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.
co-developed
A toolkit four of us built to ask: if you ask a model what it prefers in four different ways, do you get the same answer? Mostly, no — stated preference and actual choice diverge sharply at 14B parameters. Along the way we also caught a silent parsing bug that had been quietly misreading responses. Full paper linked below.
I tested whether quantum kernel SVMs (via PennyLane) actually beat classical PCA+SVC baselines for probing transformer representations. They did not (98.0% vs. 98.7%, p=1.0) — another clean null result, and a useful one to have on record.
An 11-task Python/ML taskphase, capped off with a scikit-learn multiclass pipeline at 99.39% test accuracy.