让大模型自己校准自信度，无需标注数据。
Self-Calibrating Language Models via Test-Time Discriminative Distillation
- 利用模型对答案正确性的判别概率作为无监督标签，实现测试时自校准。
- 在四个领域上将期望校准误差降低56%至78%，优于现有方法。
- 适合需要高置信度可靠性的部署场景，如医疗、金融决策支持。
大语言模型普遍存在过度自信问题：它们常在错误回答时仍表现出高置信度。现有校准方法或需标注验证数据，或在分布偏移下性能下降，或带来显著推理开销。近期研究表明，大模型内部已存在比其输出更可靠的校准信号——当被问及“该答案是否正确？”时，“True” token的概率 $P(\text{True})$ 比模型自述置信度表现更优，这一差距在理论上由生成误差不低于对应判别误差的两倍所支撑。我们提出 $ extbf{SECL}$（$ extbf{SE}$lf-$ extbf{C}$alibrating $ extbf{L}$anguage Models），一种测试时训练（TTT）流程，以该差距作为无标签自监督信号，完全无需标注数据或人工干预。SECL仅在输入分布偏移时进行适配，在6%–26%的问题流上训练，成本低于其所蒸馏的基线模型。在三个模型家族的四款小型语言模型及四个多样化领域中，SECL将期望校准误差（ECE）降低56%–78%，优于其自身监督信号，并匹配或超越近期推理时方法。SECL是首个将测试时训练应用于校准的方法；七组消融实验涵盖信号质量、门控策略、权重累积、损失设计、领域顺序、超参数敏感性和层选择，证实各组件均关键且在不同配置下稳健。代码：https://anonymous.4open.science/r/secl-emnlp26-submission-C890
原文摘要 · Abstract (English)
Large language models (LLMs) are systematically overconfident: they routinely express high certainty on questions they often answer incorrectly. Existing calibration methods either require labeled validation data, degrade under distribution shifts, or incur substantial inference costs. Recent work has shown that LLMs already contain a better-calibrated signal than the one they verbalize: the token probability of "True" when the model is asked "Is this answer correct?" ($P(\text{True})$) consistently outperforms their stated confidence, a gap that is theoretically grounded as generative error is lower-bounded by roughly twice the corresponding discriminative error. We introduce $\textbf{SECL}$ ($\textbf{SE}$lf-$\textbf{C}$alibrating $\textbf{L}$anguage Models), a test-time training (TTT) pipeline that exploits this gap as label-free self-supervision, requiring no labeled data or human supervision. SECL adapts only when the input distribution shifts, training on just 6--26% of the question stream at lower cost than the baseline it distills from. Across four small language models from three model families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56--78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods. SECL is the first method to apply TTT to calibration; seven ablations covering signal quality, gating strategy, weight accumulation, loss design, domain ordering, hyperparameter sensitivity, and layer selection confirm that each component is crucial and robust across configurations. Code: https://anonymous.4open.science/r/secl-emnlp26-submission-C890
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。