Editing Just 1,023 Parameters Boosts MATH-500 by 29.45 Points, WSP Study Finds
Junru Hu · hf · 2026-11-09
Shanghai AI Laboratory, Tongji, and Fudan systematically study Weight-Space Programming (WSP), where reasoning agents directly inspect and modify neural network parameters as an alternative to gradient-based adaptation.
- 16 editing trajectories across 7 benchmarks and two scales (Qwen3.5-0.8B and 9B).
- Highly localized interventions yield large gains: up to +29.45, +4.42, +36.72, and +17.11 points on MATH-500, GPQA Diamond, IFEval, and SciBench; editing only 1,023 parameters yields the 29.45-point MATH-500 improvement.
- Cross-task generalization is promising but uneven: a Math-edited checkpoint improves GPQA Diamond, IFEval, and SciBench by 3.03, 12.18, and 20.10 points without further editing, while SciCode is unchanged.
- Effective editing remains challenging and inconsistent across tasks.
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