改 1000 个参数，AI 成绩涨 29 分
改大模型，通常要动几亿个参数。这篇论文发现，只改其中 1023 个，数学成绩就能涨 29.45 个百分点——相当于在 100 道题里多答对近 30 道。做法是让 AI 直接「读」和「改」自己的权重，而不是靠梯度反传。但别急着兴奋：它改完数学，顺带把物理、指令遵循也提了几分，编程却纹丝不动；而且 16 次尝试里有 6 次因为改完没法验证效果，干脆原样退回。也就是说，这技术目前像蒙眼拆弹：偶尔一剪子剪对，但没人保证下一剪。
📄 原文摘要(英文)
Weight-Space Programming (WSP) enables reasoning agents to directly inspect and modify neural network parameters, offering an alternative to gradient-based model adaptation. However, its capabilities, cross-task generalization, and practical limitations remain insufficiently understood. We systematically investigate WSP through 16 editing trajectories across seven benchmarks and two model scales (Qwen3.5-0.8B and Qwen3.5-9B). Our findings reveal three key insights. First, WSP can achieve substantial performance gains through highly localized interventions. On full held-out evaluations, accuracy improves by up to 29.45, 4.42, 36.72, and 17.11 percentage points on MATH-500, GPQA Diamond, IFEval, and SciBench, respectively. Notably, modifying only 1,023 parameters yields a 29.45-point improvement. Second, WSP exhibits promising but uneven cross-task generalization. A Math-edited checkpoint improves GPQA Diamond, IFEval, and SciBench by 3.03, 12.18, and 20.10 points, respectively, without further editing, while leaving SciCode performance unchanged. Third, effective editing remains challenging and inconsistent. Outcomes vary substantially across trajectories, and six of the 16 trajectories retain the original weights when candidate improvements cannot be reliably validated. Together, these results highlight WSP's potential for highly localized model adaptation while revealing important challenges in intervention discovery, reliability, and cross-task transfer. Homepage: https://wsp-scope.github.io/ GitHub: https://github.com/WSP-Scope/WSP-Scope-Project