Recursa: Toward Recursively Self-Improving Foundation Model Systems

Recursa: Toward Recursively Self-Improving Foundation Model Systems

Foundation models are often dropped into environments filled with unfamiliar documents, tools, terminology, and local rules. Further training can help, but it usually requires substantial data and dependable rewards or expert supervision—resources that a new domain may not provide. The authors of Recursa propose a different starting point: improve the surrounding agent system through its own work.

The central idea is to treat adaptation as foundation model system learning. Instead of updating model weights alone, Recursa learns a knowledge repository containing instructions, reusable skills, and memories of prior successes and failures. The model, harness, and repository work together, with the repository serving as an immediately editable layer between a general-purpose model and a specialized domain.

Two gaps to close

The paper distinguishes between a knowledge gap and a behavior gap. The knowledge gap appears when an agent does not know a domain’s facts, vocabulary, document layout, or available resources. The behavior gap appears when the agent has access to relevant information but applies it incorrectly—for example, finding a related document but choosing the wrong evidence or using an overly narrow interpretation of a requirement.

Recursa addresses these gaps in two phases:

  • Bootstrapping: agents inspect a domain corpus, create summaries and navigation structures, and practice on tasks generated from the material.
  • Recursive self-improvement: agents solve training tasks, inspect execution traces and feedback, revise instructions and skills, and validate those changes before retaining them.

The same underlying model can act as solver and learner. Crucially, the system does not depend on a stronger teacher model to write the revisions, according to the authors.

A harness designed for long tasks

Recursa’s adaptation process depends heavily on its execution harness. Rather than forcing every tool result into the model’s context, the agent writes Python code in a persistent read-eval-print loop. Search results, intermediate variables, and reusable computations can remain in the execution environment while only selected evidence is returned to the model.

The harness also maintains an append-only event log. When earlier interactions fall outside the active context, an agent can query the log programmatically and retrieve the relevant records instead of loading an entire history. This design is intended to reduce context pressure during long research tasks and while learners diagnose previous attempts.

Multi-agent orchestration provides the other major component. A learner can divide a collection of tasks among child learners, run solver attempts in parallel, merge their proposed changes, and reflect on conflicts or unproductive work. At the leaf level, the process follows a repeated cycle: solve, diagnose, edit, and validate. An edit is kept only when it fixes targeted failures without causing regressions on sampled tasks that already passed.

What the system learned

For evaluation, the authors used Harvey LAB’s Firm Knowledge benchmark: a synthetic law-firm environment with 250 long-form tasks over a large corpus. The tasks require searching across documents, reconciling distributed evidence, and satisfying task-specific criteria. Recursa first built a navigation skill and 13 additional skills from corpus exploration and self-practice. These skills covered behaviors such as tracing document versions, determining task scope, and checking numbers and dates against conflicting sources.

Bootstrapping primarily improved efficiency. With DeepSeek-V4.1-Flash, the authors report that the bootstrapped system used fewer interaction turns and input tokens than the no-skills baseline, while the gain in accuracy was modest. The result suggests that domain maps and procedural notes can prevent repeated exploration even when a capable model can eventually locate the required evidence on its own.

The larger gains came from multi-learner self-improvement. On held-out test tasks, the all-pass rate increased from 30.0% to 55.0% with DeepSeek-V4.1-Flash, and from 35.0% to 46.0% with Qwen3.8-Flash. The authors attribute the difference to improvements in complete task execution rather than just partial rubric coverage. Their single-learner comparison performed worse than the multi-learner approach in these experiments, supporting the value of decomposing learning across specialized subsets.

The qualitative examples are revealing. After learning, an agent searched for multiple formulations of a legal concept rather than relying on a literal phrase. In another task, it used matter-level practice-group records to establish scope before deciding which settlements belonged in the answer. These are not merely new facts; they are changes in search strategy and evidence judgment.

Important caveats

The results are promising, but they do not demonstrate unrestricted autonomous improvement. The experiments learned only the repository of instructions, skills, and memory. The foundation models and harness remained fixed, while broader recursive improvement of those components is described as future work.

The evaluation is also concentrated in one synthetic legal environment and uses model-based judging. The authors revised 46 potentially inconsistent benchmark criteria after comparing them with task statements and source documents. They report results under both rubric versions, but this intervention makes careful interpretation important: some measured changes reflect the chosen evaluation criteria as well as the agent’s behavior.

Recursa also requires substantial inference and parallel-agent computation. Its performance depends on the quality of the source corpus and the capabilities of the underlying models. Errors, bias, or misleading material can be recorded in skills and memory and then reused. The authors further note privacy and security risks because the system explores software environments, executes programs, and retains observed information, even though their experiments used isolated sandboxes and controlled access.

Practical implications

For teams building domain-specific agents, Recursa points toward a practical progression: first build a searchable map of the workspace, then capture procedural skills, and finally turn task traces into validated improvements. This can be useful where expert-written examples are scarce but documents, execution histories, and partial feedback are available.

The broader lesson is that adaptation may be a systems problem before it is a weight-training problem. A carefully designed repository, execution environment, validation loop, and memory policy can make a general model more reliable in a local domain. Whether that approach transfers beyond the benchmark—and whether it can remain safe as the system improves itself—will require evaluations across real domains, stronger safeguards, and evidence that learned changes remain correct as the underlying environment evolves.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close