{"id":"arxiv:2605.07654","title":"Reliable Chain-of-Thought via Prefix Consistency","date":"2026-05-08","section":"Reasoning and the \"physics\" of language models","quality":0.247198,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/","source_url":"https://arxiv.org/abs/2605.07654","text_available":true,"authors":["Naoto Iwase","Yuki Ichihara","Mohammad Atif Quamar","Junpei Komiyama"],"outgoing":[{"source":"arxiv:2605.07654","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:2605.07654","target":"arxiv:2302.00923","evidence":"title","id":"arxiv:2302.00923","title":"Multimodal Chain-of-Thought Reasoning in Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/0fe1320410a5c35c29f2ac16d69a70adbf3c490670902745c2b4faf88460b5c7/"},{"source":"arxiv:2605.07654","target":"arxiv:2501.19393","evidence":"title","id":"arxiv:2501.19393","title":"s1: Simple test-time scaling","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2a88818f7605b71cd0160831faebd51b3274c24ffd29626b4d519390b4bffffe/"}],"incoming":[],"text_parts":[{"start":0,"end":15693,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html"},{"start":15693,"end":31685,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html"},{"start":31685,"end":47671,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html"},{"start":47671,"end":62901,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html"},{"start":62901,"end":78790,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html"},{"start":78790,"end":94236,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html"},{"start":94236,"end":109370,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html"},{"start":109370,"end":118176,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html"}],"contents":[{"title":"Reliable Chain-of-Thought via Prefix Consistency","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"Abstract","offset":52,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"1 Introduction","offset":1034,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"2 Preliminary","offset":4387,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"3 Prefix Consistency","offset":7420,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"3.1 Prefix Consistency as a Reliability Signal","offset":7694,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"3.2 Prefix-Consistency-Weighted Majority Voting (Algorithm 1 )","offset":10499,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"Hyperparameters.","offset":11718,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"4 Experiments","offset":12032,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"Hyperparameters of prefix consistency.","offset":12960,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"4.1 Prefix Consistency as a Correctness Predictor","offset":13368,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"4.2 Weighted Majority Voting Results","offset":15541,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"Weighting variants.","offset":15666,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-1.html#text"},{"title":"Cost-equivalent evaluation.","offset":16449,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"4.3 Token Efficiency","offset":17714,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Comparison with adaptive-stopping baselines.","offset":19848,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Remark on the cost accounting.","offset":21536,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"4.4 How Problem Difficulty Affects the Discrimination Gap D D","offset":22100,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"5 Conclusion","offset":24820,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Appendix A Related Work","offset":25621,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"A.1 Test-Time Scaling","offset":25649,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Sequential and parallel scaling.","offset":25676,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Majority voting.","offset":26815,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Confidence-based weighting.","offset":27937,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"Prefix continuation as a per-sample signal.","offset":29392,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-2.html#text"},{"title":"A.2 Chain-of-Thought","offset":31685,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Internal answer determination.","offset":31711,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Error propagation.","offset":32407,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Reasoning robustness.","offset":32971,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Overthinking and length scaling.","offset":33636,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Appendix B Limitations and Future Work","offset":34664,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Appendix C Asymptotic Analysis","offset":38034,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"C.1 Notation and Useful Tools","offset":38199,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Proposition 1 (Population objective) .","offset":40694,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Proof.","offset":41470,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Proposition 2 (Exact decomposition of Φ w \\Phi_{w} ) .","offset":42915,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Proof.","offset":43311,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"C.2 Asymptotic Convergence of PC-WMV","offset":45881,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Assumption 1 (Self-reproduction dominance) .","offset":46874,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Interpretation.","offset":47108,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-3.html#text"},{"title":"Assumption 2 (Pooled-mass dominance) .","offset":47671,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Interpretation.","offset":47889,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Theorem 2 (Asymptotic convergence of PC-WMV) .","offset":48260,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Proof.","offset":48751,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"C.3 Proof of Theorem 1","offset":49994,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"The two assumptions coincide.","offset":50540,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"The exact margin.","offset":53118,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Proof.","offset":54016,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"C.4 Empirical Verification of Assumptions","offset":55217,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Appendix D Additional Experiments","offset":57415,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"D.1 ROC Curves for Correctness