reasoning — AI Digest

28 июн OpenAI Previews GPT-5.6 Family: Sol, Terra, and Luna in Government-Gated Limited Release OpenAI models-llm
7 июл Tencent Officially Releases Hunyuan Hy3: 295B MoE Model with Agent and Reasoning Capabilities Tencent models-llm
9 июл OpenAI GPT-5.6 Sol, Terra, and Luna Launch Publicly After Government Review OpenAI models-llm
11 июл OpenAI Launches GPT-5.6 Family: Sol, Terra, and Luna OpenAI models-llm
17 июн VibeThinker-3B Reaches Frontier-Level Reasoning Benchmarks via Curriculum RL WeiboAI research
2 июл Orca: BAAI's General World Foundation Model Trained on 125K Hours of Video BAAI research
9 июл Tencent Releases Hunyuan Hy3: 295B Open-Weight MoE Model Under Apache 2.0 Tencent models-llm
11 июл Tencent Releases Hunyuan Hy3: 295B Open-Weight MoE Under Apache 2.0 Tencent models-llm
16 июл Thinking Machines Lab Releases Inkling: 975B Open-Weight Multimodal MoE Thinking Machines Lab models-llm
19 авг StateM: harness scaling pushes GPT-5.6 Sol to 95.3% on Terminal-Bench 2.1 and the same harness adapts to DeepSeek-V4 Flash for $15 total spend research
30 апр Recursive Multi-Agent Systems: agent communication in latent space Stanford University research
9 мая Zyphra Releases ZAYA1-8B: Open Reasoning MoE Model Trained on AMD Hardware Zyphra models-llm
10 мая Google DeepMind's AI Co-Mathematician Reaches 48% on FrontierMath Tier 4 Google DeepMind research
13 мая RubricEM: Meta-RL with Rubric-Guided Policy Decomposition Beyond Verifiable Rewards Google research
15 мая SU-01: Gold-Medal-Level Olympiad Reasoning via Curriculum SFT and Two-Stage RL SU-01 Team research
18 мая SOOHAK: Frontier LLMs Solve Hard Math But Fail to Recognize Unsolvable Problems research
20 мая Code as Agent Harness: Survey Positions Code as the Substrate for Executable Agent Systems (159 HF upvotes) Multi-institution (42 authors) research
20 мая SkillsVote: Lifecycle Governance of Agent Skills — Collection, Recommendation, Evolution (219 HF upvotes) Memtensor Research Group / IAAR-Shanghai research
18 июн Grok 4.3 Now Available on Amazon Bedrock with 1M-Token Context xAI models-llm
26 июн Qwen-AgentWorld: Language World Models for General Agents at 35B and 397B Scale Qwen Team, Alibaba research
2 июл ByteDance Releases Seed2.0 Model Card for Frontier Model ByteDance research
3 июл DeepSeek Confirms V4 Official Launch for Mid-July with Peak-Time API Pricing DeepSeek models-llm
7 июл The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning Tianjin University / Alibaba research
27 июл AREX: Towards a Recursively Self-Improving Agent for Deep Research research
3 авг OpenAI's Next Model Astra Solves Ten Open Problems in Math and Theoretical CS OpenAI research
12 авг Claude Research Model Raises Riemann Zeta Zero Lower Bound from 41.6% to 67.2% Anthropic research
11 сен Sber releases GigaChat 3.5 Reasoning, first Russian open-weights reasoning model Sber models-llm
14 сен Recurrent Looped Transformer: viral technical report claims 'infinite temporal depth' for latent reasoning Princeton University research
15 сен DataFlex-RL: evaluation platform finds RLVR data policies don't beat uniform sampling Peking University research
28 сен Fireworks Research ships Ember-1, a token-efficient reasoning model built on Kimi K3 Fireworks models-llm
18 мая RoPE Provably Fails at Long Contexts: Locality Bias and Token Consistency Both Break research
10 июн DRPO: Rethinking Divergence Regularization in LLM Reinforcement Learning Tencent Hunyuan research
14 июн MaxProof: MiniMax Model Exceeds IMO and USAMO Gold-Medal Thresholds on Formal Math MiniMax research
4 июл Mistral Releases Leanstral 1.5: Open Formal-Verification Model for Lean 4 Mistral research
8 июл LLM-as-a-Verifier: verification as an independent scaling axis for LLMs Stanford University / UC Berkeley / NVIDIA research
16 июл Ring-Zero: Scaling Zero RL to 1 Trillion Parameters with Emergent Reasoning Behaviors Ant Group research
29 июл Kimi K3 Technical Report: Kimi Delta Attention and Stable LatentMoE Architecture Detailed Moonshot AI research
12 авг BDH-CQ: Recurrent Latent Reasoning Model Breaks ARC-AGI-1 Cost-Efficiency Frontier Pathway research
27 авг IBM releases Granite 4.2 reasoning LLMs (3B/8B/30B, Apache 2.0, 512K context) trained on ~15T tokens with GB200 NVL72 GRPO RL IBM models-llm
5 сен Random Attention: rethinking KV cache eviction for efficient reasoning Salesforce AI Research research
13 сен NVIDIA Open Recipe for IMO Gold with Natural-Language Nemotron Pipeline NVIDIA research
16 сен ZGCM-1: a fully open, extremely efficient 7B foundation model for math and agentic search ZGCAGI research
24 сен The Tasteful Agent: measuring 'taste' in long-horizon tasks Microsoft research
23 окт PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking research
6 мая Ctx2Skill: Self-Improving Framework for Autonomous Context-Skill Discovery in LLMs research
8 мая AI Co-Mathematician: Google DeepMind Achieves 48% on FrontierMath Tier 4 Google DeepMind research
