Recursa lets a foundation model system rewrite itself: all-pass on held-out legal tasks rises from 30.0% to 55.0%

Recursa lets a foundation model system rewrite its own instructions, skills and memory from its task experience, improving performance in a new domain without a human expert or a stronger teacher model. Earlier methods learn only parts of the system, such as instructions, harness or tools, and often rely on a stronger model to drive learning; methods aimed at the knowledge gap study only a domain corpus, and methods aimed at the behavior gap learn actions only from task feedback. On Harvey LAB's Firm Knowledge benchmark, all-pass rate on held-out test tasks rises from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

The system first reads domain materials and records its findings as skills that keep links to their sources and state their uncertainty and applicable scope, then practices on tasks generated from those materials to form an initial knowledge repository. Without this step every task begins with exploration, and that cost compounds in a multi-agent system. Across all 250 tasks, bootstrapping reduces interaction turns and input tokens by roughly a fifth to a quarter and raises the all-pass rate slightly, while the mean criteria pass rate stays unchanged.

The same model then serves as both solver and learner: a top-level learner splits the 150 training tasks by topic into subsets, parallel child learners diagnose failures, edit skills and validate, and the parent merges the revisions and reflects on the children's learning traces. A single learner that does not split tasks and runs the self-improvement loop directly on all 150 training tasks serves as the control isolating the effect of multiple learners. With DeepSeek-V4.1-Flash, multi-learner learning raises training all-pass from 34.7% after bootstrapping to 48.7% and test from 33.0% to 55.0%, while the single learner reaches 43.0% on test; with Qwen3.8-Flash, multi-learner learning goes from 37.3% to 47.3% on training and from 34.0% to 46.0% on test, while the single learner falls to 30.0% on training and reaches 37.0% on test.

The demonstrated setting is Harvey LAB's Firm Knowledge synthetic law firm: 46 fictional clients, 266 matters across 15 practice areas, 9,288 files, and 250 tasks split into 150 training and 100 test tasks, judged by Qwen3.8-Max. Only the instructions, skills and memory in the knowledge repository are learned; the harness and foundation models stay fixed, and extending recursive self-improvement to model weights and the harness is ongoing work.

The authors note that performance relies on the quality and availability of domain corpora, requires substantial compute, and that incorrect or biased data and flaws in the underlying models can propagate through learning; learning effectiveness depends on the capability of the underlying models, which they have not investigated, and experiments cover only a limited set of domains. They also note that Recursa autonomously explores software environments, executes programs and retains observed information in memory, raising privacy and security risks; experiments ran in strictly isolated sandboxes with no network access except through controlled channels for model inference and training.