Learning Self-Interpretation from Interpretability Artifacts: Training Lightweight Adapters on Vector-Label Pairs

Researchers have developed a method to make large language models reliably interpret their own internal reasoning without being retrained. By training a tiny 'adapter' with just a few parameters on interpretability data, they force a frozen model to explain its hidden states. The adapter achieved 94% accuracy at identifying topics and surfaced implicit reasoning steps in 70B-scale models, outperforming the original training labels (71% vs 63%). This shows self-interpretation improves with model scale.