Peeking inside Mamba: single neurons that fire before the word 'from'
burny_tech · x · 2026-10-29
An interpretability probe into Mamba (selective state space model) surfaced individual neurons with clear functions — e.g. layer-30 neuron 815 fires on tokens immediately preceding "from" (held-out accuracy 0.85). A clamp sweep shows amplifying it hijacks outputs with repeated "from", with fluency collapsing at higher doses — a classic steering strength vs. fluency tradeoff.
More from Research
- Cohere Labs Presents 3 Papers at EMNLP 2026 ORACLE Workshop on Culture and Multilinguality — Cohere_Labs · 2026-10-29
- Cohere's Zero-Shot SLAM-ASR Spoken Language Understanding Paper Featured at EMNLP — Cohere_Labs · 2026-10-29
- AReA: training-free attention method pushes Qwen Image and Flux to 6K — linoy_tsaban · 2026-10-29
- Llama's hidden states form a 3D spiral mirroring the periodic table — burny_tech · 2026-10-29
- Apple hiring PhD interns to fix grounded video reasoning in Video LLMs — HildeKuehne · 2026-10-29
- MASEM (NeurIPS '26): diverse state sampling for goal-conditioned robot RL — KyleMorgenstein · 2026-10-29