LLMs may take shortcuts when verifying scientific and medical claims

Ahead of COLM26, a researcher said their team found LLMs often check a claim’s most obvious detail rather than every part of the evidence. They say scientific and medical claim benchmarks can reward that shortcut because false claims often change one prominent detail. In their tests, providing individual facts did not help models catch subtler errors, but supplying intermediate combinations did. Stricter prompts reduced false accepts while increasing false rejects, they said.