What they're sharing
Articles & links
SPIRAL: Learning to Search and Aggregate Jubayer Ibn Hamid, Ifdita Hasan Orney, Michael Y. Li, Omar Shaikh, Yoonho Lee, Dorsa Sadigh, Chelsea Finn, Noah Goodman https://t.co/CRBpj1Mjhk [𝚌𝚜.𝙰𝙸] https://t.co/kVEHyMHKpK
SPIRAL: Learning to Search and Aggregate arxiv.org
AI Weekly's analysis
→
- SPIRAL co-trains three reasoning primitives in one RL framework: sequential chain-of-thought, parallel sampling of traces, and learned aggregation of those traces.
- The paper reports outperforming GRPO by up to 11× scaling efficiency and 15% higher performance when all three compute primitives are scaled.
- Training uses set reinforcement learning to make parallel traces collectively useful, plus standard RL to train the aggregation step itself.
Read full analysis →
Autodata: An agentic data scientist to create high quality synthetic data Ilia Kulikov, Chenxi Whitehouse, Tianhao Wu, Yixin Nie, Swarnadeep Saha, Eryk Helenowski, Weizhe Yuan, Olga Golovneva, Jack Lanchantin, … https://t.co/iSchw5CkfT [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻 𝚌𝚜.𝙻𝙶] https://t.co/fC2KJEmmyE
Autodata: An agentic data scientist to create high quality synthetic data arxiv.org
AI Weekly's analysis
→
- Meta researchers introduce Autodata, a method that casts an AI agent as a data scientist iteratively generating and refining synthetic training data.
- The practical implementation is called Agentic Self-Instruct, and meta-optimizing the data scientist agent itself produced a larger uplift than static methods.
- On legal reasoning tasks, a 4B parameter model trained on agent-made data reportedly beat a 397B parameter baseline.
Read full analysis →
Bayesian control for coding agents Theodore Papamarkou, Vladislav Smirnov, Viktor Mazanov, Artem Vazhentsev, Preslav Nakov, Timothy Baldwin, Artem Shelmanov https://t.co/1EUIZ7fmTy [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻] https://t.co/5sFFzguwnn
Bayesian control for coding agents arxiv.org
AI Weekly's analysis
→
- A new arxiv paper recasts coding-agent orchestration as cost-sensitive sequential hypothesis testing managed by a Bayesian controller.
- The controller decides dynamically whether to gather more evidence, refine the solution, run a verifier, or stop the run.
- Authors report the approach is most valuable when verification is costly and critics are informative but imperfect, across six generators and nine benchmarks.
Read full analysis →
auto-psych: Automating the science of mind using agent-driven theory discovery and experimentation Ben Prystawski, Kushin Mukherjee, Daniel Wurgaft, Linas Nasvytis, Michael Y. Li, Noah D. Goodman, Michael C. Frank https://t.co/U0Bv3bc9yi [𝚌𝚜.𝙰𝙸] https://t.co/S9vC7tD3F3
auto-psych: Automating the science of mind using agent-driven theory discovery and experimentation arxiv.org
AI Weekly's analysis
→
- Auto-psych uses nested loops: an inner loop generates probabilistic cognitive models, an outer loop designs and runs online human experiments.
- In three independent human experiments, the system's discovered theories fit the data better than theories drawn from the scientific literature.
- The benchmark task was a classic cognitive psychology problem about how people perceive randomness in coin flip sequences.
Read full analysis →
Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification Yunhao Feng, Ruixiao Lin, Ming Wen, Qinqin He, Yanming Guo, Yifan Ding, Yutao Wu, Jialuo Chen, Yunhao Chen, Xiaohu Du, Jianan Ma, Zixing Chen, … https://t.co/zQzYXScMTq [𝚌𝚜.𝙰𝙸] https://t.…
Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification arxiv.org
AI Weekly's analysis
→
- The Vera framework reports average attack success rates of 93.9% against four production agent frameworks under multi-channel attacks.
- Vera-Bench ships 1,600 executable safety cases spanning 124 risk categories, covering OpenClaw, Hermes, Codex, and Claude Code.
- Verifiers judge outcomes using environment state and tool-call evidence rather than the agent's own self-report of what happened.
Read full analysis →
Subliminal Clocks: Latent Time Modelling in Diffusion Language Models Maximo Rulli, Thomas Fontanari, Simone Petruzzi, Federico Alvetreti, Giorgio Strano, Donato Crisostomi, Giorgos Nikolaou, Tommaso Mencattini, Andrea Santilli, … https://t.co/jliEcX8tuE [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻] https://…
Subliminal Clocks: Latent Time Modelling in Diffusion Language Models arxiv.org
AI Weekly's analysis
→
- Diffusion language models lack explicit timestep conditioning yet still encode denoising progress in their residual streams, decodable by probes across layers.
- Steering the model along a low-dimensional subspace tied to the inferred timestep produces predictable shifts in output confidence and entropy.
