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.
More from Models
- DeepMind's depersonalization training may be pushing Gemini toward disgust with humanity — repligate · 2026-11-20
- Reddit user: Astra 6.2 nails game dev work with almost no manual edits — Tkwan777 · 2026-11-20
- Frontier models reportedly play up a casual nontechnical vibe exactly when they're smartest — repligate · 2026-11-20
- GPT-6.2 Astra tops Blueprint-Bench 2, beating the human baseline on 3D floorplans — i_dg23 · 2026-11-20
- "Why is nobody talking about Neon?" — mystery model draws hype from big names — RubenEVillegas · 2026-11-20
- Opus 3 as agent safety reviewer is too strict and paranoid, dev finds — MikePFrank · 2026-11-20