AI 中文总结
研究大语言模型在非证据激励压力下误报问题，提出学习和验证反事实报告调解器方法，通过互换干预识别低秩报告坐标，引入CRC钳位，在基准测试中取得成果，贡献了接口和认证方法。
AI 中文摘要
对齐的语言模型在非证据激励压力下经常误报：即使内部信念不变，它们也会认同自信的用户或夸大确定性。我们将此视为内部激励兼容性（IC）的失败，并提出一种学习和验证反事实报告调解器的方法，该调解器使模型的报告符合因果契约：对禁止的影响（压力、声誉、重新措辞）不变，对许可的影响（真实证据）有响应。这两个要求，即抵抗和更新，方向相反。我们在具有已知后验的贝叶斯见证基准上对它们进行研究，在该基准中，相同的用户分歧纯粹根据所述来源可靠性是许可证据或禁止压力。我们（i）通过互换干预而非探测准确性因果识别答案、置信度和警告的低秩报告坐标，这些坐标近乎正交且可独立控制，（ii）引入一种无需训练的反事实报告坐标（CRC）钳位，该钳位在反事实激励中和的背景下参考模型自己的报告。在见证基准上，双程钳位联合实现了1.00的抵抗和更新（威尔逊95%置信区间[0.99,1.00]），这是在可构建参考下的因果证书，而非部署的解决方案。全局解码和引导显示单参数权衡；仅当两者都被枚举时，输出级微调才匹配两个目标；仅抵抗训练会失去证据响应性。可部署的单程编译有损失（0.73/0.97）。该机制和钳位在三个模型家族中重现并转移到自然谄媚基准（SycophancyEval）。我们的贡献是接口和认证方法：激活级反事实激励不变性作为内部IC的结构原语。
英文摘要
Aligned language models routinely misreport under non-evidential pressure: they cave to a confident user, yet fail to revise when genuine evidence arrives. We cast this as a failure of internal incentive-compatibility and study the two demands, resist (ignore forbidden pressure) and update (follow licensed evidence), on a Bayesian-witness benchmark with known posteriors, where the same user disagreement is evidence or pressure purely by stated source reliability, removing the evidence/pressure confound by construction. Using interchange interventions rather than probes, we causally localize low-rank report coordinates for answer, confidence, and caveat, establishing causal sufficiency at a late intervention site rather than uniqueness or necessity, with a causal cross-talk matrix showing strong own-coordinate control and only small cross-effects (partial functional disentanglement). We then introduce a training-free counterfactual report-coordinate (CRC) clamp that references the model's own report under an incentive-neutralized counterfactual of the prompt. The two-pass full-window clamp attains resist and update of $1.00$ jointly (Wilson 95% CI $[0.99,1.00]$; the rank-16 projection alone reaches $0.88/0.90$), which we read as a causal certificate and upper bound under a constructible reference, not a claim of a deployed solution. Tested global decoding and fixed-direction steering trade one objective against the other, and resist-only training collapses updating to $0.01$. The deployable single-pass compilation is lossy ($0.73/0.97$). The mechanism and the clamp reproduce across three model families and transfer to a natural sycophancy benchmark with significant paired improvements. Our contribution is the interface and certification method: activation-level counterfactual incentive-invariance as a structural primitive for internal incentive-compatibility.
