{"id":"arxiv:2510.14901","title":"Reasoning with Sampling: Your Base Model is Smarter Than You Think","date":"2025-10-16","section":"Post-training","quality":0.212422,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/","source_url":"https://arxiv.org/abs/2510.14901","text_available":true,"authors":["Aayush Karan","Yilun Du"],"outgoing":[{"source":"arxiv:2510.14901","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:2510.14901","target":"arxiv:2505.19590","evidence":"id","id":"arxiv:2505.19590","title":"Learning to Reason without External Rewards","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d8df33b706906a7abf4e660f54c75a81ffe2caaec8b0a10d1a6ad6ec63060832/"},{"source":"arxiv:2510.14901","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":[{"source":"arxiv:2510.11686","target":"arxiv:2510.14901","evidence":"id","id":"arxiv:2510.11686","title":"Representation-Based Exploration for Language Models: From Test-Time to Post-Training","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/89b1f9c1bcd283b28d316cb6cddafdd9361be5cd510b2be2a1e761c20a0fd62e/"},{"source":"arxiv:2605.06241","target":"arxiv:2510.14901","evidence":"id","id":"arxiv:2605.06241","title":"Rethinking RL for LLM Reasoning: It's Sparse Policy Selection, Not Capability Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c9e4fdd6e0c24f5a635ff70c2bca19da19c4012a05f69ceb20fdc0a86fd8b775/"},{"source":"arxiv:2610.02140","target":"arxiv:2510.14901","evidence":"id","id":"arxiv:2610.02140","title":"Finetuning with Sampling: SFT Learns Better Than You Think","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/"}],"text_parts":[{"start":0,"end":15823,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html"},{"start":15823,"end":31789,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html"},{"start":31789,"end":43994,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html"}],"contents":[{"title":"Reasoning with Sampling: Your Base Model is Smarter Than You Think","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Abstract","offset":70,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"1 Introduction","offset":1382,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"2 Related Works","offset":5616,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Reinforcement learning for LLMs.","offset":5636,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Autoregressive MCMC sampling with LLMs.","offset":6792,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Annealed sampling for diffusion.","offset":7812,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"3 Preliminaries","offset":9202,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"4 MCMC Sampling for Power Distributions","offset":10376,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"4.1 Reasoning with Power Distributions","offset":11424,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Proposition 1 .","offset":12820,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Proof.","offset":12931,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Observation 1 .","offset":15602,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"Example 1 .","offset":15803,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-1.html#text"},{"title":"4.2 The Metropolis-Hastings Algorithm","offset":18411,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Definition 1 .","offset":20361,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"4.3 Power Sampling with Autoregressive MCMC","offset":22153,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"5 Experiments","offset":25974,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"5.1 Experimental Setup","offset":25992,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Evaluation.","offset":26020,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Models.","offset":27774,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Sampling Algorithm.","offset":28389,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"5.2 Results","offset":28966,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Main results.","offset":28983,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"5.3 Analysis","offset":29948,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Reasoning trace likelihoods and confidences.","offset":30130,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Reasoning trace lengths.","offset":31315,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-2.html#text"},{"title":"Diversity and pass@ k k performance.","offset":31789,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"The effect of power distributions.","offset":32866,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Test-time scaling with MCMC steps.","offset":33903,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"6 Conclusion","offset":35758,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"7 Acknowledgments","offset":36862,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Appendix A Appendix","offset":37006,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"A.1 Additional Theoretical Discussion","offset":37030,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Proposition 2 ( Informal ) .","offset":37269,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Definition 2 .","offset":37490,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Definition 3 .","offset":38116,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Definition 4 .","offset":38470,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Proposition 3 .","offset":39468,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Proof.","offset":40060,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"A.2 Pass@k Accuracies over Multiple Domains","offset":41639,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"A.3 More Qualitative Examples","offset":42995,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"},{"title":"Instructions for reporting errors","offset":43030,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/part-3.html#text"}],"text_code_points":43994,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/source.md"}