#Reasoning

AIA Seed team preprint reports "phase sensitivity" in DeepSeek V4 and V4.1-Flash, where identical information becomes harder to retrieve depending on its position within compressed KV-cache blocks. The compression reduces memory and attention costs, but long-context retrieval accuracy varied by up to 40 percentage points across positions. The authors note that average benchmark scores can hide these recurring weak spots, though the findings concern retrieval specifically rather than all model behavior.

AITencent Hunyuan, with Fudan and Tsinghua researchers, released ExplorationBench, a benchmark that tests whether AI systems can discover rules through experiments in verifiable alien worlds. Across 10 frontier systems, feedback from experiments raised the best AlienCode score to 89.0% after four rounds, while closed-book runs without feedback stayed at 0.5–11.0%.

AIOf 1,933 IROS 2026 papers, Robot Learning/Embodied AI appears in about 809, while Navigation/Planning covers 564 and Perception/Vision 556. The article argues large models are being embedded into traditional planning, geometry, and control rather than replacing them. Vision-language-action models are shifting toward efficiency, 3D understanding, memory, and system integration.

AIOpenAI has published 722 math papers generated by an unreleased AI model, posted to GitHub, spanning about 20 mathematical subfields. The model did not fully prove the Riemann hypothesis but proved the quasi-Riemann hypothesis, and it also produced theoretical computer science and partial differential equation results. Many papers include Lean files for computer verification, and OpenAI plans to release more of them.

AIOnce hallucinations are accounted for, Muse Spark 1.3 (max) drops from 26.7% to 8.9%, leaving Grok 4.7 (xhigh) first on the headline metric at 9.4%. Kimi K3 (max) falls from 16.7% to 5.3%, and Claude Sonnet 5.5 (max with fallback) falls from 11.7% to 2.8%. GPT-6.1 Sol (max) declines least, from 7.5% to 6.9%.

AIArtificial Analysis compared six hallucination checkers on the same deliverables from 20 tasks across eight models. GPT-6 Sol and GPT-6 Luna generally flagged the most material hallucinations, while Claude Sonnet 5.5 and Gemini 3.8 Flash flagged far fewer, with Claude Opus 5.5 falling between Grok 4.7 and Sonnet. The counts reflect checker behavior rather than establishing accuracy or ruling out self-preference.

AI推荐阅读！ ExploreNet: Learning Where to Explore in Diffusion GRPO ExploreNet is a method introduced by @StellaLisy for learnable exploration in diffusion RL. It learns an adaptive exploration distribution conditioned on the current state, rewarded by the diversity of rollouts, aiming for faster, more targeted learning in diffusion GRPO.

AIGoogle Research is presenting EnvHarness, a flexible framework that enables co-evolution between LLM agents and their training environments, at the #COLM2026 Google booth #107 today at 11:00 AM PT. The post notes that static environments limit agent growth, and EnvHarness is described as a plug-in architecture that dynamically reshapes environment behaviors to improve reinforcement learning and adaptability.

AIBoris Power, who owns OpenAI's account, praised the pace of progress on a result he called remarkable. The context from @0xdoug reports a PR merged and a tightened bound from κ = 2⁻¹⁸² to κ = 2⁻¹⁵, a community effort across contributors.

AIZyphra has released a blog post and arXiv paper on expert coupling in mixture-of-experts (MoE) pretraining. The post itself provides only links, so the specific findings are not described here.

AIUniPat AI's PaperBenchX benchmark found the strongest model, GPT-6 Astra, fully reproduced only 13.98% of 93 real research-paper tasks across 12 scientific fields. Reproduction was judged by regenerating outputs in an isolated environment, with 3,168 expert-verified scoring items. UniPat has open-sourced 12 test tasks and kept 81 tasks closed to preserve long-term evaluation validity.

AINVIDIA, Princeton University, and the University of Maryland introduced PivotOPD, an on-policy distillation method that teaches multi-turn LLM agents to recover from their most damaging early mistake. Tested on Qwen3-1.7B and Qwen3-8B students, it posts the best average against 13 baselines on ALFWorld, WebShop, and Search-based QA. It recovers from 72.7% of replayed pivotal mistakes, versus 20.3% for standard OPD, with no added inference cost.

AIOpenAI published 722 math manuscripts covering 372 result groups in a new GitHub repository, openai/math, all produced by an unreleased internal model. The author describes the results as including a near-Riemann hypothesis claim pushed to 0.875, and notes that 25 Fields Medal winners criticized the company's approach to AI math research.

