Alibaba Research's Recursa adapts agents to new domains without changing model weights
Alibaba Research introduces Recursa, which builds a validated skill repository from task experience so a model can adapt to a new domain with fixed weights.
Alibaba Research has built a system called Recursa that lets a language model get better at an unfamiliar domain without changing its weights. Instead of changing the model, Recursa keeps a knowledge repository. It fills that repository with domain skills, instructions on when to apply each skill, memories of past tasks and reflections on what to improve. On Harvey LAB's Firm Knowledge legal benchmark, the paper reports that the all-pass rate on held-out test tasks rises from 30.0% to 55.0% with DeepSeek-V4.1-Flash. With Qwen3.8-Flash, it rises from 35.0% to 46.0%.
The repository is built in two phases. First, Recursa bootstraps it from domain materials. Then a recursive self-improvement loop refines it. In that loop, a learner splits its training tasks into subsets and hands them to child learners. The leaf learners run solvers in parallel, diagnose failures and propose edits. Only edits that pass validation are merged back into the repository. The abstract says the system does this without human experts or a stronger teacher model. The paper's curator summary adds that the same model that does the tasks also writes the revisions, which separates the approach from distillation.
The gains are measured on a single benchmark with held-out tasks, and the model stays the same while the repository changes. The 30.0% to 55.0% jump with DeepSeek-V4.1-Flash is the larger gain. The Qwen3.8-Flash result, which moves from 35.0% to 46.0%, is smaller and starts from a different baseline. The paper's own caveats matter here. The authors say results depend on the quality of the domain corpus, and the curator summary notes that experiments cover one legal domain. The paper also says that extending self-improvement to the model weights and harness is ongoing work, so the weights-fixed design is a scope choice for now.
The approach is aimed at teams that cannot easily get expert-labeled training data for a new domain. The arXiv abstract frames the problem that way: new domains rarely provide large amounts of expert-created training data or reliable supervisory signals. For such teams, the result is a repository of skills, instructions, memories and reflections, with only validated edits merged in, rather than a fine-tuned checkpoint.
What changed
Recursa builds and validates a domain skill repository from task experience so the model can adapt to a new domain while its weights stay fixed.
Type: Research · Applies to: DeepSeek-V4.1-Flash, Qwen3.8-Flash
Why it matters
Teams adapting agents to specialised domains may be able to try a validated skill repository before investing in fine-tuning.
What you can do
Try building a validated domain skill repository before committing to fine-tuning for a new agent domain.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works