releases the 2026 AI Index — a 400+ page annual report tracking AI’s technical performance, investment, labor market effects, policy landscape, and public sentiment across nine chapters.

Gemini Robotics ER 1.6: Enhanced Embodied Reasoning

Google DeepMind releases Gemini Robotics-ER 1.6, a robotics-specialized reasoning model with upgraded spatial reasoning, multi-view success detection, and instrument reading (93% accuracy with agentic vision).

Introducing Muse Spark: Scaling Towards Personal Superintelligence

Meta’s Superintelligence Labs launches Muse Spark — a natively multimodal reasoning model with multi-agent “Contemplating” mode that achieves 58% on Humanity’s Last Exam.

Gemini 3.1 Flash TTS: the next generation of expressive AI speech

Google launches Gemini 3.1 Flash TTS, a text-to-speech model with natural-language audio tags for granular vocal control across 70+ languages, available via Gemini API, Vertex AI, and Google Vids.

Introducing MAI-Image-2-Efficient: Faster, More Efficient Image Generation

Microsoft releases MAI-Image-2-Efficient — 22% faster and 4x more GPU-efficient than MAI-Image-2, targeting high-volume and real-time image generation workloads.

Introducing routines in Claude Code

Anthropic launches Routines in Claude Code — serverless automations triggered by schedule, API call, or GitHub webhook events, with daily limits of 5–25 runs depending on plan tier.

MLOps/LLMOps

Multimodal LLM Evaluation: A Developer’s Guide to Multimodal Language Models

A guide to evaluating multimodal LLMs, highlighting why text-only metrics fall short for image, audio, and video inputs, while outlining methods for grounding outputs and using LLM-based evaluation to measure real-world performance.

Learning

Should AI Step Aside?: Teaching Agents When Humans Want to Intervene

A research blog post introducing CowCorpus and PlowPilot — a dataset and intervention-aware web agent system that predicts when users want to take over, yielding a 26.5% improvement in user-rated usefulness over a fully autonomous baseline.

MirrorCode: Evidence that AI can already do some weeks-long coding tasks

A research report from Epoch AI introducing MirrorCode, a long-horizon coding benchmark, showing Claude Opus 4.6 can autonomously reimplement a 16,000-line bioinformatics toolkit estimated to take a human engineer 2–17 weeks.

8 Tips for Writing Agent Skills

A practical guide on authoring effective agent skills, covering description precision, instruction conciseness, layered context loading, and when to retire skills as model capabilities advance.

The AI Revolution in Math Has Arrived

A Quanta Magazine feature documenting how AI has become a genuine research accelerator, with mathematicians using it to discover and prove new results in days rather than months.

Gemma 4 Fine-tuning Guide

Unsloth’s technical guide for fine-tuning Google’s Gemma 4 family covering VRAM requirements, critical bug fixes for KV-sharing and gradient accumulation, and recipes for SFT, vision, audio, and GRPO training.

Towards developing future-ready skills with generative AI

A Google blog post introducing Vantage, a GenAI-powered assessment platform that places students in AI-simulated multi-party conversations to measure “future-ready” skills.

Libraries & Code

comet-ml/opik

An open-source LLM evaluation tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

Papers & Publications

Introspective Diffusion Language Models

Abstract:

Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR models agree with their own generations, while DLMs often do not. We define the introspective acceptance rate, which measures whether a model accepts its previously generated tokens. This reveals why AR training has a structural advantage: causal masking and logit shifting implicitly enforce introspective consistency. Motivated by this observation, we introduce Introspective Diffusion Language Model (I-DLM), a paradigm that retains diffusion-style parallel decoding while inheriting the introspective consistency of AR training. I-DLM uses a novel introspective strided decoding (ISD) algorithm, which enables the model to verify previously generated tokens while advancing new ones in the same forward pass. From a systems standpoint, we build I-DLM inference engine on AR-inherited optimizations and further customize it with a stationary-batch scheduler. To the best of our knowledge, I-DLM is the first DLM to match the quality of its same-scale AR counterpart while outperforming prior DLMs in both model quality and practical serving efficiency across 15 benchmarks. It reaches 69.6 on AIME-24 and 45.7 on LiveCodeBench-v6, exceeding LLaDA-2.1-mini (16B) by more than 26 and 15 points, respectively. Beyond quality, I-DLM is designed for the growing demand of large-concurrency serving, delivering about 3x higher throughput than prior state-of-the-art DLMs.

