Arguments that Alter Minds: LLM Rationales Sway Human (and LLM) Notions of Plausibility
Shramay Palta, Peter Rankel, Sarah Wiegreffe, Rachel Rudinger
摘要
We investigate the degree to which human (and LLM) plausibility judgments of multiplechoice commonsense benchmark answers are subject to influence by (im)plausibility arguments for or against an answer, in particular, using rationales generated by LLMs. We collect 3, 000 plausibility judgments from humans and another 13, 600 judgments from LLMs. Overall, we observe increases and decreases in mean human plausibility ratings in the presence of LLM-generated PRO and CON rationales, respectively, suggesting that, on the whole, human judges find these rationales convincing. Experiments with LLMs reveal similar patterns of influence. Our findings demonstrate a novel use of LLMs for studying aspects of human cognition, while also raising practical concerns that, even in domains where humans are "experts" (i.e., common sense), LLMs have the potential to exert considerable influence on people's beliefs. 1 Pro Rationale: A bedroom is a private space where a car-less person can use a radio, smartphone, or other devices to tune into talk radio without external disturbances, ensuring an undisturbed listening experience. Additionally, bedrooms are typically associated with comfort and quiet, reinforcing the ability to focus on the content. Con Rationale: The bedroom is an implausible choice because it is a communal space shared with others in many households, which makes it difficult to ensure full privacy for listening to talk radio. Additionally, a person might not have access to a radio or leisure time in their own bedroom if they share living accommodations. Question: If a car-less person want to listen to talk radio in private, where might they listen to it? Choice: bedroom (gold label) Pro Rationale: A bedroom is a private space … Con Rationale: The bedroom is an implausible choice …
问问这篇 Paper
智能体会读完全文。
Lune 把这篇 Paper 索引到了每一个公式,引用它的顶会 Paper 也一样。你提问,回答直接引用原文。
它引用的顶会 Paper12
- LLM Evaluators Recognize and Favor Their Own GenerationsArjun Panickssery, Samuel R. Bowman, Shi FengNeurIPS 2024 · 被引用 865 次
- Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team PerformanceGagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok 等CHI 2021 · 被引用 713 次
- Abductive Commonsense ReasoningChandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi 等ICLR 2020 · 被引用 521 次
- Learning to Rationalize for Nonmonotonic Reasoning with Distant SupervisionFaeze Brahman, Vered Shwartz, Rachel Rudinger, Yejin ChoiAAAI 2021 · 被引用 46 次
- Back to the Future: Unsupervised Backprop-based Decoding for Counterfactual and Abductive Commonsense ReasoningLianhui Qin, Vered Shwartz, Peter West, Chandra Bhagavatula 等EMNLP 2020 · 被引用 25 次
相关 Paper
- Are Machine Rationales (Not) Useful to Humans? Measuring and Improving Human Utility of Free-text RationalesBrihi Joshi, Ziyi Liu, Sahana Ramnath, Aaron Chan 等ACL 2023 · 被引用 6 次
- Beyond Accuracy: Experts See AI Fact-Checks as Accurate but Less UsefulChenyan Jia, Apoorva Gondimalla, Angie Zhang, David Joseph Mullings 等CHI 2026 · 被引用 1 次
- Accommodation and Epistemic Vigilance: A Pragmatic Account of Why LLMs Fail to Challenge Harmful BeliefsMyra Cheng, Robert D. Hawkins, Dan JurafskyACL 2026 · 被引用 6 次
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMsLaura Ruis, Akbir Khan, Stella Biderman, Sara Hooker 等NeurIPS 2023 · 被引用 87 次
- Social Dynamics as Critical Vulnerabilities that Undermine Objective Decision-Making in LLM CollectivesChanggeon Ko, Jisu Shin, Hoyun Song, Huije Lee 等ACL 2026 · 被引用 1 次
