GEAR: Scikit-learn Style Modular Framework for Recursive Self-Improvement of LLM Agents

Research Significance: A Modular Toolkit for RSI

GEAR (General Evolutionary Architecture for Recursive Self-Improvement) is an open-source framework designed to provide a universal toolbox for Recursive Self-Improvement (RSI) of LLM agents. Positioned as toolkit for RSI , it enables developers to compose, swap, and compare different evolution strategies while reusing existing task execution, evaluation, and version management capabilities. The framework decomposes parent version selection, modification planning, task sampling, scoring, and candidate ranking into replaceable modules, with shared infrastructure recording each candidate's code version, run conditions, and evaluation evidence for inspection and further experimentation.

The preprint GEAR: An Open General Evolutionary Architecture for Recursive Self-Improvement of LLM Agent Harness (October 26, 2026) is authored by Yehui Tang, Guanqi Zhu, Shao-Jun Xia, Xin Chen, Mengzhe Geng, Hongyang Du, and Yuhao Tang. Resources: project homepage rsi-gear.github.io, source code github.com/rsi-gear/gear, paper https://huggingface.co/spaces/WizardTY/GEAR.

Background: Three Core Problems in Self-Evolution

Regardless of the specific evolution approach (prompt optimization, workflow evolution, code modification, model training), all RSI methods face three common challenges:

Sampling: The search space is large but compute budget is limited. Deciding which existing versions to start from, which improvement directions to explore, and which tasks to prioritize for feedback determines how far a search can go.

Variant Evaluation: Generating a new version is only the beginning; the critical step is judging whether it actually improved. Score increases may stem from task difficulty variance, reasoning tier fluctuations, or run noise. Evaluation must control cost while providing trustworthy feedback for retention or rejection decisions.

Diversity Management: Keeping only the highest-scoring version each round risks premature convergence. A temporarily lagging variant might excel on a subset of tasks or use a different solution strategy. Preserving valuable differences so they can seed future evolution is essential for sustained exploration.

These three problems — sampling, variant evaluation, and diversity management — form a common basis for understanding and comparing self-evolution methods. Agent harness evolution serves as a concrete research scenario: prompts, tools, skills, workflows, and context management are mutable while the target model weights stay fixed.

Framework Design: Meta-Learning Based Two-Layer Architecture

GEAR models agent self-improvement as a meta-learning problem: improving the method for solving future tasks from execution experience on a batch of tasks. The system splits into two layers:

Inner layer (task-level execution and adaptation): The task agent uses a given model and harness to interact with the environment, call tools, and adjust the current task's plan based on feedback. Execution trajectories, final results, and evaluation feedback are recorded. Temporary adjustments in this layer serve only the current task and do not automatically become rules for subsequent tasks.

Outer layer (cross-task meta-learning): The meta agent aggregates trajectories and evaluation evidence across multiple tasks, identifies recurring problems, and proposes modifications to prompts, tools, skills, workflows, or context management. Candidate versions undergo evaluation and selection before becoming the new harness for future tasks.

The two layers form a loop: task execution produces experience, cross-task experience drives framework improvement, new framework participates in subsequent tasks and produces new experience. A temporary fix in one task only enters this continuous accumulation process after being codified as a verified framework modification.

The aforementioned sampling, variant evaluation, and diversity management can all be configured with different strategies in the outer layer: selecting which historical versions to branch from, how to evaluate candidates, and which versions to retain for further exploration. The paper implements a concrete instance using phased GEPA (Genetic-Pareto) search, but these strategies are swappable within the shared framework. This meta-learning manifests as a two-layer structure where harness evolution keeps target model weights fixed and improves via search and selection over the runtime framework, without requiring the inner solving process to be differentiable.

Experimental Results and Analysis

Fixed Model and Reasoning Tier: Gains Across Three Domains

Experiments used AutomationBench (100 Marketing tasks, 100 Support tasks) and Terminal-Bench 2.1 (89 command-line tasks). Each model-harness configuration was evaluated once per task. The primary controlled comparison fixed GPT 5.6 Luna at medium reasoning tier, comparing initial DeepSeek Harness (DSH) against DSH evolved for each dataset.

Marketing pass rate: 27.0% → 40.0% (+13 tasks out of 100)

Support pass rate: 17.0% → 25.0%

Code pass rate: 52.8% → 62.9%

Goal completion rate (average proportion of satisfied scoring criteria per task) also improved: Marketing 75.4% → 83.9% , Support 72.9% → 81.3% . Both full-task pass rate and partial completion improved.

Evolved Harness Transfers Across Models

Harnesses evolved on Luna were tested on GPT 6 Astra at max reasoning tier. Marketing and Support pass rates reached 61% and 39% versus Codex baselines of 57% and 35% , providing preliminary evidence that improved working methods transfer across models. On Terminal-Bench 2.1, evolved harnesses with Luna max and Astra max achieved 84.3% and 87.6% pass rates respectively.

Explicit Acceptance Rules Prevent Regressions

The Marketing evolution trajectory (Figure 2 right) shows why explicit acceptance mechanisms are necessary. Initial version: 27.0% pass, 75.4% goal completion. Round 7 final version: 40.0% pass, 83.9% goal completion. However, the path was not monotonic: rounds 1, 5, and 6 candidates were rejected.

Round 5: Goal completion rose from 81.2% to 82.3% but full-task pass rate dropped from 36% to 34%. Higher average completion does not guarantee more fully completed tasks. This candidate was not adopted.

Round 6: Pass rate remained 36% but goal completion fell to 80.6%; also rejected.

The system's explicit rules blocked these regressions, retaining only versions satisfying acceptance criteria. This illustrates GEAR's recording and evaluation infrastructure: letting researchers see gains and losses during improvement, inspect what a modification actually changed, rather than just receiving a final score.

Future Directions: From Research Toolbox to Business Process Optimization

For researchers, GEAR serves as an experimental platform to compare self-improvement methods. On shared task execution and evidence recording infrastructure, researchers can try different search strategies, task sampling schemes, and acceptance rules, analyzing which methods are more effective under the same budget.

For developers, a promising direction is integrating business test suites into this loop. For example, building repeatable task sets for customer support or DevOps workflows, using failure trajectories to improve skills and workflows, then deciding on new versions via regression evaluation. Human experts can review modification rationales and run logs to participate in analysis and annotation.

For real business scenarios, the framework's toolkit for RSI architecture enables rapid construction of targeted algorithms across different scenarios, achieving performance comparable to top-tier large models with smaller models on customer service, marketing, and coding tasks.