CodeTracer: Towards Traceable Agent States

Abstract:

Code agents are advancing rapidly, but debugging them is becoming increasingly difficult. As frameworks orchestrate parallel tool calls and multi-stage workflows over complex tasks, making the agent’s state transitions and error propagation hard to observe. In these runs, an early misstep can trap the agent in unproductive loops or even cascade into fundamental errors, forming hidden error chains that make it hard to tell when the agent goes off track and why. Existing agent tracing analyses either focus on simple interaction or rely on small-scale manual inspection, which limits their scalability and usefulness for real coding workflows. We present CodeTracer, a tracing architecture that parses heterogeneous run artifacts through evolving extractors, reconstructs the full state transition history as a hierarchical trace tree with persistent memory, and performs failure onset localization to pinpoint the failure origin and its downstream chain. To enable systematic evaluation, we construct CodeTraceBench from a large collection of executed trajectories generated by four widely used code agent frameworks on diverse code tasks (e.g., bug fixing, refactoring, and terminal interaction), with supervision at both the stage and step levels for failure localization. Experiments show that CodeTracer substantially outperforms direct prompting and lightweight baselines, and that replaying its diagnostic signals consistently recovers originally failed runs under matched budgets. Our code and data are publicly available.

You may also enjoy Claude Managed Agents, Evaluating alignment of behavioral dispositions in LLMs, a paper on TriAttention: Efficient Long Reasoning with Trigonometric KV Compression, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Qwen: Qwen3.6-Plus: Towards Real World Agents

Alibaba launches Qwen3.6-Plus, a frontier agentic coding model that matches or beats Claude Opus 4.5 on SWE-bench and Terminal-Bench 2.0.

Claude Managed Agents: get to production 10x faster

Anthropic launches Claude Managed Agents in public beta — a suite of composable, cloud-hosted agent APIs that abstract away sandboxing, state management, permissioning, and orchestration, enabling teams to ship production agents in days instead of months.

Gemma 4: Our most capable open models to date

Google releases Gemma 4, a family of four open models — with the 31B ranking #3 among open models on Arena AI and outcompeting models 20x its size.

Modus secures $85M to expand AI-powered audit and accounting partnerships

Modus Audit raises $85M to deploy AI across audit and accounting firm workflows.

Ollama is now powered by MLX on Apple Silicon in preview

Ollama 0.19 launches MLX-powered inference on Apple Silicon, delivering ~2x gains in prefill and decode speed on M5 chips, with NVFP4 quantization support and smarter KV cache reuse for agentic workloads.

MLOps/LLMOps

AI Agent Evaluation: Building Reliable Systems Beyond Simple Testing

A practical guide on why standard LLM evaluation breaks for agentic systems, covering compounding failures, process vs. outcome metrics, multi-turn state tracking, and the trace-evaluate-optimize loop needed for production agents.

Simulate realistic users to evaluate multi-turn AI agents in Strands Evals

A technical blog about ActorSimulator in AWS’s Strands Evals SDK, which generates persona-consistent, goal-driven simulated users to automate multi-turn agent evaluation at scale.

Learning

Components of A Coding Agent

A breakdown by Sebastian Raschka of the six architectural components that make coding agents (Claude Code, Codex CLI) meaningfully more capable than raw LLMs in a chat UI

Quantization from the ground up

A highly interactive, ground-up explainer on LLM quantization covering floating point formats, symmetric vs. asymmetric compression, outlier handling, and empirical quality/speed tradeoffs.

Evaluating alignment of behavioral dispositions in LLMs

A blog post on evaluating behavioral alignment across 25 LLMs, finding frontier models hit ~80–83% alignment with human consensus but are systematically overconfident in ambiguous scenarios and inconsistent between self-reported and revealed behavior.

