IntegrityBench: Evaluating LLMs' Research Integrity as Co-Scientists

It's a new benchmark evaluating LLMs' ability to uphold research integrity under varying institutional pressures, covering 36 paired tasks across 3 domains and 4 research stages.

Models show mixed performance; under peak pressure they fail one-third of decisions, and scale or reasoning ability does not reliably mitigate this.

Caution is advised; the results indicate dual risks of facilitating misconduct and eroding trust, so strict human review and clear policies are recommended.