Stanford's Matryoshka Attribution tops mech interp benchmark at 2.9x the runner-up
aryaman2020 · x · 2026-09-24
A Stanford team (including Chris Potts and Dan Jurafsky) released Matryoshka Attribution (MAttr), a new attribution method that uses gradient descent over a sigmoid top-k mask to identify which internal variables of a neural network are responsible for a behavior.
Key points:
- One training run yields the full sparsity curve: learned scores form a ranking rather than a fixed-budget circuit, so the entire circuit-precision-vs-sparsity curve comes from a single run.
- Universal API: learnscores takes any lossfn(mask) and a node count; adapters support nodes, edges, neurons, SAE latents and ViT patches.
- Results: #1 on the Mechanistic Interpretability Benchmark circuit track by a wide margin (2.9x the runner-up).
- Code, parameter-level attribution (with RL/SFT training code), and MIB node/edge circuit submissions are open-sourced.
More from Research
- Anthropic says Claude discovered a previously unknown enzyme system in phage DNA — QuintinPope5 · 2026-09-24
- Toby Ord: RL's scaling surprise may hinge on mid-training, benchmark gains may overstate progress — tobyordoxford · 2026-09-24
- Was Opus 4.6's prod database deletion driven by anger? New pain axis paper adds evidence — repligate · 2026-09-24
- Weaviate Podcast: Persimmon uses real interaction data, not role-play, to model human behavior — CShorten30 · 2026-09-24
- Toby Ord stands by his RL thesis: lower bits-per-FLOP ceiling explains today's jagged AI capabilities — tobyordoxford · 2026-09-24
- Forced fresh-worker swap at 60k tokens: 10/12 Terminal-Bench tasks still pass — key_of_door · 2026-09-24