ReCode revolutionizes LLM agents by unifying planning and action via recursive code generation, enabling adaptive, human-like decision-making. It targets researchers and developers seeking universal granularity control, allowing agents to dynamically decompose complex tasks from high-level strategies into executable code primitives. This approach yields significant performance gains and superior data efficiency.

How It Works

The system employs a divide-and-conquer strategy, organizing partial programs in a tree structure where each node represents a sub-task. LLMs recursively expand placeholder functions into specific calls or subroutines using environment-specific prompts and few-shot examples. A dynamic execution loop immediately runs each node, with fresh observations guiding further expansion, retries, or completion. A constrained Python executor manages shared state, validates code, and exposes tools, facilitating robust, adaptive execution from strategic planning to concrete actions.

Quick Start & Requirements

Setup requires Python 3.10+ within a conda environment; separate environments are advised for ALFWorld, ScienceWorld, and WebShop due to potential conflicts. Key steps include:

• Creating/activating a conda environment (e.g., conda create -n recode-envname python=3.10).
• Configuring environment-specific assets (ALFWorld data, WebShop setup.sh, pip install -e ., conda install openjdk=11, pip install en_core_web_lg).
• Setting up LLM access via configs/profiles.yaml.
• Running evaluations with python run.py (e.g., python run.py -a recode -e alfworld -n 1 --profile default). Prerequisites include specific dataset paths and LLM API access.

Highlighted Details

• ReCode achieved significant performance improvements across ALFWorld, WebShop, and ScienceWorld, outperforming baselines by 20.9% on average.
• It reached a perfect 100 score in ALFWorld using claude-4-sonnet.
• Supervised fine-tuning (SFT) demonstrated exceptional data efficiency, with ReCode+SFT outperforming ReAct+SFT and CodeAct+SFT.

Maintenance & Community

Contact zhaoyangyu713@gmail.com for technical assistance. No community channels (Discord, Slack) or public roadmap are listed.

Licensing & Compatibility

The README omits a software license, preventing an assessment of its compatibility for commercial use or integration into closed-source projects.

Limitations & Caveats

Users may face dependency conflicts between environments, requiring separate conda installations. The WebShop requirements.txt might be incomplete, potentially necessitating direct contact with the maintainer. The absence of a stated license is a significant adoption blocker.