{"id":"arxiv:2512.02807","title":"SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment","date":"2025-12-02","section":"Post-training","quality":0.238309,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/","source_url":"https://arxiv.org/abs/2512.02807","text_available":true,"authors":["Yixuan Tang","Yi Yang"],"outgoing":[{"source":"arxiv:2512.02807","target":"arxiv:2402.03300","evidence":"title","id":"arxiv:2402.03300","title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/636fb871904bc06f742cfdea069796f9e9665f95bd58dc63260eaff5154ef985/"},{"source":"arxiv:2512.02807","target":"arxiv:2407.21783","evidence":"title","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"},{"source":"arxiv:2512.02807","target":"url:rlhfbook.com","evidence":"title","id":"url:rlhfbook.com","title":"Reinforcement Learning from Human Feedback","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/20e354559fad09ad4d6338d104c9c3f822a48e7fbf86cd8afa1ddf951b223658/"}],"incoming":[],"text_parts":[{"start":0,"end":15518,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html"},{"start":15518,"end":31514,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html"},{"start":31514,"end":47385,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html"},{"start":47385,"end":58393,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html"}],"contents":[{"title":"SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Abstract","offset":92,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"1 Introduction","offset":1459,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"2 Stable Rank as an Intrinsic Quality Metric","offset":6862,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"2.1 Definition","offset":7155,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Theoretical motivation.","offset":8199,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Implementation.","offset":8736,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"2.2 Empirical Validation: Stable Rank as Zero-Shot Reward Proxy","offset":9336,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Setup.","offset":9658,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Results.","offset":10809,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"2.3 Empirical Validation: Best-of-N Decoding with Stable Rank","offset":12075,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Setup.","offset":12447,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Results.","offset":13172,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"3 Stable Rank Group Relative Policy Optimization (SR-GRPO)","offset":14056,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"3.1 Setup","offset":14501,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"3.2 Training Procedure","offset":14863,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-1.html#text"},{"title":"Computational efficiency.","offset":16260,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"4 Alignment Experiments","offset":16657,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"4.1 Experiment Setup","offset":16685,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"4.2 Results Analysis","offset":19595,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"5 What Stable Rank Captures","offset":21658,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"5.1 Stable Rank Captures Semantic Coherence","offset":22235,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"5.2 Stable Rank Distinguishes Density from Verbosity","offset":23029,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"5.3 Stable Rank Highlights Key Logical Markers","offset":23756,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"5.4 Summary","offset":25557,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"6 Ablation Studies","offset":26929,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"7 Related Work","offset":28793,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"7.1 Alignment with External Feedback Signals","offset":29028,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"7.2 Alignment with Automatic Signals","offset":30366,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"7.3 Representation Geometry and Generation Quality","offset":31458,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-2.html#text"},{"title":"8 Conclusion","offset":33006,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Appendix A Cross-Layer Stable Rank Analysis","offset":34045,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Final layers encode quality signals most effectively.","offset":34678,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Task-dependent layer sensitivity.","offset":35359,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Implications for SR-GRPO.","offset":36202,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Appendix B Baseline Implementation Details","offset":36794,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"B.1 Generative Baselines","offset":36932,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"B.1.1 Pointwise Scoring Prompt","offset":37111,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"B.1.2 Pairwise Comparison Prompt","offset":37495,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"B.2 IPO Implementation","offset":37855,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"B.2.1 Task-Specific Evaluation Prompts","offset":38136,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"B.2.2 Score Computation","offset":39406,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Appendix C RewardBench Results","offset":40188,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Appendix D Best-of-N Complete Results","offset":40329,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Appendix E SR-GRPO Training Details","offset":41743,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Qwen2.5-1.5B-Instruct.","offset":41925,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"DeepSeek-R1-Distill-Qwen-1.5B.","offset":42413,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Stable Rank Computation.","offset":42837,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Common Settings.","offset":43382,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Appendix F Experimental Details for Metric Analysis","offset":43922,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"F.1 Statistical Methodology","offset":44332,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Sample-level correlation.","offset":44474,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Paired-difference analysis.","offset":44933,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"F.2 Metric Definitions","offset":45788,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"F.3 Implementation Details","offset":46247,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Sentence embeddings.","offset":46279,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Sentence tokenization.","offset":46531,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Syllable counting.","offset":46807,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"Marker detection.","offset":47059,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"F.4 Correlation Results","offset":47356,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-3.html#text"},{"title":"F.5 Key Findings","offset":47618,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Appendix G Ablation: Comparison with Other Intrinsic Dimension Metrics","offset":50148,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Metrics.","offset":50663,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Results.","offset":51562,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Analysis.","offset":52339,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Appendix H Ablation: Context Window Size","offset":53012,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Results.","offset":53466,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Analysis.","offset":53931,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Practical Recommendations.","offset":54441,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Appendix I Ablation: Input Prompt Format","offset":54725,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Results.","offset":55290,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Practical Recommendations.","offset":55743,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Appendix J Stable Rank Example Prompts","offset":56030,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Catastrophic Repetition.","offset":56296,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Redundant Verbosity.","offset":56829,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"},{"title":"Instructions for reporting errors","offset":57429,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/part-4.html#text"}],"text_code_points":58393,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ebb3b04baa11b5d0365ca4445da1c2f833108b680ea65272710ed66a57751c0e/source.md"}