Detecting deception in language models
Probe-based analysis of deceptive behavior and internal model signals.
Current focus
Using probes and evals to study deceptive behavior in language models.
Recent writing
Two Kinds of Deception, Two Kinds of Signal
Replicating Apollo Research's linear-probe pipeline on Qwen 3.5-4B. The original paper already flagged insider trading as layer-sensitive; on Qwen the cross-domain transfer collapses belo...
First Notes on Extending Deception Detection to Qwen 3.5
Initial results from extending a deception-detection pipeline to Qwen 3.5-4B, including rollout quality, grading noise, and early probe performance.