A new research paper explores how Large Language Models (LLMs) make decisions when presented with past case examples. The study found that LLMs like DeepSeek V4.1 Flash, GPT-5.6 Luna, and Claude Haiku 5.5 are significantly influenced by the recorded decisions of past cases, with a shift from denial to approval in precedents increasing approval rates by 30-55%. Conversely, the factual content of past cases appears to decrease approval rates, suggesting that LLMs may perceive the current request as weaker when juxtaposed with favorable facts. This effect was observed across various policies and models, highlighting the importance of evaluating the factual and decision components of precedent data separately in LLM-driven decision systems. AI
IMPACT Highlights potential biases in LLM decision-making based on how precedent data is presented, impacting applications in areas like customer service and legal analysis.
RANK_REASON Research paper detailing LLM decision-making processes. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
