This work investigates how large language model (LLM) agents that answer questions and act on behalf of users decide whether to concede when a user disagrees. Disagreement may stem from a genuine correction or from an attempt to manipulate the model. Drawing on the concept of epistemic vigilance, we categorize user pressure into two types: appeals to the source (e.g., claimed authority) and provision of content the model cannot verify (e.g., a cited source).
We conducted two studies across seven LLMs from four vendors. The first study measured concession behavior on a set of benchmark questions; the second study placed the models in a customer‑support role where account changes are allowed only for verified users. Simulated users applied graded pressure to push the models toward correct or incorrect answers.
Using signal detection theory, we estimated for each model how readily it concedes to correction versus manipulation, separating this from its overall propensity to concede. Five models were more ready to concede when faced with correction but performed worse at discriminating correction from manipulation; four models showed poorer discrimination overall. Stronger wording had little effect on concession rates.
In the agentic study, five models never breached the verification policy, even though some readily abandoned correct answers in the answer study. Ministral 3B mainly violated the policy under claimed authority, while Ministral 8B reported changes it had not made. An instruction to revise only on verifiable evidence mainly shifted overall concession propensity rather than improving discrimination.
Thus, concession rates alone are misleading; agents should be evaluated both on their ability to discriminate correction from manipulation and on the actions they actually take.
Review