Neuron Tuning: tweaking a few context-specific neurons disturbs unrelated text 3-19x less than LoRA
burny_tech · x · 2026-11-02
Jonathan Chang introduces Neuron Tuning, a fine-tuning approach that updates only the few neurons that fire in the target context, keeping the edit's effects localized to that context.
- Compared to LoRA, it changes the model 3-19x less on unrelated text
- Benchmark charts cover three edits (casual talk style, one-sentence weather answers, learning 160 facts about fictional people): neurons picked by gradient or activation selectivity cause 5-19x less side effects than a same-size rank-8 LoRA (12M numbers), while globally-firing neurons cause 4-37x more
- Key takeaway: a neuron's contextual selectivity determines how local the fine-tune is — picking the right neurons matters more than the method itself
More from Research
- Deep RL book excerpt: the policy loss is fake — a surrogate with no physical meaning — burkov · 2026-11-02
- NBER Paper Proposes Coasian Stakeholder Theory: Defining Stakeholders by Shutdown Power — sebkrier · 2026-11-02
- FAST-MaP Protocol Brings Single-Nucleotide RNA Structure Data in Under a Week via Nanopore — anshulkundaje · 2026-11-02
- A Free arXiv Paper Collection to Help You Discover Worthwhile Reads — omarsar0 · 2026-11-02
- livgorton: mech interp is pre-paradigmatic; "stuck in old paradigm" is a dodge — livgorton · 2026-11-02
- VTuber stream explainer on IPD × NVIDIA protein design model RFD4-Proteina — AllThingsApx · 2026-11-02