开放后训练配方如何改变 LLM 事实核查者相信谁
From Leniency to Suspicion: How an Open Post-Training Recipe Changes Whom LLM Fact-Checkers Believe
AI 导读
一篇 arXiv 论文研究了后训练如何让大语言模型在事实核查时因发帖者职业产生怀疑。结果显示只有开放的 Tülu-3 配方（OLMo-2 32B、Tülu-3 70B）会因发帖者职业地位低而更多判其为假，效应小于一个百分点但在独立声明集上可复现；Meta、阿里、Mistral 的配方无此现象。
正文 · 原文
Abstract:LLMs increasingly fact-check online content with the poster's identity in view. However, it remains unclear which post-training changes how much a model's truth judgement depends on the poster's social class, and how such a change happens inside the model. This paper investigates job-based doubt, i.e., judging a claim false more often when the poster has a lower-status job, using fact-checked claims with randomised poster profiles in base and post-trained checkpoints of six model families. Our results reveal that: 1) post-training creates such doubt only with the open Tülu-3 recipe (OLMo-2 32B, Tülu-3 70B), not with Meta's, Alibaba's or Mistral's; the effect is under one percentage point but statistically reliable, exceeding training drift and replicating on independent claim sets; 2) in OLMo-2 32B, the model does not start noticing the poster's job more, but starts interpreting it differently: a lower-status job pushes the base model toward "true" and the post-trained model toward "false"; and 3) this flip is carried by a single residual-stream direction and happens at supervised fine-tuning. The mechanism was confirmed on 400 new claims with pre-specified predictions. None of seven deployed commercial models doubts lower-status posters more. Our work underscores the need to audit speaker bias per training recipe.
| Comments: | 10 pages, 4 figures, 6 tables. Preprint, under review |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY) |
| Cite as: | arXiv:2610.38532 [cs.AI] |
| (or arXiv:2610.38532v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.38532 arXiv-issued DOI via DataCite |
Submission history
From: Yasir Zaki [view email]
[v1]
Wed, 28 Oct 2026 09:39:56 UTC (60 KB)
来源：arXiv cs.AI · arxiv.org