跳到正文
arXiv cs.AI· Jan Kirin·· 2 小时前AI 评分35

循环语言模型的操作性原内省:在答案出来前读取正确性信号

Operational Proto-Introspection in Looped Language Models: Reading Correctness Before the Answer

AI 导读

Jan Kirin 在 arXiv 论文中研究能否在 2.6B 参数循环模型 Ouro-RLTT 推理过程中预测答案正误。仅读前 48 个 token 的隐藏状态,探针在合成逻辑任务上达到 AUROC 0.758,而 token 置信度仅 0.515;同一问题的两个答案中,探针 74% 的情况给正确者更高分。但该信号对难度可预测,且未能帮助模型生成更好答案,在 GSM8K 和第二检查点上结果未定。

正文 · 原文

View PDF HTML (experimental)

Abstract:Can one tell, while a language model reasons, whether its answer will be right? Looped models are a good place to ask: they run the same layers several times per token, leaving one hidden state per loop for a probe to read. I probe Ouro-RLTT, a 2.6B-parameter looped model, for what these states reveal about how its reasoning will end. I call such information operational proto-introspection: operational because it is defined by what a probe can read, and proto because nothing here shows that the model itself uses it.
In part, yes. Reading the first 48 tokens of reasoning, from answers that state no answer yet, a probe on the hidden states reaches an AUROC of 0.758 on a synthetic logic task, against 0.515 for token confidence. Question difficulty is a rival explanation: on 1,164 further questions, token confidence plus simple features of the question predict about as well. It is not only difficulty: between two answers to the same question, the probe scores the correct one higher 74% of the time. A fine-tuned language model reading the same tokens does about as well. On GSM8K and on a second checkpoint this within-question test is unresolved.
A probe that ranks finished candidate answers is weak at layer 8 on the first loop, much better on the second, and as accurate as the last layers by the third, on four Ouro checkpoints. The direction it reads is largely shared across loops: a probe trained on the last loop reads the first nearly as well as a probe trained there.
The signal helps decide what to do with finished outputs. It improves abstention over surface features and picks correct answers among candidates better than chance, though surface features pick about as well. Once several samples are finished, a baseline that knows their agreement leaves the answer-free probe nothing measurable to add. None of my attempts to use it to make the model generate better answers worked.
Comments: Version 4 is substantially rewritten and adds fixed-prefix correctness tests, comparisons against visible-text readers and between answers to the same question, and cross-loop probe transfer. It revises the interpretation of the earlier final-answer-marker tests, corrects the ranking and pruning analyses, and removes unsupported preference claims. Appendix I details the changes
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Cite as: arXiv:2607.18553 [cs.LG]
  (or arXiv:2607.18553v4 [cs.LG] for this version)
  https://doi.org/10.48550/arXiv.2607.18553

arXiv-issued DOI via DataCite

Submission history

From: Jan Kirin [view email]
[v1] Mon, 20 Jul 2026 22:40:36 UTC (375 KB)
[v2] Sat, 25 Jul 2026 23:42:03 UTC (681 KB)
[v3] Tue, 28 Jul 2026 16:28:19 UTC (864 KB)
[v4] Wed, 18 Nov 2026 17:33:48 UTC (179 KB)

来源:arXiv cs.AI · arxiv.org