Reliable Chain-of-Thought via Prefix Consistency

Project page: https://naoto-iwase.github.io/prefix-consistency-page Code: https://github.com/naoto-iwase/prefix-consistency Raw outputs and aggregated weighted-majority-voting (WMV) results used in the paper "Reliable Chain-of-Thought via Prefix Consistency". The archive covers all 5 models (GPT-OSS-20B, GPT-OSS-120B, Nemotron2-9B, Nemotron3-30B, Ministral3-14B) on all 4 benchmarks (AIME 2025, Brumo 2025, FrontierScience-Olympiad, HMMT Feb 2026) under self-judge, and the same 4 benchmarks under external-judge (GPT-OSS-120B as judge) for 4 of those 5 models (Ministral3-14B omitted on the external-judge side). Each cell contains the N initial samples, one regeneration per sample at the default tau=0.75 / K=1 setting, and aggregated WMV results for all evaluated methods. GPT-OSS-20B cells additionally include a 3x3 tau-K sensitivity grid and marker-based regenerations. Together with the code, this archive reproduces every figure and table in the paper without GPUs. See README.md inside the archive for the full layout. Note: model outputs included in the JSONLs are subject to each LLM's license or terms of use; please use under the most permissive applicable terms.