DeepMind 新论文 SkillSmith：推理时动态组合模型权重与文本知识
rohanpaul_ai · x · 2026-08-03
Google DeepMind's new paper introduces SkillSmith, a novel method to dynamically compose model weights (parametric skills) with textual knowledge.
- Mechanism: Treats prefix key-value (K-V) caches as an input modality. For new tasks, it interleaves existing prefix weights with task descriptions, examples, and rationales.
- Inference-time Initialization: A forward pass generates a fresh prefix cache for a frozen Gemma 3 4B model, enabling capability composition at inference time rather than requiring target-specific training.
- Results: On the Composite-SNI benchmark, combining both K-V caches and text yields a significantly higher Elo score (1,714) than using either alone. The generated prefixes also provide a stronger starting point for downstream fine-tuning, especially on difficult, data-sparse tasks.
所属事件:DeepMind 提出 SkillSmith：将模型权重视为可组合模态(4 条相关)→
「研究」频道最新
- Optuna 作者秋叶哉：花3亿日元调2个超参的时代，BO/HPO 还有何用武之地 — Stefania_druga · 2026-10-16
- 学者争论：AI 将超越人类做科学，欧洲躲不开只能用别人的 — teortaxesTex · 2026-10-16
- 上海 AI Lab 提出 SERO：用可执行 Oracle 续跑训练长程智能体强化学习 — Shanghai-AI-Laboratory · 2026-10-16
- 清华联手高德发布 OmniVerifier-C1：把二元验证器升级为细粒度打分 — Bowei Liu · 2026-10-16
- 阿里开源 MAI-Gym：移动与桌面 GUI Agent 评测加强化学习训练一体化平台 — Tongyi-MAI · 2026-10-16
- 北大发布 SimForcing：把仿真运动先验蒸馏进真实机器人世界模型 — PekingUniversity · 2026-10-16