Steering single neurons in a molecule-generating model changes one property while staying valid
alex_peys · x · 2026-10-30
Inspired by the viral Qwen single-neuron steering experiments, alexpeys tested whether the ideas transfer to autoregressive models that speak a non-human language—sending Codex to hunt for neurons in a decoder-only model trained to generate molecules.
It worked: single neurons, when steered, keep the model generating valid molecules but change just one property—for example, one neuron produces molecules with more rings. Chart data: mean ring count rose from 2.87 to 3.57 (unconditioned) and 2.79 to 3.58 (conditioned), 1,024 steered outputs per arm per condition, with logP/TPSA conditioning bounds still satisfied.
Takeaway: mech interp techniques work beyond natural language, on molecular SMILES too—"maybe mech interp is fun."
More from Research
- Chip-integrated Pockels laser emits 780nm with 1GHz tuning, locked to rubidium D2 line — jwt0625 · 2026-10-30
- CVPR's new prior-work rule for arXiv preprints sparks fairness debate among researchers — CSProfKGD · 2026-10-30
- Eterna to reveal OpenKnot round 9 RNA design results on Nov 5, plus Struct2SeQ-OA method talk — chaitjo · 2026-10-30
- UCLA's ChairNav brings long-horizon autonomous wheelchair navigation to real sidewalks — zhoubolei · 2026-10-30
- Architecture breakdown: wider and deeper than ML3, top-4/128-expert MoE, MLA keeps KV cache tiny — eliebakouch · 2026-10-30
- Google's sleep framework lets LLMs keep learning, retaining up to 10M tokens — ZeroStateReflex · 2026-10-30