Reading Calibrated Uncertainty from Language Model Trajectories

Aliai Eusebi, Alexander Herzog, Xiaoyu Liang, Marie Vasek, Enrico Mariconti, Lorenzo Cavallaro

arXiv:2605.22864 · 2026-05-25 공개 · arXiv · PDF

llm calibration trajectory-analysis error-localization softmax-uncertainty activation-features sparse-probing mlp-updates

Abstract

The maximum softmax probability (MSP) represents a default approach when evaluating uncertainty quantification for language model generation with structured output. Although cheap, it is often miscalibrated. Methods that probe the model's internal activations feed raw hidden states into opaque classifiers, reading activations as static snapshots and leaving implicit the layer-wise trajectory by which a representation is formed. Yet, similar endpoints can arise from very different paths, and how evidence accumulates, reinforces, or reverses across depth might reveal uncertainty that final probabilities obscure. We extract eleven scale-invariant geometric features, tracing the cumulative path of per-layer MLP updates, and feed them to a sparse linear probe. The probe outperforms MSP under selective abstention, with gains scaling with baseline miscalibration up to 21 AURC points. Because every feature has a closed-form geometric meaning, the probe's coefficients trace how and where along depth errors take shape -- which layers commit prematurely, which contradict the running state, where trajectories drift away from their endpoint.

한국어 요약

📋 한 줄 요약

**[LM Uncertainty / Layer Trajectory]** Per-layer MLP update의 cumulative 경로에서 11개 scale-invariant geometric feature를 추출해 sparse linear probe — MSP 대비 selective abstention에서 최대 21 AURC point 개선·각 layer의 commit·contradiction·drift 해석 가능.

🎯 핵심 기여도

💡 핵심 아이디어

LM uncertainty는 final activation snapshot이 아니라 layer-wise update trajectory에 인코딩되며, scale-invariant 기하 특성으로 trajectory를 압축하면 MSP를 능가하면서 동시에 어느 layer에서 오류가 형성되는지 해석 가능한 신호를 얻는다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Uncertainty quantification에 layer trajectory라는 새 정보 차원 도입, 해석 가능성과 성능을 동시 달성, miscalibrated 모델에서 가장 큰 효과로 실용 가치. **한계**: 11 geometric feature 선택의 도메인 의존성, MLP update 외 attention 경로 미커버, 매우 큰 모델에서의 효율성 추가 검증.

🚀 실용적 활용