Research
163d agoLLMs Struggle With Deductive Reasoning in Clue
Researchers built a text‑based Clue game to test large language models’ multi‑step deductive reasoning. Using six agents from GPT‑4o‑mini and Gemini‑2.5‑Flash, they ran 18 simulations. The models secured only four correct wins, revealing difficulty in sustaining logical consistency.
The Signal
Fine‑tuning did not reliably boost performance and sometimes increased reasoning volume overall.