arXiv cs.AI· Ming Shen, Tianjun Yao·· 12 小时前AI 评分42
并非所有测试时计算都是证据：LLM 推理中的模型特定偏差
Not All Test-Time Compute Is Evidence: Model-Specific Bias in LLM Reasoning
AI 导读
研究指出，同一推理模型重复采样、加深推理深度及自我评估产生的错误彼此相关，测试时扩展所假设的“更多计算即更多独立证据”并不成立。为此提出 DIA（Disagreement-triggered Independent Arbitration），将中间阶段与最终阶段预测的分歧视为推理不稳定的逐题信号，交由一组异构小模型在不见候选答案的情况下独立求解并直接聚合。
正文 · 原文
Abstract:Test-time scaling methods implicitly assume that additional computation from the same model provides independent evidence. We show that this assumption fails within a single reasoning model: errors are correlated across repeated samples, across reasoning depth, and in the model's own self-evaluation, so more compute from the same model is not more evidence. We introduce DIA (Disagreement-triggered Independent Arbitration), which treats disagreement between an intermediate-stage and the final-stage prediction as a per-question signal of reasoning instability and resolves it by asking a small panel of heterogeneous models to solve the question independently, without seeing either candidate, and aggregating their answers directly. Across seven benchmarks and two base models, DIA outperforms full chain-of-thought in 12 of 14 settings, including a 7.00% gain on MuSiQue, while activating on only $\sim 17\%$ of examples and using up to 47.2% fewer tokens than a multi-agent baseline. Ablations show that both independent solving and solver heterogeneity are necessary for these gains.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.23213 [cs.AI] |
| (or arXiv:2610.23213v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.23213 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ming Shen [view email]
[v1]
Fri, 16 Oct 2026 09:48:02 UTC (1,352 KB)
来源：arXiv cs.AI · arxiv.org