Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model (Ring-1T)
Technical report on scaling reinforcement learning to a trillion-parameter reasoning model.
Publications论文
You can also find my articles on my Google Scholar profile.完整论文列表也可以在我的 Google Scholar 主页 查看。
Technical Reports技术报告
Conference & Journal Publications会议与期刊论文
InftyThink+: Effective and Efficient Infinite-Horizon Reasoning via Reinforcement Learning
Scales infinite-horizon reasoning with reinforcement learning for more effective and efficient long thinking.
InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models
Breaks context-length limits by turning long reasoning into iterative, bounded-length thinking.
MathFimer: Enhancing Mathematical Reasoning by Expanding Reasoning Steps through Fill-in-the-Middle Task
Expands reasoning chains via a fill-in-the-middle task to strengthen mathematical reasoning.
VerifyBench: Benchmarking Reference-based Reward Systems for Large Language Models
A benchmark for evaluating reference-based reward systems that verify LLM answers.
S³cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners
Spontaneous step-level self-correction makes LLMs stronger mathematical reasoners.
Code-A1: Adversarial Evolving of Code LLM and Test LLM via Reinforcement Learning
Adversarially co-evolves a Code LLM and a Test LLM with opposing RL objectives.
Test-Time Reinforcement Learning for GUI Grounding via Region Consistency
Improves GUI grounding at test time via label-free region-consistency reinforcement learning.
Let LRMs Break Free from Overthinking via Self-Braking Tuning
Teaches reasoning models to stop overthinking by learning when to brake.
Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning
Detects and bridges thought leaps in chains of thought to improve reasoning fine-tuning.
GSM8K-V: Can Vision Language Models Solve Grade School Math Word Problems in Visual Contexts
Tests whether vision-language models can solve grade-school math word problems shown visually.
IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis
Decouples deep-search agents into a Planner and a Synthesizer over an evolving summary state, trained with role-specific RL.
Learning from Reliable Negatives: Confidence-Anchored Test-Time Adaptation for GUI Grounding
Label-free test-time adaptation for GUI grounding that learns from reliable negative samples anchored by coordinate-token confidence.
Pause or Fabricate? Training Language Models for Grounded Reasoning
Multi-turn RL that teaches models to pause and ask for clarification instead of fabricating missing premises.
Milestone-Guided Policy Learning for Long-Horizon Language Agents
Guides long-horizon language agents with milestone signals to make policy learning more stable.
SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models
Progressive training that builds up spatial reasoning ability in vision-language models.
CoVerRL: Breaking the Consensus Trap in Label-Free Reasoning via Generator-Verifier Co-Evolution
Escapes the consensus trap of label-free RL by co-evolving a single model as both generator and verifier.
Teaching LLMs According to Their Aptitude: Adaptive Switching Between CoT and TIR for Mathematical Problem Solving
Adaptively switches between chain-of-thought and tool-integrated reasoning based on problem difficulty.
GUI-SAGE: Enhancing GUI Automation with Self-Explanatory Learning
Self-generated, in-distribution explanations and entropy-modulated credit assignment to escape zero-advantage traps in GUI RL.
Do Large Language Models excel in Complex Logical Reasoning with Formal Language?
Investigates whether LLMs can perform complex logical reasoning expressed in formal language.
AskToAct: Enhancing LLMs Tool Use via Self-Correcting Clarification
Enhances LLM tool use by self-correcting clarification of ambiguous requests.
SVGenius: Benchmarking LLMs in SVG Understanding, Editing and Generation
A benchmark for LLM understanding, editing, and generation of SVG graphics.
EasySteer: A Unified Framework for High-Performance and Extensible LLM Steering
A vLLM-based, extensible framework for LLM steering with 10.8-22.3x speedup over existing tools.
LAPO: Internalizing Reasoning Efficiency via Length-Adaptive Policy Optimization
Two-stage RL that lets models internalize an appropriate reasoning length for each problem.
UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models
A comprehensive benchmark for undergraduate-level physics reasoning in LLMs.
LogicPro: Improving Complex Logical Reasoning via Program-Guided Learning
Improves complex logical reasoning by learning from program-guided supervision.
ViewSpatial-Bench: Evaluating Multi-Perspective Spatial Localization in Vision-Language Models
A benchmark for camera- and human-perspective spatial localization in vision-language models.
Triad: A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering
A multi-role LLM agent framework for knowledge base question answering.
Preprints预印本
Cooper: Co-Optimizing Policy and Reward Models in Reinforcement Learning for Large Language Models
Co-optimizes the policy and reward models together in reinforcement learning for LLMs.
PaperGym: Rubric-Centered Evolution for Research-Plan Generation
Turns each research paper into an RL environment with leakage-resistant rubrics for training research-plan generation.
Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning
Scales zero RL to a trillion-parameter model and studies the emergent reasoning behaviors that arise.
KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation
An online Android benchmark for mobile agents that must infer user preferences and decide when to act proactively.
Hierarchical Budget Policy Optimization for Adaptive Reasoning
Budget-constrained hierarchical RL that teaches models problem-specific reasoning depth.
Double-Checker: Enhancing Reasoning of Slow-Thinking LLMs via Self-Critical Fine-Tuning
Fine-tunes long-CoT models to iteratively critique and refine their own solutions.
Workshop PapersWorkshop 论文
OmniEAR: Benchmarking Agent Reasoning in Embodied Tasks
Benchmarks agent reasoning in embodied tasks across diverse environments.