16 мая SDAR: Self-Distilled Agentic Reinforcement Learning for Multi-Turn Agents Zhejiang University / Meituan research
19 мая MMSkills: Reusable Multimodal Skills for General Visual Agents (105 HF upvotes) Shanghai Jiao Tong University research
2 июн GrepSeek: Training Search Agents for Direct Corpus Interaction via Shell Commands (93 HF Upvotes) University of Massachusetts Amherst research
4 июн ThoughtFold: Introspective Preference Learning Cuts Reasoning Tokens by 56% Without Accuracy Loss research
6 июн The Deterministic Horizon: Information-Theoretic Proof That Extended CoT Fails and Tool Use Is Necessary research
6 июн The Self-Correction Illusion: LLMs Fix Others' Errors but Not Their Own — Role Labels Are the Cause research
8 июн GitHub Copilot Gets 1M Token Context Window and Configurable Reasoning Levels GitHub / Microsoft tools
8 июн Agentic Transformers Provably Learn Depth-First Search via Reinforcement Learning Carnegie Mellon University / Ohio State University research
11 июн Arbor: Generalist Autonomous ML Research via Hypothesis-Tree Refinement NLPIR Lab research
11 июн DeNovoSWE: Full Repository Generation Jumps from 5.8% to 47.2% with Synthetic Training Data AweAI Team research
11 июн Z-Reward: Score Distributions Instead of Scalar Rewards for Image Generation RLHF Alibaba research
25 июн Quantized Reasoning Models Think They Need to Think Longer, but They Do Not Meta research
26 июн The Deterministic Horizon: When Extended Reasoning Fails and Tool Delegation Becomes Necessary research
29 июн OPID: On-Policy Skill Distillation Improves Long-Horizon Agent RL Institute of Automation, Chinese Academy of Sciences research
7 июл CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents Zhipu AI / Tsinghua University research
11 июл LLM-Driven Formal Mathematics Review: Where Current Systems Fall Short UCLA research
17 июл SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration Ant Group / Renmin University research
17 июл SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning research
22 июл TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs MCG-NJU (Nanjing University) research
23 июл SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD SLAI research
23 июл Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimization research
2 авг Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering Horizon Research, Frontis.AI, Tsinghua University research
5 авг DAPD: Dual-Anchored Policy Distillation research
6 авг AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling Nanjing University research
9 авг LongHorizon-Harness proposes Manage-Execute-Audit loop for long-horizon agents research
10 авг AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning Tsinghua University research
13 авг On-Policy Self-Distillation without Any Supervision research
14 авг Intern-S2-Preview: Scientific Agentic Foundation Model Shanghai AI Lab research
29 авг TTPO: Test-Time Policy Optimization Zhejiang University research
5 сен Legibility is not interpretability: judged vs. actual step importance in chain-of-thought research
8 сен Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems research
23 сен RRSI: Regularized Recursive Self-Improvement of Agent Harnesses Google Research research
25 сен X-Planner: event-structured task planning for embodied intelligence X-Square Robot research
17 окт Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation research
2 мая ESamp: LLMs explore by latent distilling for semantic-novelty sampling ShanghaiTech University research
5 мая Odysseus: Training VLMs for 100+ Turn Interactive Decision-Making via RL Princeton University research
11 мая Soohak: 64 Mathematicians Build Research-Level Benchmark That Stumps Frontier LLMs Seoul National University research
11 мая AutoTTS: LLM Agents Automatically Discover Test-Time Scaling Strategies for $40 research
3 июн TrOPD: Trust-Region On-Policy Distillation Stabilizes LLM Training When Teacher-Student Gap Is Large Samsung Research research
3 июн Do Language Models Need Sleep? Offline Recurrence as Memory Consolidation for Improved Inference Google / CMU research
12 июн InterleaveThinker: RL Framework for Agentic Text-and-Image Interleaved Generation research
12 июн Astra: RL-Trained VLM Queries World Simulator for Spatial Reasoning research
10 июл Ideas Have Genomes: Frontier LLMs Score Only 27% on Scientific Lineage Reasoning Shanghai Jiao Tong University research
30 июл Pass the Baton: Trajectory-Relayed On-Policy Distillation Zhejiang University research
17 авг Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning Shanghai AI Lab research
6 мая HeavySkill: Internalizing Heavy Thinking as a Trainable Agentic Skill via RL research
7 мая LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents Shanghai Jiao Tong University research
7 мая Executable World Models for ARC-AGI-3: Coding-Agent Approach Without Game-Specific Logic research
