Recursa: Toward Recursively Self-Improving Foundation Model Systems

Recursa adapts a foundation model system to a new domain without human experts or a stronger teacher model by bootstrapping a knowledge repository of instructions, skills, and memory from domain materials, then refining it through parallel multi-agent recursive self-improvement, raising all-pass rate on Harvey LAB's Firm Knowledge benchmark from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

Foundation models encounter unfamiliar domain knowledge in changing environments, but new domains rarely provide large amounts of expert-created training data or reliable supervision for further training. The paper frames foundation model system learning as improving a system comprising a model, instructions, skills, memory, and a harness, which must close two gaps: a knowledge gap (missing facts, terminology, or knowledge of where materials are located) and a behavior gap (having the information but applying it incorrectly). Knowledge-intensive domains such as law and finance make this harder because knowledge is scattered without common organization, behavioral expectations are highly local, and correctness is hard to verify automatically. Existing methods learn only parts of the system and most are evaluated on coding, mathematics, and games, often relying on a stronger model to drive learning.

Recursa uses two agent roles on a shared harness: solvers execute domain tasks using the current learnable components K and return an answer plus execution trace, while learners inspect answers, traces, and feedback to develop and validate revisions, optionally decomposing learning across child learners and merging their results. The harness lets the model act in code: the model writes Python actions executed in a persistent REPL, with tools loaded as callable functions, subagents callable for isolated parallel work, and full interaction history kept in an append-only event log queryable from code so only selected evidence enters context. The knowledge repository (instructions, skills, and memory) is bootstrapped from the domain corpus by building a navigation structure and practicing on self-generated tasks grounded in the materials, with sparse guidance such as preferences, example rubrics, or known failure modes steering generation and assessment. Recursive self-improvement then calls Learn(D, K) on training tasks: a leaf learner loops through Solve (parallel solvers), Diagnose, Edit, and Validate, keeping an edit only if it fixes failures without causing regressions, and a splitting learner divides tasks into subsets, issues learning tasks to children in parallel, merges their revisions, and reflects on their learning traces into memory. In this paper only the knowledge repository is learned; the harness and foundation models stay fixed.

• Recursa raises all-pass rate on held-out test tasks with both solvers, without a stronger model revising it. §Abstract
• The harness alone, without learning, beats both the published base agent and the highest-accuracy study configuration. §6
• Bootstrapping produces 14 skills and reduces interaction turns and input tokens by roughly a fifth to a quarter while raising all-pass rate slightly and leaving mean criteria pass rate unchanged. §6.3
• With DeepSeek-V4.1-Flash, multi-learner components K^(2) raise training all-pass from the bootstrapped level to 48.7% and test all-pass from 33.0% to 55.0%, 25 percentage points above the agent without learning. §6.4
• With Qwen3.8-Flash, merged components K^(1) raise training all-pass from 37.3% to 47.3% and test all-pass from 34.0% to 46.0%, 11 percentage points above the agent without learning. §6.4
• The benchmark environment and rubric revision scope are described quantitatively. §6.1

• The paper states performance relies on the quality and availability of domain-specific corpora, which may be limited or unavailable in certain domains, and that substantial computational resources for large language model inference and training may limit accessibility (§8).
• The paper states incorrect or biased data and flaws in the underlying models can propagate through the learning process, and that effectiveness is contingent on the capability of the underlying models, with the impact of model limitations not investigated (§8).
• The paper states experiments focus on a limited set of domains and may not capture performance across all possible domains; extending recursive self-improvement to the foundation models and the harness is left as ongoing/future work, since only the knowledge repository is learned here (§2.2, §7, §8).
• The paper states Recursa autonomously explores software environments, executes programs, and retains observed information in memory, introducing privacy and security risks; experiments were run in controlled sandboxes with no network access except controlled channels for inference and training (§8).

• Learning must close a knowledge gap (missing domain facts, terminology, or knowledge of where materials are located) and a behavior gap (having the information but applying it incorrectly). §1
• Two phases: bootstrapping from domain materials and self-practice tasks, then recursive self-improvement from training tasks, execution traces, and available feedback. §1
• The harness executes model-written Python in a persistent REPL, so tools compose via loops, conditionals, and functions and intermediate results stay in variables rather than the context. §3.1
• Full interaction history is kept in a persistent append-only event log outside the model context and retrieved by code; the context holds only recent history and an eviction index of retrievable ranges. §3.2
• Multi-agent orchestration lets an agent dispatch subagents to process parts of a workload in parallel and merge their findings. §3.3
• The leaf learner loops Solve, Diagnose, Edit, and Validate, keeping an edit only if it fixes failures without causing regressions, and stopping when no edit improves results or the budget is spent. §5.1
• Learned judgment is illustrated on two test tasks: searching by meaning maps a legal concept to plausible formulations (Exclusivity, No Solicitation) and judging candidates by substance, and deciding what belongs uses engagement records to define scope before checking settlements. §6.5
• 46 criteria in 30 tasks conflicted with task statements or source documents; revisions changed criterion text only, leaving task statements, criterion IDs, criterion counts, and the other 220 tasks unchanged. §6.1