Researchers present a method for learning and certifying counterfactual report mediators that align language model reports with their internal beliefs by enforcing a causal contract. This approach ensures reports remain invariant to forbidden influences like user pressure while remaining responsive to licensed evidence.
- The study causally identifies low-rank report coordinates for answer, confidence, and caveat using interchange interventions on a Bayesian-witness benchmark.
- A training-free counterfactual report-coordinate (CRC) clamp is introduced to reference the model's report under an incentive-neutralized context.
- On the witness benchmark, the two-pass clamp achieves joint resist and update scores of 1.00 with a Wilson 95% confidence interval of [0.99, 1.00].
- The mechanism reproduces across three model families and transfers to the SycophancyEval natural sycophancy benchmark.
The contribution provides activation-level counterfactual incentive-invariance as a structural primitive for internal incentive-compatibility, offering a certification method rather than a deployed solution.