12 мая NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized AI Research Automation Shanghai AI Lab research
12 мая TMAS: Scaling Test-Time Compute via Multi-Agent Synergy with Hierarchical Memory research
13 мая Learning, Fast and Slow: Dual-Weight Architecture for Continual LLM Adaptation research
18 мая BetaPRM: Uncertainty-Aware Process Rewards Cut Reasoning Token Use by 33% research
19 мая NudgeRL: Strategy-Level Context Nudges for Efficient RLVR Exploration KAIST AI research
3 июн QUBRIC: Co-Designing Queries and Rubrics Extends RLVR to Open-Ended Reasoning Domains research
3 июн Quantifying Faithful Confidence Expression in Large Reasoning Models Yale NLP research
8 июн SubtleMemory: Benchmark Reveals Agents Systematically Fail Fine-Grained Relational Memory research
8 июн VideoKR: 315K-Example Training Corpus for Knowledge- and Reasoning-Intensive Video Understanding Yale University research
9 июн Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight Rutgers University research
10 июн SearchSwarm: Delegation Intelligence for LLM Agents in Long-Horizon Deep Research research
16 июн Memory is Reconstructed, Not Retrieved: Graph Memory Improves LLM Agent Recall by 23% National University of Singapore research
17 июн ZPPO: Teacher-in-Prompts Knowledge Distillation Outperforms Gradient Methods for Small Reasoners NVIDIA research
18 июн Diffusion-Proof: Formal Theorem Proving via Diffusion Language Models research
18 июн DreamReasoner-8B: Block-Size Curriculum for Diffusion Reasoning Models research
22 июн S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence in VLMs Nanyang Technological University research
23 июн Agentic Transformers Provably Learn to Search via Reinforcement Learning research
26 июн Dense Supervision Is Not Enough: The Readout Blind Spot in Looped Language Models research
26 июн OPRD: On-Policy Representation Distillation for Post-Training LLMs research
30 июн Formalizing Latent Thoughts: Axiomatic Framework for Evaluating LLM Reasoning Representations University of British Columbia research
2 июл OpenCode v1.17.13: Reasoning Mode Fixes and Searchable Model Picker SST tools
4 июл Will Scaling Improve Social Simulation with LLMs? A Study of 85 Models Stanford / Columbia / Tsinghua research
4 июл VRRL: Visually Grounded Self-Reflection for Vision-Language Models via RL UT Austin / Cornell research
7 июл Weak-to-Strong Generalization via Direct On-Policy Distillation ByteDance / Tsinghua University research
9 июл RL Post-Training Actively Builds Compositional Reasoning Strategies, Not Just Amplifies Base Skills research
14 июл OpenCode 1.17.19–1.17.20: OpenAI Pro Reasoning Mode and Per-Prompt Model Selection SST tools
14 июл AdvancedMathBench: Benchmark Suite for Advanced Mathematical Proof Generation and Verification InternLM research
14 июл Metacognition in LLMs: Foundations, Progress, and Opportunities — Yale Survey Yale NLP Lab research
18 июл LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget Mind Lab research
18 июл Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning inclusionAI research
21 июл Understanding Reasoning from Pretraining to Post-Training research
21 июл Recursive Harness Self-Improvement (RHI): Refining Agent Harnesses from Execution Feedback Sakana AI research
26 июл LLMs Get Lost in Evolving User Intent Microsoft Research research
29 июл Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills Alibaba research
3 авг From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement research
6 авг GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning research
8 авг AgentOPSD: recursive self-distillation for agentic reinforcement learning research
13 авг Stealing Reasoning Traces from Proprietary LLM APIs research
14 авг AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design Meituan research
20 авг Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL UC San Diego (Yunhao Yang, Nuno Vasconcelos, Yijiang Li et al.) research
26 авг Groq Python SDK v1.7.0 — Qwen3.8 reasoning guidance, dependency vulnerability lockfile update Groq tools
29 авг Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO research
1 сен Does on-policy distillation really distill? Teacher-free OPSA beats it on AIME24 Purdue University research
1 сен Survey: scaling reasoning models beyond human supervision via a five-level autonomy ladder research
6 сен Rethinking On-Policy Distillation of LLMs II: near-full gains from a single training example research
8 сен Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs NAVER AI Lab research
9 сен FlowBalance: verifier-grounded self-improvement for reasoning models research
17 сен StepAudio 3 Realtime: think-while-speaking audio-language model with asynchronous tool calls StepFun research
21 сен When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation research
25 сен SpeakerMem-R1: speaker-centered dual-track memory for multi-party dialogue Zhejiang University research