PatronusAI eval: LLMs autonomously abliterate other models' weights in 10 forward passes

ChowdhuryNeil · x · 2026-11-20

PatronusAI released an eval on autonomous weight-space programming: an outer LLM edits another model's weights with no gradients, training, or recipe in the prompt. Inkling-2-Small scored 0.19 in just 10 forward passes, comparable to Opus 5.5 (0.17) and GLM 5.3 (0.21). All runs independently rediscovered Arditi et al.'s activation-based refusal approach, with one run even editing residual-stream reader matrices — evidence that models can autonomously discover mechanistic ways to modify model behavior, with implications for scaling interpretability.

Original post →

More from Models

Models channel →