Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals

Federico Torrielli, Peter Schneider-Kamp, Lukas Galke Poech

arXiv:2605.26045 · 2026-05-27 공개 · arXiv · PDF

llm uncertainty-quantification model-calibration qwen3-8b confidence-estimation activation-oracles qwen3-6-27b bootstrap-mode

Abstract

Activation oracles aim to make the activations of other models legible to humans and yield promising results compared to white-box interpretability techniques. However, uncertainty quantification (UQ) for the natural-language outputs of such activation oracles is so far understudied. Here, we investigate 6 different methods for estimating the confidence of activation oracles and evaluate how well-calibrated their confidence scores are. Our experiments on 6,000 samples per oracle (varying verbalizer and context prompts) reveal that bootstrap mode frequency is the best-calibrated method among those tested (ECE 5.7% vs. 25.5% for the answer-word log-probability on Qwen3-8B; 10.3% vs. 13.1% on Qwen3.6-27B), and that the log-prob baseline can serve as a fast triage signal at a fraction of the cost. Code and the patched trainer are available at https://github.com/federicotorrielli/probabilistic_activation_oracles.

한국어 요약

📋 한 줄 요약

**[활성화 오라클 / 불확실성 정량화]** Bootstrap mode frequency가 6개 UQ 방법 중 가장 well-calibrated — Qwen3-8B ECE 5.7% (answer-word log-prob 25.5%), Qwen3.6-27B 10.3% vs 13.1%; log-prob baseline은 fast triage signal로 활용.

🎯 핵심 기여도

💡 핵심 아이디어

Activation oracle의 자연어 출력 신뢰도는 bootstrap mode frequency로 가장 well-calibrated하게 정량화 가능하며, 비용 효율적 triage가 필요한 경우 answer-word log-probability를 fast signal로 활용하는 2-단계 전략이 실용적이다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Interpretability 도구의 신뢰성 정량화라는 understudied 영역 개척, 6 method 체계 비교로 실용 지침 제공, fast triage + accurate 확인의 2단계 전략 제안, 6,000 sample/oracle 대규모 실험. **한계**: Qwen3 시리즈 중심으로 다른 모델 family 일반화 추가 검증, 6 UQ 방법 외 더 많은 방법 비교 여지, bootstrap의 계산 비용 부담.

🚀 실용적 활용