Libraries & Code

comet-ml/opik

An open-source LLM evaluation tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

NousResearch/hermes-agent

The self-improving AI agent built by Nous Research. It’s the only agent with a built-in learning loop.

Papers & Publications

VOID: Video Object and Interaction Deletion

Abstract:

Existing video object removal methods excel at inpainting content “behind” the object and correcting appearance-level artifacts such as shadows and reflections. However, when the removed object has more significant interactions, such as collisions with other objects, current models fail to correct them and produce implausible results. We present VOID, a video object removal framework designed to perform physically-plausible inpainting in these complex scenarios. To train the model, we generate a new paired dataset of counterfactual object removals using Kubric and HUMOTO, where removing an object requires altering downstream physical interactions. During inference, a vision-language model identifies regions of the scene affected by the removed object. These regions are then used to guide a video diffusion model that generates physically consistent counterfactual outcomes. Experiments on both synthetic and real data show that our approach better preserves consistent scene dynamics after object removal compared to prior video object removal methods. We hope this framework sheds light on how to make video editing models better simulators of the world through high-level causal reasoning.

TriAttention: Efficient Long Reasoning with Trigonometric KV Compression

Abstract:

Extended reasoning in large language models (LLMs) creates severe KV cache memory bottlenecks. Leading KV cache compression methods estimate KV importance using attention scores from recent post-RoPE queries. However, queries rotate with position during RoPE, making representative queries very few, leading to poor top-key selection and unstable reasoning. To avoid this issue, we turn to the pre-RoPE space, where we observe that Q and K vectors are highly concentrated around fixed non-zero centers and remain stable across positions -- Q/K concentration. We show that this concentration causes queries to preferentially attend to keys at specific distances (e.g., nearest keys), with the centers determining which distances are preferred via a trigonometric series. Based on this, we propose TriAttention to estimate key importance by leveraging these centers. Via the trigonometric series, we use the distance preference characterized by these centers to score keys according to their positions, and also leverage Q/K norms as an additional signal for importance estimation. On AIME25 with 32K-token generation, TriAttention matches Full Attention reasoning accuracy while achieving 2.5x higher throughput or 10.7x KV memory reduction, whereas leading baselines achieve only about half the accuracy at the same efficiency. TriAttention enables OpenClaw deployment on a single consumer GPU, where long context would otherwise cause out-of-memory with Full Attention.

You may also enjoy Mistral AI’s Voxtral, How Kimi, Cursor, and Chroma Train Agentic Models with RL, a paper on Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Gemini 3.1 Flash Live: Making audio AI more natural and reliable

Google launches Gemini 3.1 Flash Live, its highest-quality real-time audio model, scoring 90.8% on ComplexFuncBench Audio and 36.1% on AudioMultiChallenge

Speaking of Voxtral

Mistral launches Voxtral TTS, a 4B-parameter multilingual text-to-speech model supporting 9 languages with 70ms latency, voice cloning from 3-second samples, and more.

SAM 3.1: Faster and More Accessible Real-Time Video Detection and Tracking With Multiplexing and Global Reasoning

Meta updates SAM 3 to SAM 3.1, adding object multiplexing to double video processing speed to 32 FPS on a single H100 for its open-source text-prompted segmentation and tracking model.

Cohere Transcribe: state-of-the-art speech recognition

Cohere launches Transcribe, a 2B-parameter open-source ASR model that tops the HuggingFace Open ASR Leaderboard with a 5.42% average word error rate across 14 languages,

Granola raises $125M at $1.5B valuation for its AI note-taking app

Granola raises $125M Series C at a $1.5B valuation led by Index Ventures, following a quarter of 250% revenue growth, with plans to expand its AI meeting notes app toward agentic task automation.

MLOps/LLMOps

Deploying Disaggregated LLM Inference Workloads on Kubernetes

A technical guide to deploying disaggregated LLM inference (separate prefill, decode, and router services) on Kubernetes using NVIDIA Grove, KAI Scheduler, and NVIDIA Dynamo.

