arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.01528cs.CYcs.LG
收藏

消除偏见的可预测性:通过强化学习训练鲁棒的大语言模型推理

Making Bias Non-Predictive: Training Robust LLM Reasoning via Reinforcement Learning

Qian Wang, Xuandong Zhao, Zirui Zhang, Zhanzhi Lou, Nuo Chen, Dawn Song, Bingsheng He

AI总结:

本文提出Epistemic Independence Training (EIT)框架,通过使偏见提示非预测性来提升大语言模型的推理鲁棒性,实验表明其在对抗性偏见下表现更优,并能泛化至多种偏见类型。

摘要:

大型语言模型(LLMs)日益成为推理和自动化评估工具,但它们仍易受认知偏见影响,特别是在面对虚假提示级线索(如共识主张或权威诉求)时会改变推理方式。现有通过提示或监督微调的缓解方法无法泛化,因为它们只修改表层行为而不改变使偏见线索吸引人的优化目标。我们提出Epistemic Independence Training (EIT),一种基于关键原则的强化学习框架:为了学习独立性,偏见提示必须成为奖励的非预测性因素。EIT通过平衡冲突策略实现这一点,其中偏见信号等可能支持正确或错误答案,结合一种惩罚跟随偏见而不奖励偏见同意的奖励设计。在Qwen3-4B上的实验表明,EIT在对抗性偏见下同时提高了准确性和鲁棒性,同时在偏见与事实一致时保持性能。值得注意的是,仅在带wagon偏见上训练的模型能泛化到未见过的偏见类型,如权威和干扰,表明EIT诱导的是可转移的epistemic独立性而非偏见特定的启发式方法。EIT进一步在多个基准(MedQA, HellaSwag)、模型家族(Llama-3.2-3B)和规模(Qwen3-8B)上泛化,并在不需环境标签的情况下优于分布偏移方法(GroupDRO, IRM)。代码和数据可在https://anonymous.4open.science/r/bias-mitigation-with-rl-BC47获取。

英文摘要:

Large language models (LLMs) increasingly serve as reasoners and automated evaluators, yet they remain susceptible to cognitive biases -- often altering their reasoning when faced with spurious prompt-level cues such as consensus claims or authority appeals.} Existing mitigations via prompting or supervised fine-tuning fail to generalize, as they modify surface behavior without changing the optimization objective that makes bias cues attractive. We propose \textbf{Epistemic Independence Training (EIT)}, a reinforcement learning framework grounded in a key principle: to learn independence, bias cues must be made non-predictive of reward. EIT operationalizes this through a balanced conflict strategy where bias signals are equally likely to support correct and incorrect answers, combined with a reward design that penalizes bias-following without rewarding bias agreement. Experiments on Qwen3-4B demonstrate that EIT improves both accuracy and robustness under adversarial biases, while preserving performance when bias aligns with truth. Notably, models trained only on bandwagon bias generalize to unseen bias types such as authority and distraction, indicating that EIT induces transferable epistemic independence rather than bias-specific heuristics. \revised{EIT further generalizes across benchmarks (MedQA, HellaSwag), model families (Llama-3.2-3B), and scales (Qwen3-8B), and outperforms distribution-shift methods (GroupDRO, IRM) without requiring environment labels.} Code and data are available at https://anonymous.4open.science/r/bias-mitigation-with-rl-BC47

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