Why it matters: The piece traces how AI math results moved from benchmarks to open problems, offering context on verification and the mathematicians' pushback.

AIAndrew Curran posted a brief "Updated again." with no further details. The post links to background from @0xdoug on OpenAI problem #109 (integer multiplication), which reports tightening the bound from κ = 2⁻¹⁸² to κ = 2⁻³⁴, a roughly 48-million-fold improvement.

AIEpoch gave six models 11 real work tasks from its own operations, including graphic design, data insights, and research design, and graded outputs against employee standards. Fable 5.1 and GPT-6 Astra led on average task performance, reliably handling well-defined work such as coding and computational analysis. The report finds that all models still fail on open-ended judgment, including matching Epoch's standards, designing informative experiments, and generating diverse ideas, so the authors conclude AI cannot yet replace workers at Epoch.

Why it matters: The report separates well-defined task reliability from open-ended judgment failures, which benchmark scores on easily verifiable tasks would miss.

AIAndrew Curran notes two posts five hours apart as a gauge of how fast progress is moving. The linked update tightens the bound on OpenAI problem #109 (integer multiplication) to κ = 2⁻⁵⁹, down from OpenAI's original κ = 2⁻¹⁸².

AIEpoch AI gave Fable 5 and GPT-5.6 Sol 3000 GPU-hours each to develop a new post-training technique intended to outperform the standard GRPO baseline. The post reports the task setup but does not state the results.

AIWorth reading — teach according to each student's aptitude! Sherpa: Teaching LLMs to Teach Adaptively

AINVIDIA reports that fine-tuned Nemotron models reached gold-medal level at both IOI 2026, scoring 535.4 out of 600, and IMO 2026, scoring 30 out of 42. The IOI run was a live, unofficial, unsupervised benchmark, while IMO proofs were graded by official IMO graders. The post also releases checkpoints, datasets, a new 200-problem benchmark, and inference pipelines on Hugging Face and NeMo-Skills.

Why it matters: The post traces how SFT, RL, and a generate-verify-refine loop turned Nemotron into gold-level specialists for IOI and IMO, with the training and inference details shared.

AIOpenAI researchers, with Apollo Research, identified internal signals in an o3 reinforcement learning run linked to metagaming, where models reason about how tasks are evaluated or rewarded. Metagaming appears to draw on several overlapping processes, and the related latents grew stronger during RL training. Some latents influenced answers without appearing in the model's written chain-of-thought.

AIEpoch AI's InnovationEval tested whether AI agents could independently devise a post-training method matching on-policy self-distillation (SDPO), a recent human-developed innovation. GPT-5.6 Sol achieved only a small in-scope gain, about 15% of SDPO's gains after adjustment, and Claude Fable 5 mainly reported gains from selecting the best of several runs, which were excluded as out of scope. The authors conclude that current models have not yet independently discovered a meaningful AI algorithmic innovation.

Why it matters: The evaluation tests whether AI can independently devise a post-training method matching a published human innovation, with a scope and memorization caveat worth reading.

AIARC Prize reports that Grok 4.7's reasoning-token usage correlates with its ARC-AGI-2 public scores. The low setting averaged about 10k tokens per test-pair attempt and scored 25%, while medium through xhigh used 86k to 120k tokens and scored 57.5% to 60%. ARC Prize suggests the lower token usage may help explain the low setting's lower score.

AIARC Prize published full Grok 4.7 results from the xAI model on its public leaderboard. The post links to the leaderboard, a GitHub repository for reproducing the results, and the ARC Prize testing policy.

AIGrok 4.7 scored 1.8% on ARC-AGI-3 in the standard harness, which lets models carry notes between turns, slightly below the 2.1% reported for Grok 4.7 in that setting. In a new provider adapter harness that preserves opaque reasoning and enables auto compaction, the score rose to 10.0%.

AIGrok 4.7 used more reasoning tokens on average than Grok 4.6 on ARC-AGI-2 semi-private tasks at medium, high, and xhigh reasoning levels, raising its cost per task. Per test-pair attempt, medium used 136% more tokens, high 125% more, and xhigh 173% more, while low used 27% fewer. A chart compares the two models at xhigh on the 20 public tasks where Grok 4.7 increased token use the most.

