AIBearisharXiv – CS AI · 18h ago6/10
🧠
GIScholarBench: Benchmarking LLM Overconfidence in GIS Research
Researchers introduced GIScholarBench, a benchmark testing whether large language models exhibit overconfidence when performing academic research tasks. Evaluating Claude, Gemini, and ChatGPT on 10,865 GIS papers, the study found all models generate confident outputs even when knowledge is incomplete, particularly in citation generation and research ideation tasks.
🧠 ChatGPT🧠 Claude🧠 Sonnet