Recursa 论文：不改模型权重，靠自建知识库提升 DeepSeek 与 Qwen 在法律任务上的表现
这篇论文提出 Recursa，不微调、不换模型，靠系统自己写知识库，把 DeepSeek-V4.1-Flash 法律任务成绩从 30% 拉到 55%，做领域适配的可以照着试。
论文提出 Recursa 框架，通过系统自动撰写知识仓库（存领域技能、使用说明、任务记忆和反思）实现 RSI。在 DeepSeek-V4.1-Flash 上，held-out 法律任务的 all-pass 成绩从 30.0% 提升到 55.0%，模型权重保持不变。在 Harvey LAB 的 Firm Knowledge 基准上，Qwen3.8-Flash 从 35.0% 升至 46.0%。流程是先从领域材料构建仓库，再由并行学习者执行任务、提出修复，经验证的修复才合并回去。作者建议在微调之前先尝试经过验证的技能仓库。
Banger paper on RSI with foundation model systems. (bookmark it) Seriously, I highly recommend this paper if you want to start experimenting with RSI in your own company. Lots of great insights and tips. They propose Recursa, which raises DeepSeek-V4.1-Flash from 30.0% to 55.0% all-pass on held-out legal tasks, without a stronger teacher or human experts. The model weights stay the same. The gain comes from a knowledge repository the system writes for itself. The repository stores domain skills, instructions on when to use them, memories of past tasks, and reflections. Recursa first builds it from domain materials. Then parallel learners run tasks and propose fixes, and only validated fixes get merged back. On Harvey LAB's Firm Knowledge benchmark, Qwen3.8-Flash goes from 35.0% to 46.0% with the same loop. If you are adapting an agent to a new domain, try a validated skill repository before reaching for fine-tuning. arxiv.org/abs/2611.04035 Chat with Paper: academy.dair.ai/papers/recursa… 💬 17 🔄 1 ❤️ 28 👀 2645 📊 21 ⚡