Learning

Best Embedding Model for RAG 2026: 10 Models Compared

A practical benchmarking guide comparing 10 embedding models across four production-critical RAG dimensions — cross-modal, cross-lingual, long-document retrieval, and MRL compression

How Kimi, Cursor, and Chroma Train Agentic Models with RL

A technical synthesis of three recent agentic RL training reports — Kimi K2.5, Cursor Composer 2, and Chroma Context-1 — distilling shared patterns around production-environment training, context management, and reward design.

Multimodal Embeddings and RAG: A Practical Guide

A practical guide to multimodal embeddings and RAG covering the core theory (contrastive learning, modality gap, MRL), three concrete build patterns (audio, PDF, video), and when multimodal actually outperforms text-only pipelines.

Five techniques to reach the efficient frontier of LLM inference

A practical guide to LLM inference optimization framed around the “efficient frontier” concept — five techniques that move production systems toward the latency/throughput Pareto boundary without additional hardware spend.

Libraries & Code

comet-ml/opik

An open-source LLM evaluation tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

katanemo/plano

Plano is an AI-native proxy and data plane for agentic apps — with built-in orchestration, safety, observability, and smart LLM routing so you stay focused on your agents core logic.

Papers & Publications

Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis

Abstract:

Unified multimodal models provide a natural and promising architecture for understanding diverse and complex real-world knowledge while generating high-quality images. However, they still rely primarily on frozen parametric knowledge, which makes them struggle with real-world image generation involving long-tail and knowledge-intensive concepts. Inspired by the broad success of agents on real-world tasks, we explore agentic modeling to address this limitation. Specifically, we present Unify-Agent, a unified multimodal agent for world-grounded image synthesis, which reframes image generation as an agentic pipeline consisting of prompt understanding, multimodal evidence searching, grounded recaptioning, and final synthesis. To train our model, we construct a tailored multimodal data pipeline and curate 143K high-quality agent trajectories for world-grounded image synthesis, enabling effective supervision over the full agentic generation process. We further introduce FactIP, a benchmark covering 12 categories of culturally significant and long-tail factual concepts that explicitly requires external knowledge grounding. Extensive experiments show that our proposed Unify-Agent substantially improves over its base unified model across diverse benchmarks and real world generation tasks, while approaching the world knowledge capabilities of the strongest closed-source models. As an early exploration of agent-based modeling for world-grounded image synthesis, our work highlights the value of tightly coupling reasoning, searching, and generation for reliable open-world agentic image synthesis.

IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse

Abstract:

Long-context agentic workflows have emerged as a defining use case for large language models, making attention efficiency critical for both inference speed and serving cost. Sparse attention addresses this challenge effectively, and DeepSeek Sparse Attention (DSA) is a representative production-grade solution: a lightweight lightning indexer selects the top-k most relevant tokens per query, reducing core attention from O(L2) to O(Lk). However, the indexer itself retains O(L2) complexity and must run independently at every layer, despite the fact that the resulting top-k selections are highly similar across consecutive layers. We present IndexCache, which exploits this cross-layer redundancy by partitioning layers into a small set of Full layers that run their own indexers and a majority of Shared layers that simply reuse the nearest Full layer’s top-k indices. We propose two complementary approaches to determine and optimize this configuration. Training-free IndexCache applies a greedy search algorithm that selects which layers to retain indexers by directly minimizing language modeling loss on a calibration set, requiring no weight updates. Training-aware IndexCache introduces a multi-layer distillation loss that trains each retained indexer against the averaged attention distributions of all layers it serves, enabling even simple interleaved patterns to match full-indexer accuracy. Experimental results on a 30B DSA model show that IndexCache can remove 75% of indexer computations with negligible quality degradation, achieving up to 1.82× prefill speedup and 1.48× decode speedup compared to standard DSA. These positive results are further confirmed by our preliminary experiments on the production-scale GLM-5 model.

You may also enjoy What 81,000 people want from AI \ Anthropic, Evaluating agentic search in OpenSearch, a paper on OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Introducing Composer 2 · Cursor

Cursor launches Composer 2, a frontier-level coding model trained via continued pretraining and long-horizon RL that scores 61.3 on CursorBench and 73.7 on SWE-bench Multilingual

What 81,000 people want from AI \ Anthropic

Anthropic’s largest-ever qualitative study — 80,508 Claude users across 159 countries and 70 languages — reveals what people want from AI, what they’ve already gotten, and what they fear.