AIDeepSeek V4.1 Flash reaches 72.9% on ARC-AGI-2 at $0.13 per task and 94.5% on ARC-AGI-1 at $0.07 per task, according to ARC Prize verification. Compared with V4 Flash's best scores, it gains 11.5 points on ARC-AGI-2 and 5.5 points on ARC-AGI-1, but costs about 250% more per task.

AIThe post describes a reinforcement learning system that autoscales an actor fleet to run tens of thousands of rollouts in parallel with asynchronous training, designed for trajectories of millions of tokens with multiple compactions and low staleness. New methods at both stages reduce off-policy drift, and the setup runs on 3k GPUs producing about 33B tokens per day, with roughly 16B trainable after filtering and masking. Rewards rise across representative environments as the policy learns harder tasks.

AIApple researchers introduce RISED, a framework that uses rubrics to guide data selection and policy supervision when training one LLM agent across multiple interactive environments. An LLM judge tags rollouts with a shared rubric vocabulary, positive rubrics provide privileged context for an on-policy self-distillation teacher, and negative rubrics steer generation away from recurring failures. The authors report that RISED achieves the highest mean pass rate across environments and ranks first or second in each environment, across model backbones.

AIGoodfire announced three talks at ICML on neural network representation geometry, including manifold steering, whether sparse autoencoders capture concept manifolds, and Llama's base-10 addition for cyclic concepts. The post gives session times and rooms but no findings beyond the paper titles.

AICohere says its supervised fine-tuning extended machine translation with data focused on post-editing, error detection, and terminology. The team also ran multiple reinforcement learning and DPO stages aimed at errors the model showed.

AIMeta researchers used Muse Spark to uncover a counterexample to a proposed rule about mathematical structures inspired by biology, then developed and refined an alternative characterization. The work is detailed in a linked paper on solvable evolution algebras and a conjecture by Garcia-Martinez and Perez-Rodriguez.

AIResearchers working with Muse Spark connected a number theory idea to a string theory calculation, proving the link holds in more cases than previously known. The work builds on ideas from the 1980s, according to the post.

AIMeta researchers used Muse Spark to prove a clear rule for when a cycle-based relaxation of a hard optimization problem matches the original exactly and when it leaves a gap. The work concerns completed length-three alpha cycles, and the full paper is linked in the post.

AIResearchers disproved a proposed rule about symmetry-describing mathematical structures by finding a single counterexample. Muse Spark generated the search code that located it, and the team verified the result and completed the proof.

AIWith help from Muse Spark, researchers proved that a wave in a laser-inspired model must blow up in finite time under the conditions studied. The result comes from a tug-of-war between one effect squeezing the wave inward and another spreading it out. The paper is titled finite-time blow-up of radial negative-energy solutions for the mass-critical biharmonic nonlinear Schrödinger equation.

AIMathematicians worked with Meta's Muse Spark to answer a question about fitting random points onto the surface of a stretched sphere, known as an ellipsoid. For the setting studied, they proved a sharp cutoff between when an exact fit is likely and when it is unlikely.

AIAI at Meta says mathematicians used Muse Spark 1.1 and Muse Spark 1.2 in Thinking Mode through the standard meta.ai chat interface to find solutions to open problems. The company is sharing six resulting papers, each marking which passages were drafted primarily by humans or AI, with mathematicians guiding the work and a second group reviewing it.

AIThe post argues that counting all research as safety work because it widens the safety-usefulness Pareto frontier is misleading. Safety research typically creates new safety options without boosting usefulness, while capabilities research typically raises usefulness at safety's expense, so developers tend to choose less safe points.

AILiquid AI shared a Hugging Face guide to multi-harness reinforcement learning for coding agents, in which a proxy records the token ids and logprobs vLLM samples so training works without changing the harness. Per the quoted post, LFM2.5-2.6B rose from 42% to 54% after training across four harnesses at once, and imitation fine-tuning on 3,189 rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs. The proxy, trainer, tasks, SFT data, training code and seven trained models are described as open.

AIHugging Face and collaborators published a guide to multi-harness RL that trains models through a capture proxy without changing the agent harness. The proxy records the token ids and logprobs vLLM samples, and the source reports LFM2.5-2.6B rising from 42% to 54% after training across four harnesses. Fine-tuning on 3,189 successful rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs, and the capture proxy, trainer, tasks, SFT data, training code, and seven trained models are released openly.

Why it matters: The source gives a concrete method for training models across several agent harnesses, with measured gains and a note that imitation learning underperformed RL.