Techniques

Rubric-Supervised Critic (2603.03800) — Critic model trained on 24 behavioral features achieves +15.9 improvement on SWE-bench via reranking. 83% fewer task attempts with early stopping.

Optimal Transport Refusal Ablation (2603.04355) — Achieves up to 11% higher attack success than SOTA across six models. Layer-selective intervention at 1-2 layers outperforms full-network approaches. Refusal mechanisms are more localized than assumed.

SynthID Watermark Vulnerabilities (2603.03410) — First formal analysis of Google's production watermarking. Mean score function is inherently vulnerable to "layer inflation attack." Watermarking-based IP protection has fundamental limits.