This week in deep learning, we bring you FLUX Upscale: 2K and 4K for Video, Empty shelves or lost keys? Recall is the bottleneck for parametric factuality and a paper on VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?.
You may also enjoy Pacing model development in an era of cyber-critical capabilities, Introducing the Conceptual Reasoning Index, a paper on Zetta: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical 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
FLUX Upscale: 2K and 4K for Video
Black Forest Labs launches FLUX Upscale, a standalone tool and API endpoint that regenerates video at up to native 4K while repairing generation artifacts, offered in Precise and Creative modes.
Pacing model development in an era of cyber-critical capabilities
OpenAI paused its largest frontier RL run and slowed scaling after preliminary evidence that its upcoming model Astra may cross the “Critical” cybersecurity threshold, hardening research-environment security, monitoring, and alignment before proceeding.
Stripe buys AI model router OpenRouter in reported $7.5B deal
Stripe agreed to acquire AI model-routing startup OpenRouter — a single-endpoint gateway to 400+ models from 80+ providers handling 10T+ daily tokens — in a reported ~$7.5B+ deal.
Offering Zero Data Retention for frontier models
OpenAI reaffirmed Zero Data Retention for eligible API customers and previewed Private Safety Processing.
MIT CSAIL researchers identified “attribution decay” — showing that in large-scale generative models, removing any single training image often leaves outputs unchanged, meaning many AI images can’t be attributed to any specific source.
MLOps/LLMOps/AgentOps
What is AI Observability? A Complete Guide to Debugging and Monitoring Modern AI Systems at Scale
A guide on why AI observability must track system behavior, not just health — capturing telemetry across four layers (app, agent, model, retrieval) and adding evaluations as a fourth signal alongside logs, metrics, and traces to catch semantic failures that leave infra dashboards green.
The builder’s guide to GPT‑5.6
OpenAI’s builder’s guide argues GPT-5.6 collapses agent economics by pairing lower-reasoning-effort accuracy with new Responses API primitives, letting startups match frontier quality at a fraction of the cost.
Building Federated Multimodal AI Workflows with NVIDIA FLARE
A technical guide showing how NVIDIA FLARE federates multimodal VLM training, where FedUMM’s adapter-only approach cut per-client communication from 28.6 GB to 0.094 GB per round while holding ~97% of centralized performance.
Learning
The final part of the Observable Job Agent series brings the agent to life with voice. See how LangGraph, ElevenLabs, FastAPI, and Opik come together to build a voice agent that can search, rank, and tailor job applications.
Empty shelves or lost keys? Recall is the bottleneck for parametric factuality
Google Research’s knowledge profiling framework shows factual errors in frontier LLMs are overwhelmingly recall failures, not encoding failures — the facts are stored but inaccessible, shifting the factuality bottleneck from acquisition to utilization.
Introducing the Conceptual Reasoning Index
Redwood Research and Anthropic introduce the Conceptual Reasoning Index, aggregating three benchmarks that measure argumentation on unverifiable questions, where top scorer Opus 5 hits 73.6 against an estimated ceiling of 91.
Thinking of ACE? We Can Do It with Fewer Tokens
IBM Research’s ALTK-Evolve matches or beats ACE on AppWorld agent tasks by calibrating how much learned “memory” reaches the model per task rather than injecting the full playbook every step.
Hyperlaw: AI Will Change How Law Evolves
An analytical article about “hyperlaw” — how AI-driven legal productivity gains of 12%–130% cut both contracting and litigation costs, shrinking settlement ranges and accelerating precedent change fastest in contract-free domains.
Recovering Encrypted LLM Reasoning Traces
A security research post reproducing an attack that replays encrypted LLM reasoning blobs across sessions, accounts, and models to recover hidden traces — the source paper extracted 367 PII items and 182 credentials from 315,320 public blocks.
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.
Papers & Publications
VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?
Abstract:
Constructing an interactive 3D open world from a user query is important. However, existing methods are primarily evaluated on idealized, simple queries, making it difficult to systematically analyze and compare how multimodal agents understand user intent, use 3D tools, and reason over textual and visual 3D world information. To this end, we propose VibeWorlding, a unified framework for benchmarking and training vibe worlding agents: a multimodal agent that can autonomously infer user intent, plan scene layout, invoke 3D tools, and reflect on the multimodal feedback in a multi-turn agent-environment interaction process. To achieve this, we first build VWE-BENCH, a benchmark of 2,616 high-quality 3D assets, 323 human-annotated seed 3D worlds, and 6,828 reverse-synthesized multimodal user queries, split into verified queries with ground-truth and unverified queries with carefully designed rubrics. Moreover, we develop VibeWorlding-Gym, a joint multimodal RL post-training framework that integrates (1) a sandbox environment unifying asset retrieval, editing, and image rendering as MCP tools, and (2) a rubric-based verifier that combines physical feasibility and intent fulfillment verification, supporting both fair model evaluation and scalable multimodal RL reward service. Our experiments show that current frontier MLLMs are far from solving the vibe worlding agent task, with even GPT-5.5 and Qwen3.8-Max reaching below 60% success rate, and trace the bottleneck to precise 3D world editing. We further find that RL training can ease this weakness and enable open-source MLLMs to even surpass closed-source frontiers: our VibeWorlder-8B is comparable to frontier MLLMs, while our flagship VibeWorlder-30B-A3B attains the best overall Pass@1 among all evaluated models.
Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
Abstract:
Embodied agents are increasingly used to close the gap left by end-to-end policy models. Yet the agentic path has not realized closed-loop learning in physical execution: existing harnesses remain largely open-loop, following fixed skills during rollout and reflecting only after an episode completes. Such post-hoc reflection cannot govern execution as it unfolds, because physical interaction requires decisions to track rapidly changing robot-environment states at a frequency beyond today’s large agentic models. We present Zetta, a closed-loop embodied harness that evolves code-based runtime critics and recovery skills online while keeping the base policy frozen. Through three timescale-separated loops, Zetta provides action-frequency governance, rollout-level critic-recovery proposal, and validation-gated skill updates. Together with Z-Infra, a rollout infrastructure decoupling agent logic from heterogeneous execution resources, Zetta achieves state-of-the-art success on LIBERO-Pro and RoboCasa under our current rollout budget, reaching 90.8% and 93.6%, with an 11.1x inference speedup; success continues to scale with self-exploration experience; learned skills transfer zero-shot, and clear robotic “Aha Moments” emerge. These results show that closed-loop harness self-evolution opens a scaling path for reliable physical intelligence.
