Figma-style attribution graphs for SAE features. Nodes = features. Edges = feature-to-feature attribution, scored by AtP* (Kramár et al. 2024). Pick a scenario below — each is a real circuit computed on our Qwen3.6-27B paper-grade SAEs.
Attribution graphs on Qwen3.6-27B paper-grade SAEs. Upstream = L11, downstream = L31. Triangle nodes are SAE reconstruction-error terms (Marks et al. 2024). Edge thickness = |attribution|; orange = positive, cyan = negative. Each scenario uses a task-specific contrastive logit metric.
Run 15b_sfc_qwen36_27b_papergrade.ipynb for the four-scenario SFC pipeline, or 14_attribution_patching.ipynb for node-only AtP*. Both emit circuit.json in the schema this viewer consumes.
Four ways to compute attribution graphs, fastest → slowest. All output the same circuit.json format this viewer renders.
AtP* with QK fix + GradDrop (Kramár 2024). Node attribution only. Fastest — 2-3 forwards + 1 backward per prompt.
Marks et al. 2024 replication. Node + edge attribution via AtP + IG-10 early-layer fallback. Emits full DAG.
Specialized SFC pipeline for caiovicentino1/qwen36-27b-sae-papergrade. Produces the 4 scenarios above — medical · IOI · math · refusal. One JSON per scenario.
AutoCircuit library. Original NeurIPS 2023 algorithm. Independent verification. Slower than AtP but peer-reviewed.
Lindsey et al. 2024. Train a shared-dictionary SAE across L11/L31/L55. Ties multi-layer features into one feature index.