Lyria 3 Pro: Create longer tracks in more Google products

Google launches Lyria 3 Pro, an upgraded music generation model that produces tracks up to 3 minutes with structural song awareness (intros, verses, choruses, bridges).

MolmoWeb: An open agent for automating web tasks

Allen AI releases MolmoWeb, a fully open visual web agent built on Molmo 2 that scores 78.2% on WebVoyager and 73.7% on SWE-bench Multilingual, outperforming GPT-4o-based agents while releasing all weights, training data, and evaluation tools.

OpenAI to acquire Astral

OpenAI acquires Astral — maker of Python developer tools uv, Ruff, and ty used by millions of developers — to deepen its Codex ecosystem.

MLOps/LLMOps

Building an MCP Ecosystem at Pinterest

Pinterest Engineering details how they scaled MCP from concept to a production ecosystem of domain-specific servers — Presto, Spark, Knowledge — with a central registry, two-layer auth, and 66,000 monthly invocations saving an estimated 7,000 engineer-hours per month.

Run cloud agents in your own infrastructure

Cursor launches self-hosted cloud agents GA, keeping code and tool execution entirely within enterprise infrastructure while Cursor handles orchestration and inference.

Learning

TurboQuant: Redefining AI efficiency with extreme compression

Google Research releases TurboQuant, a KV cache quantization method that achieves 6x+ memory reduction to 3 bits with zero accuracy loss and 8x attention speedup on H100s.

Evaluating agentic search in OpenSearch

A technical deep-dive on how OpenSearch benchmarked its agentic search feature across search relevance (BEIR and BRIGHT datasets) and query execution accuracy (Spider dataset), powered by Claude Opus 4.6.

Scaling Karpathy’s Autoresearch: What Happens When the Agent Gets a GPU Cluster

A technical blog post on how SkyPilot scaled Karpathy’s autoresearch agent from 1 to 16 GPUs, enabling ~910 experiments in 8 hours.

How Anthropic’s Claude Thinks - ByteByteGo Newsletter

ByteByteGo breaks down Anthropic’s interpretability research into six concrete findings about how Claude actually thinks — from parallel math strategies to ahead-of-time poetry planning to a default-refusal circuit that misfires into hallucinations.

A Visual Guide to Attention Variants in Modern LLMs

A visual reference guide mapping seven attention variants — MHA, GQA, MLA, SWA, DeepSeek Sparse Attention, Gated Attention, and hybrid architectures — across the open-weight models currently using them in production.

Fast regex search: indexing text for agent tools

A technical deep-dive on how Cursor built a local sparse n-gram index to replace ripgrep for agent search — eliminating 15+ second grep latency in large monorepos by narrowing regex matches to a pre-filtered candidate set before full scanning.

Libraries & Code

comet-ml/opik

An open-source LLM evaluation tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

openai/teen-safety-policy-pack

A set of prompt-based safety policies designed to create age-appropriate protections for teens.

Papers & Publications

OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis

Abstract:

Training deep research agents requires long-horizon trajectories that interleave search, evidence aggregation, and multi-step reasoning. However, existing data collection pipelines typically rely on proprietary web APIs, making large-scale trajectory synthesis costly, unstable, and difficult to reproduce. We present OpenResearcher, a reproducible pipeline that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis and executes the search-and-browse loop entirely offline using three explicit browser primitives: search, open, and find, over a 15M-document corpus. Using GPT-OSS-120B as the teacher model, we synthesize over 97K trajectories, including a substantial long-horizon tail with 100+ tool calls. Supervised fine-tuning a 30B-A3B backbone on these trajectories achieves 54.8\% accuracy on BrowseComp-Plus, a +34.0 point improvement over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. Because the environment is offline and fully instrumented, it also enables controlled analysis, where our study reveals practical insights into deep research pipeline design, including data filtering strategies, agent configuration choices, and how retrieval success relates to final answer accuracy.

Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation

Abstract:

Agent memory systems often adopt the standard Retrieval-Augmented Generation (RAG) pipeline, yet its underlying assumptions differ in this setting. RAG targets large, heterogeneous corpora where retrieved passages are diverse, whereas agent memory is a bounded, coherent dialogue stream with highly correlated spans that are often duplicates. Under this shift, fixed top-k similarity retrieval tends to return redundant context, and post-hoc pruning can delete temporally linked prerequisites needed for correct reasoning. We argue retrieval should move beyond similarity matching and instead operate over latent components, following decoupling to aggregation: disentangle memories into semantic components, organise them into a hierarchy, and use this structure to drive retrieval. We propose xMemory, which builds a hierarchy of intact units and maintains a searchable yet faithful high-level node organisation via a sparsity--semantics objective that guides memory split and merge. At inference, xMemory retrieves top-down, selecting a compact, diverse set of themes and semantics for multi-fact queries, and expanding to episodes and raw messages only when it reduces the reader’s uncertainty. Experiments on LoCoMo and PerLTQA across the three latest LLMs show consistent gains in answer quality and token efficiency.

You may also enjoy Introducing Mistral Small 4, State of RL for reasoning LLMs, a paper on Data Agents: Levels, State of the Art, and Open Problems, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Mamba-3

Together AI releases Mamba-3, an inference-first state space model that outperforms Mamba-2, Gated DeltaNet, and Transformer-based Llama-3.2-1B on end-to-end latency at the 1.5B scale.

Introducing Mistral Small 4

Mistral releases Small 4 — an open-source, 119B-parameter MoE model unifying reasoning, multimodal, and coding capabilities, delivering 40% lower latency and 3x higher throughput than its predecessor.

Claude builds interactive visuals right in your conversation

Anthropic launches inline interactive charts, diagrams, and visualizations in Claude chat — available in beta across all plan types.

Measuring Progress Towards AGI: A Cognitive Framework

Google DeepMind releases a cognitive taxonomy paper proposing 10 human-grounded abilities to measure AGI progress, paired with a $200,000 Kaggle hackathon to crowdsource the missing benchmarks.

Gumloop reels in $50M for its AI automation platform

Gumloop raises $50M Series B led by Benchmark — with participation from Shopify Ventures and Y Combinator — bringing total funding to $70M for its no-code, drag-and-drop AI agent automation platform.

Okta unveils new framework to manage AI agents and upcoming Okta for AI Agents platform

Okta unveils a security blueprint for the agentic enterprise and announces its “Okta for AI Agents” platform – treating AI agents as governed, non-human identities with centralized access control and a kill switch for rogue agents.

MLOps/LLMOps

Agent-native Architectures: How to Build Apps After Code Ends

A technical guide on building agent-native applications — software architectures where agents are first-class citizens, using atomic tools and outcome-driven loops instead of hardcoded workflows.

Learning

State of RL for reasoning LLMs

A technical deep-dive surveying the evolution of reinforcement learning algorithms for reasoning LLMs (2024–2026), tracing the lineage from REINFORCE and PPO through GRPO and eight successor methods

Many SWE-bench-Passing PRs Would Not Be Merged into Main - METR

METR researchers found that roughly half of SWE-bench Verified PRs that pass the automated grader would not actually be merged by real repo maintainers, with automated grader scores averaging 24 percentage points higher than maintainer merge rates.

LumberChunker: Long-Form Narrative Document Segmentation

An article about LumberChunker, a RAG chunking method that uses an LLM to detect semantic boundaries in long-form narrative documents, achieving DCG@20 of 62.1% on the GutenQA benchmark — outperforming all fixed-size and recursive baselines.

VAGEN: Teaching Vision-Language Models to Build World Models Through Reinforcement Learning

A Stanford AI Lab research blog post about VAGEN, a reinforcement learning framework that trains 3B-parameter VLM agents to build internal world models via structured state estimation and transition predictions.

Libraries & Code

comet-ml/opik

An open-source LLM evaluation tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

tobi/qmd

A mini CLI search engine for your docs, knowledge bases, meeting notes, and more.

