Weight-Space Programming paper: editing 1,023 params lifts MATH-500 by 29.45 points
pmttyji · reddit · 2026-11-10
A new paper systematically studies Weight-Space Programming (WSP), where reasoning agents directly inspect and modify neural network weights as an alternative to gradient-based adaptation. The evaluation spans 16 editing trajectories across 7 benchmarks and two model scales (Qwen3.5-0.8B and 9B).
Key findings
- Large gains from highly localized edits: up to +29.45, +4.42, +36.72, and +17.11 percentage points on MATH-500, GPQA Diamond, IFEval, and SciBench on held-out evals; modifying just 1,023 parameters yields the +29.45 MATH-500 jump (0.8B Math: 8.35 → 37.80).
- Uneven but real cross-task generalization: a Math-edited checkpoint improves GPQA Diamond, IFEval, and SciBench by +3.03, +12.18, and +20.10 points with no further editing, while SciCode stays flat; gains shrink drastically at 9B (Math even -0.35).
- Editing remains unreliable: outcomes vary substantially across trajectories, and 6 of 16 trajectories keep the original weights when candidate improvements can't be validated.
Paper, code, and project page are public; WSP shows promise for localized adaptation but intervention discovery, reliability, and transfer remain open challenges.
More from Research
- AI Trials Observatory launches: 93 RCTs on how AI affects mental health and beliefs — DG_Rand · 2026-11-11
- Raschka: open LLMs have abandoned full attention — hybrids and linear models now split 50/50 — rasbt · 2026-11-11
- Strogatz & Townsend's new book Big Math tackles math's Faustian bargain with AI — stevenstrogatz · 2026-11-11
- RLVR lets small models close the coding gap; Python's dynamic typing hurts pretraining — mdancho84 · 2026-11-11
- 50 researchers from ByteDance, Alibaba, Tencent release 303-page code AI field guide — mdancho84 · 2026-11-11
- pi-H3 generates 768p video in 2-4 steps: 10s clip in 9.2s on a single B200 — GordonWetzstein · 2026-11-11