跳到正文
arXiv cs.CL· Xiaopeng Chu, Wenzheng Kong, Wei Liu, Wenyi Zhang·· 23 小时前AI 评分31

自适应推理的全账本反事实审计:FLCVA 揭示轻量控制器仅识别当前错误

A Full-Ledger Counterfactual Audit of Adaptive Reasoning

AI 导读

研究者提出全账本反事实价值审计(FLCVA),在同一答案流程下比较立即作答与四次定长续写并计入策略所用信号成本。在 GSM8K 和 MATH-500 上,随机 256-token 续写可挽回 Qwen3-8B 26.6%、DeepSeek-R1-Distill-Qwen-7B 18.9% 的错误答案,同时仅损害 1.4% 和 1.9% 的正确答案。

正文 · 原文

View PDF HTML (experimental)

Abstract:Adaptive reasoning must decide whether to answer now or spend more compute. Predicting that the current answer is wrong does not establish whether further reasoning will improve it. We introduce the Full-Ledger Counterfactual Value Audit (FLCVA), which compares an immediate answer with four fixed-length continuations from the same frozen prefix under one answer procedure and accounts for the cost of every signal used by a policy. On GSM8K and MATH-500, a random 256-token continuation recovers a wrong answer in 26.6% of Qwen3-8B cases and 18.9% of DeepSeek-R1-Distill-Qwen-7B cases, while damaging a correct answer in 1.4% and 1.9%. Correctness labels from independently sampled continuations predict a separate continuation's outcome in both primary models. In offline replay, using these labels to allocate compute improves performance on DeepSeek; the Qwen3 gain remains inconclusive. Under full cost accounting, the tested lightweight, myopic controllers mainly recognize current errors and add little continuation-specific value beyond that shortcut. A Llama-8B replication and action/cost ablations examine the robustness of these findings. FLCVA separates measuring local continuation value, predicting it from observable signals, and using those predictions to allocate computation.
Comments: Accepted to EACL 2027 (Main Conference)
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Cite as: arXiv:2611.22778 [cs.AI]
  (or arXiv:2611.22778v1 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2611.22778

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Xiaopeng Chu [view email]
[v1] Tue, 17 Nov 2026 13:44:33 UTC (1,619 KB)

来源:arXiv cs.CL · arxiv.org