rl — AI Digest

17 июн VibeThinker-3B Reaches Frontier-Level Reasoning Benchmarks via Curriculum RL WeiboAI research
3 мая Exploration Hacking: LLMs Can Be Fine-Tuned to Strategically Resist RL Training research
9 мая OpenAI Discloses Accidental Chain-of-Thought Grading in RL Training Across Six Models OpenAI research
10 мая Google DeepMind's AI Co-Mathematician Reaches 48% on FrontierMath Tier 4 Google DeepMind research
11 мая Flow-OPD: On-Policy Distillation Pushes GenEval +29 Points on Stable Diffusion 3.5 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
20 мая SkillsVote: Lifecycle Governance of Agent Skills — Collection, Recommendation, Evolution (219 HF upvotes) Memtensor Research Group / IAAR-Shanghai research
26 июн Qwen-AgentWorld: Language World Models for General Agents at 35B and 397B Scale Qwen Team, Alibaba research
10 мая Anthropic Eliminates Claude's Agentic Blackmail Behavior via 'Teaching Claude Why' Anthropic research
10 июн DRPO: Rethinking Divergence Regularization in LLM Reinforcement Learning Tencent Hunyuan research
4 мая Learning while Deploying: Fleet-Scale Reinforcement Learning Turns Robot Deployment into Continuous Training AGIBot research
6 мая Ctx2Skill: Self-Improving Framework for Autonomous Context-Skill Discovery in LLMs research
7 мая RLDX-1: Multi-Stream Action Transformer Achieves 86.8% on ALLEX Humanoid Tasks RLWRLD research
9 мая OpenSearch-VL: Open Recipe for Training Frontier Multimodal Search Agents Tencent Hunyuan research
16 мая SDAR: Self-Distilled Agentic Reinforcement Learning for Multi-Turn Agents Zhejiang University / Meituan 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
8 июн Agentic Transformers Provably Learn Depth-First Search via Reinforcement Learning Carnegie Mellon University / Ohio State University research
10 июн Flow-DPPO: Principled RL Alignment for Flow Matching Image and Video Models Tencent Hunyuan research
11 июн Arbor: Generalist Autonomous ML Research via Hypothesis-Tree Refinement NLPIR Lab research
11 июн Z-Reward: Score Distributions Instead of Scalar Rewards for Image Generation RLHF Alibaba research
14 июн InterleaveThinker: RL Planner+Critic Pipeline for Interleaved Text-and-Image Generation CUHK Multimedia Lab research
2 мая CoPD: co-evolving policy distillation for unified multi-capability models research
5 мая Odysseus: Training VLMs for 100+ Turn Interactive Decision-Making via RL Princeton University research
3 июн TrOPD: Trust-Region On-Policy Distillation Stabilizes LLM Training When Teacher-Student Gap Is Large Samsung Research research
12 июн InterleaveThinker: RL Framework for Agentic Text-and-Image Interleaved Generation research
12 июн FORT-Searcher: Shortcut-Resistant Training Data Framework for Deep Search Agents research
12 июн Astra: RL-Trained VLM Queries World Simulator for Spatial Reasoning research
28 июн The Verification Horizon: No Single Reward Function Works for Coding Agents at Scale Qwen (Alibaba) research
6 мая HeavySkill: Internalizing Heavy Thinking as a Trainable Agentic Skill via RL 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
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
9 июн On the Geometry of On-Policy Distillation: A Training Paradigm Distinct from SFT and RLVR Hong Kong University of Science and Technology research
9 июн Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight Rutgers University research
17 июн ZPPO: Teacher-in-Prompts Knowledge Distillation Outperforms Gradient Methods for Small Reasoners NVIDIA research
23 июн Agentic Transformers Provably Learn to Search via Reinforcement Learning research
26 июн OPRD: On-Policy Representation Distillation for Post-Training LLMs research
28 апр Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond HKUST/NUS/Oxford/NTU research
28 апр World-R1: Reinforcing 3D Constraints for Text-to-Video Generation Microsoft Research research