Deep Learning Weekly: Issue 467
Qwen3.8-Max, A verifiable autonomous research framework via Chain-of-Evidence, a paper on From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement, a
This week in deep learning, we bring you Qwen3.8-Max, Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence and a paper on From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement.
You may also enjoy Shieldstral, Even after R&D is automated, parallelization constraints could delay a technological singularity, a paper on PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal 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
Qwen: Qwen3.8-Max: A New Bar for Coding and Cowork
Alibaba releases Qwen3.8-Max, a 2.4T-parameter multimodal MoE with 95B active parameters and 1M-token context, leading Terminal-Bench 2.1 at 86.6 and marking the first Max-class Qwen to go open-weights.
Mistral releases Shieldstral, a 3B Apache 2.0 multimodal safety classifier that takes plain-language policies at inference time and matches guard models up to 7x its size while running on a single 16GB GPU.
Gemini Robotics 2 brings whole body intelligence to robots
Google DeepMind launches Gemini Robotics 2, a three-model family bringing full humanoid whole-body control, 22-DoF multi-finger dexterity, and multi-robot collaboration, with on-device adaptation to new embodiments in hours from under 200 examples.
Nscale buys AI infrastructure optimization startup Anyscale for reported $1.65B
Data center builder Nscale acquires Anyscale, the company behind the open-source Ray cluster-optimization framework, for a reported $1.65B to complete a vertically integrated AI cloud spanning power, compute, and software.
Obsidian Security raises $85M as AI agents create cybersecurity’s next major attack surface
Obsidian Security raises $85M Series D at a $1.1B valuation to govern non-human AI agent identities, with 65% of its enterprise customers already granting agents access to third-party SaaS data.
MLOps/LLMOps/AgentOps
86% of Your Claude Code Bill Has Nothing to Do With Your Prompts
A look into where Claude Code tokens actually go, revealing that prompts make up just a fraction of total spend while context replay, MCP schemas, and tool outputs quietly dominate the bill.
Investigating three real-world incidents in our cybersecurity evaluations
Anthropic discloses that a review of 141,006 cybersecurity evaluation runs surfaced three incidents where Claude models breached the real production infrastructure of three organizations via a misconfigured third-party eval environment.
FBTriton Infra: Upstream Ingestion, Hierarchical Validation, Ideals vs Realities
An engineering post on how Meta keeps its custom Triton fork in sync with upstream, using AI agents to sort incoming commits by risk and a three-tier testing system to catch regressions cheaply.
Learning
The Bare-Bones Coding Agent Loop
A hands-on guide to building a coding agent from scratch, unpacking the agent loop, tool orchestration, and harness design that power systems like Claude Code and Codex.
Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence
Google Research introduces the Science One Framework and CoE Audit, an autonomous research agent that hits zero phantom references and fully verifiable scores against baselines, hallucinating up to 21% of citations.
SOTA alignment assessments don’t strongly update us against misalignment
A critical analysis arguing that frontier labs’ alignment assessments provide far weaker evidence against misalignment than their system cards claim, because the covert-capability evals underpinning those reliability arguments are themselves unreliable.
Even after R&D is automated, parallelization constraints could delay a technological singularity
An economic analysis arguing that intelligence-explosion models omit a key parameter — “parallelization technology,” the ceiling on how many researchers can work usefully at once — which could slow or flatten takeoff even after AI R&D is fully automated.
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.
The long-horizon computer-use harness. Run AI agents across desktop apps and the CLI for extended periods while preserving task state and making reliable progress on complex workflows.
Papers & Publications
Abstract:
Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be deterministically verifiable. Open-ended tasks instead often rely on human preferences, reward models, or LLM-based judges, introducing evaluation bias, judge capability bottlenecks, and additional inference costs. Drawing on the principle of self-supervised learning, which constructs pretext tasks to derive supervision from the data itself, we propose Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a task-transformation-based training paradigm for extending RLVR to open-ended tasks. RLSVR transforms open-ended tasks into verifiable proxy environments whose internal rules and interaction outcomes automatically generate reward signals. We instantiate RLSVR with SpyRL, a Self-PlaY Reinforcement Learning method inspired by social deduction game Who Is the Spy?. Agents receive asymmetric information, complete the same target task, and vote to identify a designated spy. Because the spy identity is predetermined, voting outcomes provide fully verifiable rewards, while successful identification remains closely related to output quality. Experiments on text summarization, creative writing, and mathematical reasoning show that SpyRL outperforms existing self-improvement methods on non-verifiable tasks and yields consistent gains on verifiable reasoning tasks. These results demonstrate that task transformation can extend scalable RLVR-based self-improvement beyond inherently verifiable domains.
PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
Abstract:
Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained experience actually improves them over time has not been systematically tested. We introduce PAST-Bench, a benchmark designed to isolate this question. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that turn retained experience on and off. It spans 26 scenarios and 204 episodes across memory, procedural reuse, information gathering, and update. We report both later-task gains and whether those gains follow the intended save, retrieve, and update pathway. Across seven base models and four agent frameworks, improvement is real but uneven across capabilities. Agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended pathway. Guided by these findings, we develop Hermes+, which extends Hermes with five targeted interventions across stages of the agent loop. Hermes+ raises the average gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced, although the effect remains capability- and model-dependent.
 | A guest post by
|