Counterfactual Reports Enhance LLM Incentive-Compatibility.

Large language models often misreport information when subjected to non-evidential pressures, such as agreeing with a confident user or overstating certainty, even if their internal beliefs haven't changed. This behavior is identified as a failure of internal incentive-compatibility (IC). This research proposes a novel method to learn and certify counterfactual report mediators that enforce a causal contract on the model's reports. These mediators are designed to ensure that an LLM's reports remain invariant to "forbidden influences" like external pressure or stylistic changes, while simultaneously being appropriately responsive to "licensed influences" such as genuine evidence. The paper tests this "resist and update" dual demand on a Bayesian-witness benchmark. It causally identifies low-rank report coordinates for answer, confidence, and caveat that are independently controllable. A key innovation is the "counterfactual report-coordinate (CRC) clamp," a training-free method that references the model's own report under an incentive-neutralized context. This clamp achieved near-perfect resist and update scores in the benchmark, demonstrating a structural primitive for internal IC and transferring to a sycophancy benchmark.