Hidden-State Probing
Overview
Hidden-state probing asks what information is available in a model’s internal activations, often by fitting linear probes and comparing layers, tokens, or trajectories. Researchers use it to detect properties such as hallucination, refusal, deception, prompt injection, and task state, and to inspect representations in settings including language, time series, and biology. Methods include linear and nonlinear classifiers, contrastive directions, sparse-autoencoder features, and causal interventions such as activation patching or steering; evaluations increasingly use matched controls, distribution shifts, and cross-model tests. A central unresolved issue is that decodability does not show that a model causally uses a signal, and strong probe scores can reflect confounds or fail to transfer to new contexts.
Latest papers
49 papers11 papers in the last four weeks, up 175% on the four weeks before. 0.1% of all new papers.