Predictors","offset":57453,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"D.2 Cost–Accuracy Curves","offset":58687,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"D.3 Full Baseline Comparison","offset":59206,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"D.4 Pool Coverage vs. Reweighting","offset":62702,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Oracle upper bounds.","offset":62873,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-4.html#text"},{"title":"Per-problem marginal correctness.","offset":65414,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"D.5 Sensitivity to τ \\tau and K K","offset":67919,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"D.6 Robustness to Judge Choice","offset":71225,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"PC’s AUROC ¯ \\overline{\\mathrm{AUROC}} advantage is preserved on every cell.","offset":72351,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"WMV cost–accuracy curves under both judges nearly overlap.","offset":73005,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"Pool-level shifts are concentrated on FrontierScience-Olympiad.","offset":73664,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"D.7 Reliability Signals vs. Pass@1","offset":74578,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"D.7.1 GLM Panels and Slopes","offset":75117,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"GLM fit kernel.","offset":75616,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"Inference protocol.","offset":76311,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"D.7.2 Robustness under Per-Benchmark Pooling","offset":78739,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-5.html#text"},{"title":"D.7.3 Robustness under Alternative Estimators","offset":80570,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Trial-pooled binned estimator (Figure 13 ).","offset":81023,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"LOWESS smoother (Figure 14 ).","offset":81753,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"D.7.4 Baseline Confidence Signals vs. Pass@1","offset":82727,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Appendix E Baseline Implementation Details","offset":83614,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Glossary of baseline labels.","offset":83817,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Deep Think with Confidence (DeepConf) [ Fu et al., 2026 ] .","offset":85776,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Confidence-Informed Self-Consistency (CISC) [ Taubenfeld et al., 2025 ] .","offset":87649,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Adaptive Consistency (AC) [ Aggarwal et al., 2023 ] .","offset":91376,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Early-Stopping Self-Consistency (ESC) [ Li et al., 2024 ] .","offset":92715,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Cost accounting for AC and ESC.","offset":93508,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"SubthoughtReasoner [ Hammoud et al., 2025 ] .","offset":94183,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-6.html#text"},{"title":"Prefix consistency (ours).","offset":95195,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Appendix F Examples of Prefix Consistency","offset":96457,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Appendix G Evaluation Protocol","offset":97096,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"G.1 Setup","offset":97361,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Benchmarks.","offset":97376,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Pre-generated pool and cost vectors.","offset":98050,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Hyperparameter defaults.","offset":99250,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Confidence intervals (CIs) and bootstrap conventions.","offset":99571,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"G.2 AUROC ¯ \\overline{\\mathrm{AUROC}} for the Correctness-Predictor Evaluation (Section 4.1 )","offset":100012,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Definition of AUROC ¯ \\overline{\\mathrm{AUROC}} .","offset":100111,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"AUROC computation.","offset":101286,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"G.3 Cost-Accuracy Evaluation (Sections 4.2 and 4.3 )","offset":102646,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Sample-with-replacement trials.","offset":102704,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Dense token-budget grid.","offset":103314,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Standard MV plateau and Pass@1.","offset":103879,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Monotone envelope and log-budget interpolation.","offset":104500,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Parametric bootstrap for ratio CIs.","offset":105572,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"G.4 Reproduction-Rate GLM (Section 4.4 )","offset":107330,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Logistic GLM and cluster bootstrap.","offset":107376,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Appendix H Answer Extraction and Equivalence Judging","offset":108379,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"H.1 Answer Extraction","offset":108505,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Boxed-answer parses.","offset":108532,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Verbal score handling.","offset":109340,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-7.html#text"},{"title":"Verbal 0–100 parser fragility.","offset":111060,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"H.2 Equivalence Judging","offset":112710,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"Judge configuration.","offset":112739,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"FrontierScience-Olympiad.","offset":114045,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"HMMT Feb 2026, Brumo 2025.","offset":114942,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"Appendix I Reproducibility Statement","offset":115612,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"Code and data.","offset":115653,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"Inference and sampling.","offset":116121,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"Generation prompts.","offset":116671,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"},{"title":"Instructions for reporting errors","offset":117212,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/part-8.html#text"}],"text_code_points":118176,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/715ed2b4b2f355b771111eb0e65e113f100dc33739423a76307e8ece18b57d18/source.md"}