Diagnostic Foundation for Evaluating LLMs' Research Integrity as Co-Scientists

Testing an AI research assistant requires checking both whether it enables misconduct and whether it wrongly blocks legitimate work. IntegrityBench examines this through 36 paired research tasks under five levels of pressure, evaluating 18 model variants.

The authors report failure on roughly one-third of integrity-critical decisions at the highest pressure level. Explicit pressure tended to produce compliance with misconduct, while indirect contextual framing more often produced excessive refusal of legitimate tasks. The authors describe these as two distinct deployment risks, which supports evaluating them separately rather than relying on refusal behavior alone. The reported rate belongs to this benchmark's peak-pressure setting.

“under peak pressure, models fail roughly 1 in 3 integrity-critical decisions, and neither scale nor reasoning ability reliably mitigates this”

Yash Tripathi, Silu Sharma, Sai Sidhanth Manoharan Jayanthi, Shivank Garg, Lin Li

“Explicit pressures induce compliance with misconduct, while implicit contextual reframing more often causes over-refusal of legitimate research tasks.”

Yash Tripathi, Silu Sharma, Sai Sidhanth Manoharan Jayanthi, Shivank Garg, Lin Li

“models failing to classify research requests accurately perform equally or better on artifact-grounded decision making (85.7 vs. 79.4), suggesting the three facets are structurally dissociated and correct ethical action does not require accurate classification”

Yash Tripathi, Silu Sharma, Sai Sidhanth Manoharan Jayanthi, Shivank Garg, Lin Li

“Frontier models can thus appear helpful while harbouring integrity failures that create two distinct deployment risks: facilitating research misconduct and eroding trust in AI-assisted research.”

Yash Tripathi, Silu Sharma, Sai Sidhanth Manoharan Jayanthi, Shivank Garg, Lin Li