Why model self-editing matters: zero capability gap between interpreter and target
menhguin · x · 2026-11-03
Responding to the weight-editing thread, menhguin argues self-editing is interesting because: (1) target and interpreting model undergo the same kind of forward pass on the same parameters, shrinking the confound space—he believes interpretation requires the full fwd/bwd pass, hence his look into optimizers/backprop; (2) using the target model itself means zero size/capability gap and no compression needed. He also notes narrow, specific behaviors are more likely lottery tickets to switch on or off.
Related event: Researchers explore LLMs editing their own weights for interpretability(4 posts)→
More from Research
- Follow-up: Astra's weight edits only worked on Qwen's visible flaw, not on Gemma 270m — teortaxesTex · 2026-11-03
- EMBER: pretrained gene-centric foundation model predicts multigene perturbation effects — jmuiuc · 2026-11-03
- LLM edits another LLM: tweaking 0.016% of weights lifts Qwen3.5-0.8B math from 7.9% to 48.1% — teortaxesTex · 2026-11-03
- New preprint RLCE: reinforcement learning with comparative evidence for social intelligence — pliang279 · 2026-11-03
- Tau Ceti uses AI plus Lean kernel to formalize 'all of basic mathematics', open to contributors — AlexKontorovich · 2026-11-03
- New research links AI companions to less loneliness and higher self-esteem — serialchilla91 · 2026-11-03