Search papers, labs, and topics across Lattice.
This paper introduces STATEWITNESS, an activation explainer designed for auditing deceptive behavior in reasoning large language models (LLMs). By utilizing a separate decoder to analyze hidden states, STATEWITNESS provides detailed, inspectable evidence that enhances the understanding of why certain responses may be deceptive. The method achieves a mean AUROC of 0.916, significantly outperforming existing black-box monitors and activation-probe baselines, while also improving detection rates when integrated with current systems.
STATEWITNESS not only identifies deception in LLMs but also offers detailed insights into the reasoning behind suspicious responses, transforming how we audit AI behavior.
As LLMs acquire stronger reasoning capabilities, deceptive behavior becomes an increasingly serious safety concern. Existing deception monitors either score visible transcripts or derive scalar probe scores from representation vectors, leaving little inspectable evidence about why a response is suspicious. We introduce STATEWITNESS, an activation explainer for deception auditing. A separate decoder reads a target model's hidden states, then answers natural-language queries or emits structured reports about them. We evaluate STATEWITNESS on two target reasoning LLMs across seven deception datasets. STATEWITNESS reaches 0.916 mean AUROC, a relative gain of 11.6% over the best black-box text monitor and 25.0% over the best activation-probe baseline under the same evaluation protocol. When combined with existing monitors, STATEWITNESS reduces missed deceptive examples in simple threshold ensembles. Beyond scalar detection, the decoder returns query-level answers, schema reports, and token- or sentence-level evidence traces for human inspection. We view this interface as a potential building block for broader interpretability and alignment tools.
