LLM counterfactual reasoning hindered by intuitive biases, study finds

A new research paper explores how large language models (LLMs) handle counterfactual reasoning in policy evaluation, finding that "intuitiveness" of a case significantly impacts performance. Models struggle more with counter-intuitive findings, even when using advanced prompting techniques like chain-of-thought. This suggests LLMs may mimic deliberative reasoning without fully overcoming inherent biases or intuitive priors. AI

IMPACT LLMs may struggle with real-world policy evaluation when findings contradict common intuition, indicating a need for improved reasoning beyond surface-level deliberation.

RANK_REASON Research paper published on arXiv detailing findings about LLM reasoning capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• chain-of-thought
• LLM
• Yanjie He

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →