Papers & research

AIOpenAI reports that its automated system found accidental chain-of-thought grading in RL runs for several released models, including GPT-5.4 Thinking and GPT-5.4 mini. Its analysis found no clear reduction in CoT monitorability, though the company says subtler effects cannot be ruled out. OpenAI says it still avoids grading CoTs during RL and has fixed the affected reward pathways.

Why it matters: The post shows how accidental chain-of-thought grading was detected and tested, giving a concrete method for checking monitorability risks in RL training.

AIOpenAI released Auto-review in Codex, which replaces user approval at the sandbox boundary with a separate agent that approves or denies boundary-crossing actions. In internal deployment, Codex sessions stopped for human approval about 200x less often than in manual mode, and Auto-review approved around 99% of escalated actions. The post also states that Auto-review is not a guarantee of security and cannot protect against model scheming.

Why it matters: The post explains how Auto-review replaces human approval at the sandbox boundary, with internal deployment figures and stated limits that help readers judge the tradeoff for coding agents.

AICognition and Applied Compute RL-trained SWE-check, a specialized bug detection model for the Windsurf IDE. It matches frontier performance on in-distribution evals and is an order of magnitude faster with cheaper inference, though it trails frontier models on out-of-distribution evals (delta F1 0.29 versus 0.49 before training). A preview is available in Windsurf Next, with a mainstream release planned.

Why it matters: The post explains how production environment replication, reward linearization, and two-phase post-training trade bug-detection quality against latency for an IDE specialist model.

AIARC Prize has released ARC-AGI-3, a set of hundreds of interactive, turn-based environments with thousands of game-style levels, with no instructions or stated goals. Humans score 100% while frontier AI scores 0.51%. ARC Prize 2026 offers over $2 million in prizes for open-source solutions to ARC-AGI-2 and ARC-AGI-3.

Why it matters: The benchmark's human versus frontier AI gap and its interactive design show how agent evaluation is shifting from instruction-following toward exploration and adaptation.

AIGoogle researchers used the Aletheia agent, powered by Gemini 3 Deep Think, to attempt 10 FirstProof challenge problems without modification. The agent operated fully autonomously and solved 6 of the 10 problems, according to the post, with methodology and expert evaluations described in the linked arXiv paper.

Why it matters: The post gives the autonomous setup and expert-evaluated results for an AI agent on FirstProof math problems, useful for judging how far such systems go on research-level math.

AIMiniMax describes Forge, its internal reinforcement learning framework for training real-world agents, which was used during the development of MiniMax M2.5. The post explains a Windowed FIFO scheduler, prefix tree merging that the post says yields a 40x training speedup, and CISPO-based training across more than one hundred thousand agent scaffolds and environments.

Why it matters: The post details how the Forge framework balances throughput, stability, and agent flexibility, with concrete scheduling and prefix-merging methods for training agent RL at scale.

AIAnthropic reports that resource configuration alone can move Terminal-Bench 2.0 scores by up to 6 percentage points, with infra error rates falling from 5.8% under strict enforcement to 0.5% when uncapped. Above about 3x the per-task specs, extra headroom starts letting agents solve tasks they previously could not, so limits can change what the eval measures.

Why it matters: The source shows how container resource limits shift agentic coding scores, which helps readers interpret small leaderboard gaps and set up evals more consistently.

AIThinking Machines Lab describes on-policy distillation, which samples rollouts from a student model and has a teacher grade each token with reverse KL. The authors report that this matches Qwen3-style reasoning results at a fraction of RL's cost, with AIME'24 reaching 70% in about 150 steps from a 400k SFT checkpoint. The method also helps recover instruction-following behavior lost during fine-tuning on internal documents.

Why it matters: The post explains why on-policy distillation gives dense per-token feedback, letting a small model match RL results at much lower compute cost.

AILiquid AI reports STAR, an evolutionary algorithm that synthesizes tailored neural network architectures from a numerical genome representation. The authors say it produced hundreds of designs that outperform Transformer and hybrid architectures in quality, with smaller caches and parameter counts, and can optimize for latency on target hardware. The full method is described in the arXiv technical report 2411.17800.

Why it matters: The post explains how evolutionary search over a new architecture design space produced designs beating Transformers and hybrids, giving a concrete method for quality versus latency and memory trade-offs.

AICognition reports that its agent Devin resolved 79 of 570 sampled SWE-bench issues, a 13.86% success rate, without being given the files to edit. The report says this exceeds the best previous unassisted baseline of 1.96% and the best assisted result of 4.80%. It also describes the adapted evaluation setup, a 45-minute runtime limit, and cases where Devin failed on multi-file edits.

Why it matters: The report explains how SWE-bench was adapted for end-to-end agent evaluation, with failure cases that clarify where the 13.86% result comes from and its limits.