Mechanistic Interpretability

I wrote seven English principles, used a 72B model as critic-rewriter, and trained two versions of CAI on a sycophantic 8B model. SL-CAI imitated revisions: aggregate sycophancy 0.348 (worst recovery I’d tried). DPO-CAI contrasted them: 0.166 (best in the study). Same data, opposite outcomes. Imitation isn’t contrast.

GRPO achieves the lowest aggregate sycophancy (0.169) and an 8% flip rate — but linear probing finds only 7 of 36 layers above chance (AUROC 0.541). The best behavioral fix uses the most diffuse internal mechanism.

IPO restructured the model more deeply than SimPO (probe transfer 0.365 vs 0.429) but performed worse behaviorally (0.281 vs 0.176). The reference model doesn’t limit intervention depth — the loss shape does. A 2x2 framework replaces the 1D hypothesis.

DPO suppresses sycophancy but preserves the internal representation (probe transfer 0.677, p=0.005). SimPO reorganizes it — probe drops to chance (0.503, p=0.154 after correction). Same data, same model, radically different internals.

I deliberately induced sycophancy in Qwen3-8B, recovered it with DPO, then used linear probing to show the internal sycophancy representation survived alignment. The fix is cosmetic.