2026-11-09 · America/Los_Angeles · 论文 · #10

Exploring Weight-Space Programming: Capabilities, Generalization, and Limitations

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,…

热度 36.7 / 100；排名与评分保留该期记录。