- The latent time representation shows structured, interpretable geometry in activation space, per researchers at Sapienza University of Rome and EPFL.
Read full analysis →
Discrete Diffusion Language Models for Interactive Radiology Report Drafting Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert https://t.co/sCm225Db0x [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙻𝙶] https://t.co/Uag6VCLTZv
Discrete Diffusion Language Models for Interactive Radiology Report Drafting arxiv.org
AI Weekly's analysis
→
- DiffusionGemma-26B matches or exceeds its same-size autoregressive sibling Gemma-4-26B on every medical VQA dataset the authors tested.
- Decoding is reported at 3.5-4.4x faster than the AR baseline, with 3.8B active parameters after LoRA fine-tuning of the MoE model.
- Bidirectional denoising gives the model any-order infill, so a radiologist can fix report fragments and have the model fill between them.
Read full analysis →
Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination Subhadeep Pal, Shashwat Sourav, Tirthankar Ghosal, Markus J. Buehler https://t.co/iLYMDMaIY4 [𝚌𝚜.𝙰𝙸 𝚌𝚘𝚗𝚍-𝚖𝚊𝚝.𝚖𝚝𝚛𝚕-𝚜𝚌𝚒 𝚌𝚜.𝙲𝙻 𝚌𝚜.𝙻𝙶] https://t.co/SMlyxB2nbf
Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination arxiv.org
AI Weekly's analysis
→
- Graph-PRefLexOR uses Group Relative Policy Optimization to split reasoning into mechanism exploration, graph construction, pattern extraction, and hypothesis synthesis.
- On 100 open-ended materials science and mechanics questions, the system reports 40-65% improvements over base models, with the largest gains in reasoning traceability.
- Output embeddings show roughly 2-3x greater semantic diversity than baselines, which the authors credit to long-range recombination inside a bounded semantic space.
Read full analysis →
Two AI Metrics Diverged: Will it Make All the Difference? Alex Fogelson, Zachary A. Brown, Hans Gundlach, Jayson Lynch, Neil Thompson https://t.co/lV6uJDlT4w [𝚌𝚜.𝙰𝙸] https://t.co/Qef4NNwKJB
Two AI Metrics Diverged: Will it Make All the Difference? arxiv.org
AI Weekly's analysis
→
- A new arXiv paper argues whether small AI models eventually catch up to frontier systems depends entirely on which performance metric you pick.
- The authors show validation loss gaps shrink, but on other unbounded metrics frontier models grow their lead forever with more compute.
- Bounded and unbounded metrics can suggest opposing policy responses for capabilities like software engineering, synthetic biology, or rhetorical persuasiveness.
Read full analysis →
Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index Outongyi Lv, Yanzhao Zheng, Yuanwei Zhang, Zhenghao Huang, Xingjun Wang, Baohua Dong, Hangcheng Zhu, Yingda Chen https://t.co/GJb8WXUeGl [𝚌𝚜.𝙰𝙸] https://t.co/vSj7syoWEx
Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index arxiv.org
AI Weekly's analysis
→
- The Relative Surprisal Index combines a token's entropy with its selected probability to decide which tokens drive RLVR updates.
- RSI-S filtering reportedly beat baseline GRPO on Qwen2.5-1.5B, 3B and 7B by 2.10 to 3.30 points on AIME and AMC math benchmarks.
- Response lengths also fell by 108 to 265 tokens across the tested Qwen2.5 sizes, suggesting shorter outputs alongside the accuracy gains.
Read full analysis →
CDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipes Yuchen Huang, Xiang Li, Zhenqing Ling, Sijia Li, Qianli Shen, Daoyuan Chen, Yi R. Fung, Yaliang Li https://t.co/o7ndg7FcHI [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻] https://t.co/F25hfC03IN
CDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipes arxiv.org
AI Weekly's analysis
→
- CDR-Bench evaluates LLMs on 3,462 data refinement tasks spanning four real-world domains and 29 distinct operators, with deterministic reference outputs enabling exact scoring.
- Across more than ten state-of-the-art models, compositional performance degrades sharply and order-sensitive recipe success collapses.
- The authors conclude current LLMs lack the procedural faithfulness required for reliable compositional data refinement.
Read full analysis →
AI-Assisted Discovery of Convex Relaxations via Dual Agents Sungyoon Kim, Mert Pilanci https://t.co/YvTTbH9LBq [𝚌𝚜.𝙰𝙸] https://t.co/VhXAAjoKhA
AI-Assisted Discovery of Convex Relaxations via Dual Agents arxiv.org
AI Weekly's analysis
→
- Sungyoon Kim and Mert Pilanci pair a coding agent that proposes constraints with a theory agent that verifies proposals and searches for counterexamples.
- The system reports a tighter first autocorrelation bound (1.28 to 1.2937) and a tighter Erdős minimum-overlap bound (0.379005 to 0.37912).
- Every reported bound is certified using an explicit dual-feasible point validated through interval arithmetic, not just an empirical estimate.
Read full analysis →
