Stanford's Matryoshka Attribution tops mech interp benchmark, removes refusal by resetting 1% of weights
stanfordnlp · x · 2026-09-26
Stanford researchers (Aryaman Arora, Noah Goodman, Dan Jurafsky, Christopher Potts, et al.) release Matryoshka Attribution (MAttr), a new interpretability method.
- Reframes attribution as finding nested subsets of internal components that minimize a downstream loss, using a differentiable sigmoid top-k mask operator trained across all sparsities simultaneously
- Ranks #1 on the official Mechanistic Interpretability Benchmark leaderboard, identifying sparse, task-transferable circuits
- Practical demo: MAttr trained with RL on refusal judge scores shows restoring just 1% of Llama 3.1 8B Instruct's weights to base state suffices to remove refusal behavior
Related event: Stanford's Matryoshka Attribution Tops Interpretability Benchmarks(2 posts)→
More from Research
- Full paper: active learning cuts surgical video annotation effort by ~50% — santiviquez · 2026-11-11
- Active learning with weak supervision cuts surgical video annotation effort by ~50% — Laparoscopes · 2026-11-11
- Active Learning with Weak Supervision Cuts Surgical Video Annotation Effort by 50%, Penn Study Finds — Laparoscopes · 2026-11-11
- Tuva Labs takes 1st on ARC-AGI-2, 2nd on ARC-AGI-3 in ARC Prize 2026 — GregKamradt · 2026-11-11
- Sakana AI × Pathway Post-Transformers lecture on AI and the brain now online — kaixhin · 2026-11-11
- Defending METR: public data can't second-guess its AI speedup findings — peterwildeford · 2026-11-11