Papers & Publications

Attention Residuals

Abstract:

Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with depth, progressively diluting each layer’s contribution. We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs, allowing each layer to selectively aggregate earlier representations with learned, input-dependent weights. To address the memory and communication overhead of attending over all preceding layer outputs for large-scale model training, we introduce Block AttnRes, which partitions layers into blocks and attends over block-level representations, reducing the memory footprint while preserving most of the gains of full AttnRes. Combined with cache-based pipeline communication and a two-phase computation strategy, Block AttnRes becomes a practical drop-in replacement for standard residual connections with minimal overhead.

Scaling law experiments confirm that the improvement is consistent across model sizes, and ablations validate the benefit of content-dependent depth-wise selection. We further integrate AttnRes into the Kimi Linear architecture (48B total / 3B activated parameters) and pre-train on 1.4T tokens, where AttnRes mitigates PreNorm dilution, yielding more uniform output magnitudes and gradient distribution across depth, and improves downstream performance across all evaluated tasks.

Data Agents: Levels, State of the Art, and Open Problems

Abstract:

Data agents are an emerging paradigm that leverages large language models (LLMs) and tool-using agents to automate data management, preparation, and analysis tasks. However, the term “data agent” is currently used inconsistently, conflating simple query responsive assistants with aspirational fully autonomous “data scientists”. This ambiguity blurs capability boundaries and accountability, making it difficult for users, system builders, and regulators to reason about what a “data agent” can and cannot do.

In this tutorial, we propose the first hierarchical taxonomy of data agents from Level 0 (L0, no autonomy) to Level 5 (L5, full autonomy). Building on this taxonomy, we will introduce a lifecycleand level-driven view of data agents. We will (1) present the L0-L5 taxonomy and the key evolutionary leaps that separate simple assistants from truly autonomous data agents, (2) review representative L0-L2 systems across data management, preparation, and analysis, (3) highlight emerging Proto-L3 systems that strive to autonomously orchestrate end-to-end data workflows to tackle diverse and comprehensive data-related tasks under supervision, and (4) discuss forward-looking research challenges towards proactive (L4) and generative (L5) data agents. We aim to offer both a practical map of today’s systems and a research roadmap for the next decade of data-agent development.

AI Can Learn Scientific Taste

Abstract:

Great scientists have strong judgement and foresight, closely tied to what we call scientific taste. Here, we use the term to refer to the capacity to judge and propose research ideas with high potential impact. However, most relative research focuses on improving an AI scientist’s executive capability, while enhancing an AI’s scientific taste remains underexplored. In this work, we propose Reinforcement Learning from Community Feedback (RLCF), a training paradigm that uses large-scale community signals as supervision, and formulate scientific taste learning as a preference modeling and alignment problem. For preference modeling, we train Scientific Judge on 700K field- and time-matched pairs of high- vs. low-citation papers to judge ideas. For preference alignment, using Scientific Judge as a reward model, we train a policy model, Scientific Thinker, to propose research ideas with high potential impact. Experiments show Scientific Judge outperforms SOTA LLMs (e.g., GPT-5.2, Gemini 3 Pro) and generalizes to future-year test, unseen fields, and peer-review preference. Furthermore, Scientific Thinker proposes research ideas with higher potential impact than baselines. Our findings show that AI can learn scientific taste, marking a key step toward reaching human-level AI scientists.

You may also enjoy GPT-5.4, a paper on DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints, and more!

As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Gemini Embedding 2: Our first natively multimodal embedding model

Google launches Gemini Embedding 2, its first natively multimodal embedding model unifying text, images, video, audio, and documents into a single semantic space.

OpenAI to acquire Promptfoo

OpenAI is acquiring Promptfoo — an AI security platform used by 25%+ of Fortune 500 companies — to embed red-teaming, jailbreak detection, and agentic risk evaluation natively into its enterprise Frontier platform.

Introducing GPT-5.4 | OpenAI

OpenAI launches GPT-5.4 with a 1M-token context, new Tool Search API, and record scores on coding and knowledge-work benchmarks — its most capable frontier model for professional and agentic use.

