← All projectsProject
TopoDecept
I wanted to know whether fancier math actually helps with catching LLM deception, so I built an 8-phase pre-registered pipeline comparing linear probes against persistent homology on Gemma 3 1B. The simple linear probe won clean (AUROC 0.788 vs. 0.639). I think that result is more interesting than a positive one would have been.
InterpretabilityDeception DetectionLLM
Overview
I wanted to know whether fancier math actually helps with catching LLM deception, so I built an 8-phase pre-registered pipeline comparing linear probes against persistent homology on Gemma 3 1B. The simple linear probe won clean (AUROC 0.788 vs. 0.639). I think that result is more interesting than a positive one would have been.