{"id":"arxiv:2601.11061","title":"Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs","date":"2026-01-16","section":"Post-training","quality":0.396869,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/","source_url":"https://arxiv.org/abs/2601.11061","text_available":true,"authors":["Lecheng Yan","Ruizhe Li","Guanhua Chen","Qing Li"],"outgoing":[{"source":"arxiv:2601.11061","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"},{"source":"arxiv:2601.11061","target":"arxiv:2501.00656","evidence":"id","id":"arxiv:2501.00656","title":"2 OLMo 2 Furious","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1ca88dba5d5c59e8cb23b2a304e74370b8ec69077b714a30d426300f7ef2a0b9/"},{"source":"arxiv:2601.11061","target":"arxiv:2504.13837","evidence":"id","id":"arxiv:2504.13837","title":"Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b90f70269aeb02311d74ffaadf47e139683424a47118dc18619000ba72174d17/"},{"source":"arxiv:2601.11061","target":"arxiv:2506.10947","evidence":"id","id":"arxiv:2506.10947","title":"Spurious Rewards: Rethinking Training Signals in RLVR","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f48e13a9656d8eff1882eb80988e07a820c4799401298947deac735c004e8592/"}],"incoming":[],"text_parts":[{"start":0,"end":15885,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html"},{"start":15885,"end":30194,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-2.html"},{"start":30194,"end":45778,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html"},{"start":45778,"end":56464,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html"}],"contents":[{"title":"Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"Abstract","offset":109,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"Keywords:","offset":1366,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"1 Introduction","offset":1384,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"2 Related Work","offset":8307,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"3 Preliminaries","offset":10909,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"4 Empirical Study","offset":13046,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"4.1 Experiment Setting","offset":13068,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"4.2 What Does Perplexity Show","offset":15340,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-1.html#text"},{"title":"4.3 Locating the Functional Anchor: Static Analysis","offset":16575,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-2.html#text"},{"title":"4.4 Tracking the Bifurcation: Dynamic Analysis","offset":23188,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-2.html#text"},{"title":"5 Ablation Study","offset":25785,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-2.html#text"},{"title":"6 Mechanistic Intervention","offset":29825,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-2.html#text"},{"title":"6.1 Task-Relevant Neuron Identification","offset":30149,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-2.html#text"},{"title":"6.2 Causal Steering","offset":32258,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"6.3 Results","offset":33158,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"6.4 Probe-Gated Shortcut Audit","offset":35147,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"7 Conclusion","offset":36733,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"Acknowledgments","offset":38230,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"Impact Statement","offset":38410,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"Appendix A Code","offset":40156,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"Appendix B Preliminaries","offset":40267,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"Appendix C Partial Prompt Evaluation Details","offset":45398,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-3.html#text"},{"title":"Appendix D Complete Perplexity Analysis","offset":47711,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Appendix E Additional Path Patching Results","offset":48559,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Appendix F Additional JSD Results","offset":49297,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Appendix G Additional Results on Logit Lens","offset":50406,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Appendix H Additional Results on Ablation Study","offset":52548,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Appendix I Additional Results on Mechanistic Intervention","offset":54404,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Appendix J Data Sample Examples","offset":55124,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.1 MATH-500 Question 42","offset":55287,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.2 MATH-500 Question 65","offset":55317,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.3 MATH-500 Question 332","offset":55347,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.4 MATH-500 Question 77","offset":55378,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.5 MATH-500 Question 141","offset":55408,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.6 MATH-500 Question 19","offset":55439,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"J.7 MATH-500 Question 213","offset":55469,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"},{"title":"Instructions for reporting errors","offset":55500,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/part-4.html#text"}],"text_code_points":56464,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/96002027cf4c274573f7320398481a0cd73dbf72f18d7f37c783605f5158ea7b/source.md"}