Deep Learning Weekly: Issue 468
Introducing Grok 4.6, TutorMoments: Do AI tutors know when to help and when to hold back?, a paper on BDH-CQ: In-Context Learning with Recurrent Latent Reasoning, and many more!
This week in deep learning, we bring you Introducing Grok 4.6, TutorMoments: Do AI tutors know when to help and when to hold back? and a paper on BDH-CQ: In-Context Learning with Recurrent Latent Reasoning.
You may also enjoy Meta’s Muse Glimmer, One in five US workers now delegates tasks to AI instead of other humans, a paper on SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring, 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
xAI releases Grok 4.6, tuned for long-running agents and interactive builds, matching GPT-5.6 Sol at 61 on the AA Intelligence Index and leading on GDPval-AA (1753) and AA-Briefcase (1577) at unchanged $2/$6 pricing.
Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device
Meta open-sources Muse Glimmer, a 30B agentic model distilled from Muse Spark and quantized to under 20GB so it runs always-on locally, with speculative decoding delivering up to 3.1x faster decode.
Expanding Daybreak as the Cyber Defense Window Narrows
OpenAI expands Daybreak into Blue and Red access tiers and launches GPT-5.6-Cyber, which completes 95% of advanced offensive-security requests versus 1.5% for the guarded base model.
AWS Continuum integrates with OpenAI Codex and Anthropic Claude Code in major AI security push
AWS embeds its Continuum vulnerability platform directly into Claude Code, OpenAI Codex, and Kiro, positioning itself as a model-neutral security control plane that absorbs token costs behind a single price.
Anthropic says it will watermark text generated by its AI models
Anthropic will watermark all Claude-generated text and files at the model level to comply with the EU AI Act’s Transparency Code, with marks persisting through copy-paste and some editing.
Putting sign language AI into users’ hands
Google DeepMind ships SL2T, a sign-language-to-text model trained on 100,000+ hours across 50+ sign languages, hitting 70 BLEURT zero-shot and powering ASL dictation in Gboard and Live Transcribe on Pixel 11.
LTX-2.5 can generate a 10-second AI video from an image in just 6.8 seconds on Nvidia superchips
LTX releases LTX-2.5, a 22B open-weights video world model that renders a 10-second clip in 6.8 seconds on dual GB200s — roughly 7.6x faster than the quickest closed API.
MLOps/LLMOps/AgentOps
Build your own Job Agent - Part 3
Part 3 of The Observable Job Agent series continues with a practical look at LLMOps in action: using Opik traces, evals, prompt optimization, and Ollie to diagnose failures and improve agent performance.
Learning
TutorMoments: Do AI tutors know when to help and when to hold back?
Ai2 releases TutorMoments, a replay-based eval built on 462 real math tutoring transcripts and 1,500+ teacher-annotated decision points, finding LLMs systematically over-help rather than pushing students toward productive struggle.
One in five US workers now delegates tasks to AI instead of other humans
Epoch AI survey of 1,106 US workers finds 20% now delegate to AI work once handed to a coworker or contractor, with 66% of AI outputs accepted unchanged or with minor edits.
Scaling Test-Time Compute in Search Mode
Weaviate adds medium/high/ultrahigh effort tiers to Query Agent Search Mode, scaling test-time compute on query decomposition and reranking to lift BRIGHT Biology nDCG@10 from 13.0 to 57.5 over hybrid search.
A detailed blog post about how Anthropic rewrote Claude Fable 5’s biology classifier constitution to cut false-positive fallbacks by ~85%, illustrating the tradeoff between launching broad safeguards early and refining precision over time.
Libraries & Code
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.
Switchyard is a Rust proxy and library for LLM traffic. It routes requests across providers, translates between OpenAI and Anthropic APIs, records operational metrics, and provides typed, composable routing algorithms.
Papers & Publications
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
Abstract:
We introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. Inputs presented at inference time continuously update the model’s recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult. A 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of $0.0007 per task. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.
SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring
Abstract:
As AI coding agents take on increasingly complex, long-horizon software engineering tasks, existing benchmarks are rapidly saturating and their evaluation quality has come under serious scrutiny: a recent audit found that nearly 60% of unsolved SWE-bench Verified instances contain flawed tests -- either overly narrow tests that reject correct solutions or overly broad tests that check unstated requirements -- and that frontier models can verbatim reproduce gold patches from training data. Code refactoring, which requires coordinated, behavior-preserving changes across many files, offers a substantially harder and more realistic test of agent capability, yet remains underserved by current benchmarks. We introduce SWE-Bench ProMax, an expert-curated, multilingual code refactoring benchmark of 170 instances drawn from real commits across seven programming languages (Python, Java, TypeScript, Go, C, C++, and Rust). Every instance undergoes rigorous, multi-stage curation that directly addresses the quality problems identified in prior benchmarks: issue descriptions are rewritten from scratch to provide precise, unambiguous specifications, and test suites are manually reviewed to remove overly narrow and overly broad tests. Tasks with insufficient complexity or limited cross-file scope are filtered out, yielding a benchmark of challenging, large-scale refactoring tasks that average 11.4 modified files and 261.6 lines of code per instance, substantially exceeding the scale of existing benchmarks. Experiments with frontier models under two agent scaffolds show that the best model achieves only 41.2% resolve rate, confirming that SWE-Bench ProMax presents a meaningful and unsaturated challenge for current AI coding agents.
 | A guest post by
|