Talk on steering vectors: sparse propagation, circuit similarity, attention decomposition
sarahwiegreffe · x · 2026-10-25
Sarah Wiegreffe previews a talk on steering-vector interpretability covering: sparse propagation effects, functional/circuit similarities across SVs fit with different methods, decomposing SV attention effects into input-dependent and independent parts, and doing circuit discovery on multi-token generations.
More from Research
- Study: LLMs rewriting their own ARC-AGI-3 harnesss gain just 2.8 points, not significant — omarsar0 · 2026-10-25
- EMNLP 2026 paper: steering vectors run through 8-11% of the compute graph and sparsify to 85-96% — sarahwiegreffe · 2026-10-25
- EMNLP paper localizes faithfulness and citation errors in agentic deep research systems — EliasEskin · 2026-10-25
- Jev for Recursive Language Models: half the cost, 6x faster at 1M tokens — dbreunig · 2026-10-25
- LeHami: LeCun and colleagues borrow 1830s Hamiltonian mechanics to improve JEPA world models — TheTuringPost · 2026-10-25
- EleutherAI brings 6/6 accepted papers and 2 events to EMNLP 2026 — BlancheMinerva · 2026-10-25