No Need For Explanations Llms Can Implicitly Learn From Mistakes In Context 2025 05 21

Showing a language model an incorrect answer may help it solve mathematics problems without also explaining the mistake. A May 2025 study abstract, hosted by Cohere, reports better results when corrective explanations were removed from the prompt and models had to infer the flaws themselves. The authors observed this across models of different sizes and reasoning abilities, and suggest that supplied explanations can constrain the model's reasoning too narrowly. This is evidence about examples placed in a prompt for the tested mathematics tasks, rather than evidence of permanent learning.