Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals
For researchers using activation oracles to interpret LLMs, this work provides a practical calibration method to improve reliability of interpretations.
Activation oracles for interpreting language model internals lack uncertainty quantification. The authors show that bootstrap mode frequency yields the best calibration (ECE 5.7% vs 25.5% for log-prob on Qwen3-8B), while log-prob provides a fast, cheaper alternative.
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.