arXiv 论文检验幻觉的结构成因与数据的放大作用
From Architecture to Output: Testing the Structural Origins of Hallucination and the Amplifying Role of Data
AI 导读
论文通过干预实验检验幻觉的结构性成因，用 130M 到 2.8B 的五组 Pythia/Mamba 配对及四个注意力模型，在 111 条事实的植入真假陈述上测试。
正文 · 原文
Abstract:Surveys of hallucination classify errors by how an output relates to its source, and often explain them by the model's structure: attention retrieves by association, the maximum-likelihood objective has no term for truth, and left-to-right decoding commits to early tokens. Data, on this account, makes these failures worse rather than causing them. The account is usually argued rather than tested. We test it with interventions that need only sampling access. Study 1 holds pretraining data, tokenizer and scale fixed and swaps attention for a state-space mixer, using five Pythia/Mamba pairs from 130M to 2.8B parameters and four further attention models, on 111 facts presented with a planted false or a planted true statement. Swapping the mixer leaves free-recall failure unchanged (both 2.8B models answer 3 of 200 SimpleQA questions), but it changes what happens to a defect in the input. From 410M upwards, every attention model repeats the planted falsehood more often than its matched state-space model (2.8B: 650 vs. 566 of 666 items, p = 2.5e-20), also on facts both models know, while the state-space models can still read planted facts they did not know. Study 2 replaces the first wrong fact in 111 erroneous answers from three instruction-tuned models. The correct fact cuts the share of answers with a downstream error from 92.5% to 48.1%, but a different wrong fact works as well (p = 0.38), and the models give the right answer in isolation on only 2 of 111 items. The structural account survives in a narrower form. The mixing component decides whether a defect in the context becomes an output error, commitment decides how an error spreads, and on long-tail questions the first error comes from missing knowledge, not from a suppressed truth.
| Comments: | 16 pages, 4 figures, 4 tables. Substantially revised: adds a matched-architecture study (14 models) and corrected statistics for the prefix-repair study. Data and code in ancillary files |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| ACM classes: | I.2.7; I.2.6 |
| Cite as: | arXiv:2606.07537 [cs.CL] |
| (or arXiv:2606.07537v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.07537 arXiv-issued DOI via DataCite |
Submission history
From: Md. Rejaul Korim Sadi [view email]
[v1]
Wed, 29 Apr 2026 11:34:09 UTC (1,311 KB)
[v2]
Fri, 4 Sep 2026 15:29:54 UTC (103 KB)
[v3]
Mon, 9 Nov 2026 21:20:10 UTC (5,559 KB)
来源：arXiv cs.AI · arxiv.org