Inkling 2 Small matches frontier weight-space programming results with 16x fewer weight edits
mools04 · reddit · 2026-11-18
Building on "Sparks of Weight-Space Programming," alphaXiv's Inkling 2 Small shows model repairs need far fewer weight-space edits:
- Broken Car task: GPT-6 Astra changed 6,073 weights to fix a broken car policy — most merely zeroed old weights, 55 did real work. Inkling wrote a 10-weight fix from scratch, teaching a maneuver PPO never found (reverse until stalling against a wall, then launch), using a velocity-detecting neuron already in the network.
- Qwen3.5-0.8B math looping: Astra removed a "looping" direction across 64 neurons in two MLP layers; Inkling kept only the 4,096 largest-magnitude changes per layer (8k total), matching accuracy (58.2% vs 58.6%), with layer 15 alone recovering most of the gain.
600x and 16x fewer edits with no performance loss. Reproduce via a marimo notebook (molab.marimo.io/github/alphaXiv/weight-edit-minimization), built with their OpenResearch framework.
Related event: Weight-Space Programming: Tiny Weight Edits Fix Model Flaws(2 posts)→
More from Research
- Domingos on iWSP paper: just use the program instead of writing it into LLM weights — pmddomingos · 2026-11-18
- Stanford CS329Z releases two standout lectures on agent design patterns and multi-agent systems — stanfordnlp · 2026-11-18
- SC26 BoF: top HPC names to debate coding agents' threat to numerical linear algebra trust — kfountou · 2026-11-18
- Yoav Goldberg: Reject papers for low human input, not for LLM use — yoavgo · 2026-11-18
- One RTX 4090 runs Qwen3.8-Flash-Next: a hand-built inference engine, fully gated — Distinct-Pie2389 · 2026-11-18
- Indie dev's model scans raw binaries directly, scoring 0.977 on code blocks — wavefnx · 2026-11-18