Small classifiers trained on internal activations to test whether a property (valence, preference, truth) is linearly readable, and where in the network.
Used in
- Chalmers 2025 - Propositional interpretability
- Gilg et al 2026 - Probing persona-dependent preferences
- Bianco Shiller 2026 - Mechanistic tracing of pain-pleasure decisions
- Small LLMs 2026 - No reliable evidence of self-reported sentience
- Rose et al 2026 - Detecting collusion through interpretability