Google upgrades Gemini for Workspace allowing it to pull data from multiple apps to create Docs, Sheets, Slides and more

Google lets Gemini generate fully-formed Docs, Sheets, and Slides by pulling from Gmail, Drive, and Chat — turning Workspace into a single-prompt content creation engine.

Yann LeCun’s AMI Labs raises $1.03B to build world models | TechCrunch

Yann LeCun’s AMI Labs raises $1.03B at a $3.5B valuation to build JEPA-based world models — AI that learns from reality rather than language — with NVIDIA, Samsung, and Eric Schmidt among backers.

MLOps/LLMOps

Native Observability & Alerts for Your OpenClaw with Opik

A blog post announcing opik-openclaw, a native OpenClaw plugin from Comet that adds full-stack observability — tracing every LLM call, tool execution, token cost, and sub-agent delegation — to address the visibility gap in autonomous agent workflows.

Learning

Improving instruction hierarchy in frontier LLMs

A technical research post about OpenAI’s IH-Challenge — an RL training dataset that teaches models a strict trust hierarchy (System > Developer > User > Tool) to resist prompt injection, jailbreaks, and instruction conflicts.

Code Concepts: A Large-Scale Synthetic Dataset Generated from Programming Concept Seeds

A technical blog post about NVIDIA’s concept-driven synthetic data pipeline that generated 15M Python programming problems, yielding a 6-point HumanEval gain (73→79) when included in Nemotron-Nano-v3 pretraining.

Practical Guide to Evaluating and Testing Agent Skills

A practical guide about building lightweight eval harnesses for agent skills, walking through how to define success criteria, construct prompt sets, and iterate — illustrated by taking a Gemini Interactions API skill from 66.7% to 100% pass rate.

Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?

A Stanford benchmark revealing that frontier models (GPT-5.2, Gemini-3 Pro, Claude 4.5 Sonnet) all fail to build accurate, revisable cognitive maps during active spatial exploration — humans consistently outperform all of them.

Libraries & Code

comet-ml/opik

An open-source LLM evaluation tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.

alibaba/page-agent

JavaScript in-page GUI agent. Control web interfaces with natural language.

Papers & Publications

LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory

Abstract:

Feedforward geometric foundation models achieve strong short-window reconstruction, yet scaling them to minutes-long videos is bottlenecked by quadratic attention complexity or limited effective memory in recurrent designs. We present LoGeR (Long-context Geometric Reconstruction), a novel architecture that scales dense 3D reconstruction to extremely long sequences without post-optimization. LoGeR processes video streams in chunks, leveraging strong bidirectional priors for high-fidelity intra-chunk reasoning. To manage the critical challenge of coherence across chunk boundaries, we propose a learning-based hybrid memory module. This dual-component system combines a parametric Test-Time Training (TTT) memory to anchor the global coordinate frame and prevent scale drift, alongside a non-parametric Sliding Window Attention (SWA) mechanism to preserve uncompressed context for high-precision adjacent alignment. Remarkably, this memory architecture enables LoGeR to be trained on sequences of 128 frames, and generalize up to thousands of frames during inference. Evaluated across standard benchmarks and a newly repurposed VBR dataset with sequences of up to 19k frames, LoGeR substantially outperforms prior state-of-the-art feedforward methods--reducing ATE on KITTI by over 74%--and achieves robust, globally consistent reconstruction over unprecedented horizons.

DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints

Abstract:

While agent evaluation has shifted toward long-horizon tasks, most benchmarks still emphasize local, step-level reasoning rather than the global constrained optimization (e.g., time and financial budgets) that demands genuine planning ability. Meanwhile, existing LLM planning benchmarks underrepresent the active information gathering and fine-grained local constraints typical of real-world settings. To address this, we introduce DeepPlanning, a challenging benchmark for practical long-horizon agent planning. It features multi-day travel planning and multi-product shopping tasks that require proactive information acquisition, local constrained reasoning, and global constrained optimization. Evaluations on DeepPlanning show that even frontier agentic LLMs struggle with these problems, highlighting the importance of reliable explicit reasoning patterns and parallel tool use for achieving better effectiveness-efficiency trade-offs. Error analysis further points to promising directions for improving agentic LLMs over long planning horizons. We open-source the code and data to support future research.