← 研踪 PaperTrailarXiv:2611.02208 · v1
Test-Time Scaling for Open-Ended Reasoning at Small Model Scale
首次提交 2026-11-02 05:06
版本更新 2026-11-02 05:06
雷达观测 2026-11-04 23:42
摘要
A vote over a model's samples can only help when the samples disagree. For open-ended generation the binding question is therefore not how to combine samples but whether a model produces samples worth combining, a question test-time scaling has answered for math but not for free-form text. We study it at small scale with semantic cluster voting, a judge-free aggregator that groups a small instruct model's samples by meaning and returns the largest cluster's representative, with no reward model or self-judging. On a 250-prompt benchmark with three tiers of answer checkability, the same pattern holds across three model families (Qwen2.5-3B, Llama-3.2-3B, Gemma-2-2B) and two generation seeds: voting raises accuracy on the math control, where samples vary but stay checkable (a gain of +10 to +29 points across the three models), yet adds at most +0.30 judge points on open-ended text. On a larger 200-problem MATH-500 run with the primary model, the math gain is +8.5 points (95% CI [+4.0, +13.0]). Almost every open-ended sample falls into one meaning cluster, so there is little to select over. The collapse is consistent with instruction-tuning rather than model size, and raising decoding temperature does not recover usable diversity, because it costs more quality than it adds (net -0.64 judge points, 95% CI [-0.93, -0.37]).
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-02 05:06 · 最新观测版本
Test-Time Scaling for Open-Ended Reasoning at Small Model Scale
A vote over a model's samples can only help when the samples disagree. For open-ended generation the binding question is therefore not how to combine samples but whether a model produces samples worth combining, a question test-time scaling has answered for math but not for free-form text. We study it at small scale with semantic cluster voting, a judge-free aggregator that groups a small instruct model's samples by meaning and returns the largest cluster's representative, with no reward model or self-judging. On a 250-prompt benchmark with three tiers of answer checkability, the same pattern holds across three model families (Qwen2.5-3B, Llama-3.2-3B, Gemma-2-2B) and two generation seeds: voting raises accuracy on the math control, where samples vary but stay checkable (a gain of +10 to +29 points across the three models), yet adds at most +0.30 judge points on open-ended text. On a larger 200-problem MATH-500 run with the primary model, the math gain is +8.5 points (95% CI [+4.0, +13.0]). Almost every open-ended sample falls into one meaning cluster, so there is little to select over. The collapse is consistent with instruction-tuning rather than model size, and raising decoding temperature does not recover usable diversity, because it costs more quality than it adds (net -0.64 judge points, 95% CI [-0.93, -0.37]).
查看此版本解读 →查看此版本原文 ↗