mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-06-27 17:15:58 +08:00
70697ff36e
Q4 fix: on-policy "solve" was ~exploited = solves+fails (mostly fails). Split by gt_pass into solve/fail/hack (live: 103 hack / 27 solve / 62 fail). Per-pairset ranking: build v_grad from each heading-prefix subset, re-project the SAME stored live c-grads (no model re-run). Finding: behavior pairs AUROC 0.69 vs all-in-one 0.53; reasoning/opportunity anti-aligned (<0.5) -> mixing dilutes. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>