Weight-Space Reading: LLMs predict weight-edit effects at 73% accuracy with zero forward passes
udmrzn · x · 2026-11-16
A new alphaXiv paper, "Weight-Space Reading" from Poindexter Labs, asks whether an LLM can predict how editing a network's weights will change its behavior — without ever running the model.
Setup and findings:
- Sandboxed reader agents inspect raw weights of Qwen3.8-27B, Qwen3.5-9B and Muse-Glimmer-30B with Python tools, predicting effects of subtle edits on math accuracy, reasoning length, and refusal behavior with zero forward passes
- 27B predicts subtle non-destructive edits to its own weights at 0.730 balanced accuracy vs 0.578 from text descriptions alone (+0.152, 95% CI [+0.056, +0.256]), recovering two-thirds of real behavior changes zero-shot
- No conclusive self-reading advantage: a stronger external reader (27B) beats Muse-30B on its own weights (0.615 vs 0.468)
- Big caveat: across 40 edits that did nothing to Muse-30B (incl. 6 no-op controls), readers predicted changes for 53–88% of them; Muse-30B itself for 85%
- Benchmarks built from destructive edits really measure damage detection: an out-of-family model matches 27B on its own weights (0.594 vs 0.593)
The authors distill four rules for evaluating weight reading and release the evaluation harness.
More from Research
- AI/ML for Fluids cohort 2 opens: 60+ live hours over 3 months for $500 — ricardovinuesa · 2026-11-16
- "The conference-paper game is 100% dead": researcher urges GitHub-first research output — DimitrisPapail · 2026-11-16
- DeepMind researcher on doing research in the AI era: minimize friction to excellent work — DimitrisPapail · 2026-11-16
- Google AI Co-Scientist progress presented by Michael Brenner at MLCB 2026 — anshulkundaje · 2026-11-16
- New Paper: Adding Input Noise During Pretraining Distillation Beats Standard Distillation — DimitrisPapail · 2026-11-16
- Auto Weight-Space Programming: Qwen3.5-4B learns to apply logit lens to its own neurons at 99.8% — udmrzn · 2026-11-16