{"id":"arxiv:2503.02875","title":"The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models","date":"2025-03-04","section":"Post-training","quality":-0.063911,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/","source_url":"https://arxiv.org/abs/2503.02875","text_available":true,"authors":["Ke Ji","Jiahao Xu","Tian Liang","Qiuzhi Liu"],"outgoing":[{"source":"arxiv:2503.02875","target":"arxiv:2203.11171","evidence":"title","id":"arxiv:2203.11171","title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02a15b098d96a94d463bb470f59a34723abc34e59196f35c539f9796d8f3c3df/"},{"source":"arxiv:2503.02875","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:2503.02875","target":"arxiv:2501.19393","evidence":"id","id":"arxiv:2501.19393","title":"s1: Simple test-time scaling","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2a88818f7605b71cd0160831faebd51b3274c24ffd29626b4d519390b4bffffe/"},{"source":"arxiv:2503.02875","target":"arxiv:2502.03387","evidence":"id","id":"arxiv:2502.03387","title":"LIMO: Less is More for Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/714392528bb842fb3a61cee5367942f697536f0131b0fcac85c6eb8d2212125c/"}],"incoming":[],"text_parts":[{"start":0,"end":15679,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html"},{"start":15679,"end":31633,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html"},{"start":31633,"end":38794,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html"}],"contents":[{"title":"The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"Abstract","offset":129,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"1 Introduction","offset":1240,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"2 Prefix Self-Consistency","offset":4716,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"Early reasoning steps are highly consistent across reasoning trajectories.","offset":5462,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"Errors predominantly occur in later reasoning steps.","offset":6479,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"3 Methodology","offset":8780,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"3.1 Preliminary","offset":9127,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"3.2 Modeling Coverage and Accuracy","offset":11188,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"3.3 Unsupervised Prefix Fine-Tuning","offset":13818,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"Prefix Coverage and Accuracy","offset":14080,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-1.html#text"},{"title":"Sampling Only First Few Tokens","offset":16478,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Structure Tuning","offset":17832,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"4 Experiment","offset":19694,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"4.1 Experiments Setup","offset":19950,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Backbone LLMs","offset":19977,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Fine-Tuning","offset":20352,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Benchmarks","offset":21835,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Tuning Scenarios","offset":22091,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"4.2 Ablation Study","offset":22424,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Impact of Prefix Length","offset":22677,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Impact of Structure Tuning Ratio","offset":23683,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"4.3 Unsupervised Fine-Tuning","offset":24244,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"UPFT demonstrates superior performance compared to SFT in unsupervised fine-tuning.","offset":24618,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"The benefits of UPFT are more pronounced on complex reasoning tasks.","offset":25504,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"The U-Hard dataset maximizes UPFT’s potential through difficulty-focused curation.","offset":26175,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"UPFT achieves dramatic efficiency gains through minimal token training.","offset":27030,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"4.4 Supervised Fine-Tuning","offset":27711,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"UPFT achieves competitive performance with supervised methods while requiring significantly fewer tokens in both sampling and tuning.","offset":28175,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"UPFT offers seamless integration with label verification for enhanced accuracy.","offset":29693,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"5 Related Work","offset":30585,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-2.html#text"},{"title":"Unsupervised Learning","offset":33031,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html#text"},{"title":"Self-Training and Self-Improvement.","offset":34109,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html#text"},{"title":"6 Conclusion","offset":35344,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html#text"},{"title":"Appendix A Proof of Prefix Lower bound","offset":36329,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html#text"},{"title":"Appendix B Case study of Prefix Data","offset":37729,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html#text"},{"title":"Instructions for reporting errors","offset":37830,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/part-3.html#text"}],"text_code_points":38794,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b9349378a2ea40a7998ba364a4587c40dff3cc9b22750e14b3b21e296a4ecb12/source.md"}