arXiv cs.CL· Vladimir Kazei·· 11 小时前AI 评分38
小语言模型幻觉与拒答的几何结构：ConfabQA 基准与隐藏状态干预
Geometry of Confabulation and Refusal in Small Language Models
AI 导读
研究者提出 ConfabQA 基准（784 项），用线性探针、稀疏自编码器（SAE）和直接优化干预方向刻画小模型首个生成 token 前隐藏状态的几何结构。在 Qwen3-1.7B 和 Gemma 2 2B 上，探针区分即将拒答与幻觉的准确率比提示词特征基线高 7.4/16.3 个百分点；沿拒答向量方向施加不超过状态自身范数的推动，可将多数幻觉翻转为以“As...”开头的拒答。
正文 · 原文
Abstract:Faced with an unanswerable question, small language models can either refuse or confabulate: confidently produce a wrong answer. Can the hidden state before the first generated token provide a clue? Can we flip confabulations into refusals? I introduce ConfabQA, a 784-item benchmark whose obscure-pre-cutoff category breaks the typical correctness/knowledge-cutoff confound and map the geometry of the hidden state with linear probes, a sparse autoencoder (SAE), and direct optimization of intervention directions. On Qwen3-1.7B and Gemma 2 2B the probe separates upcoming refusals from confabulations well beyond prompt features (+7.4/+16.3 pp); on Llama 3.2 3B the ConfabQA margin is within fold-to-fold spread, but a direct refusal probe on PopQA and TriviaQA recovers +18-25 pp. On Qwen I closely inspect three candidate directions for the refusal vector: probe-recovered, Qwen-Scope SAE feature, and directly optimized. Although the three directions are pairwise 50-81 degrees apart, a push at or below the state's own norm along any of them flips most confabulations into refusal openings dominated by the "As..." family. Forcing the opener without touching the hidden state yields the same one-in-three judged-refusal rate as any push. The opposite push collapses refusals mostly into confabulations, yet leaves originally correct answers intact, so the direction gates the opening speech act rather than the knowledge behind it. Confabulations do not resolve into a concentrated feature in a dictionary or a token signature on Qwen3-1.7B. Scoring every probe on the margin over the strongest of four prompt-feature baselines extracts the hidden state's own contribution. Correctness signal is model-dependent and portable only for Llama 3.2 3B: 81% of the probe's margin over test majority survives transfer between ConfabQA, PopQA, and TriviaQA. Benchmark, judge, and code are released.
| Comments: | 35 pages, 16 figures, 19 tables. Code, data, and the ConfabQA benchmark: this https URL |
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.01272 [cs.LG] |
| (or arXiv:2611.01272v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2611.01272 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Vladimir Kazei [view email]
[v1]
Sat, 31 Oct 2026 21:14:26 UTC (2,935 KB)
来源：arXiv cs.CL · arxiv.org