Auto Weight-Space Programming: Qwen3.5-4B learns to apply logit lens to its own neurons at 99.8%
udmrzn · x · 2026-11-16
A new alphaXiv blog by Atharva Naik, Emmy Liu, and Varun Gangal formalizes weight-space programming — reading or writing a model's weights directly as numbers — and sets the goal of auto-WSP: a model doing it to itself.
First toy result: Qwen3.5-4B learns to apply the logit lens to 128-number slices of its own unseen neurons, scoring 99.8% on a four-way test (chance 25%). Naming tokens outright, writing weights, circuit tracing, and behavior prediction are next steps.
Related event: Researchers Propose Weight-Space Programming for Model Self-Inspection(4 posts)→
More from Research
- AI/ML for Fluids cohort 2 opens: 60+ live hours over 3 months for $500 — ricardovinuesa · 2026-11-16
- 1,000 agents write 5,000 rule-based PyBot versions in pure Python, beating a 360M LLM — LoubnaBenAllal1 · 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