Summary Trampoline-AI/predict-rlm offers a production-ready runtime for self-harnessed Recursive Language Models (RLMs), enabling LMs to invoke sub-LMs via DSPy signatures. It tackles context rot and facilitates complex, long-horizon tasks by automating control flow, yielding interpretable trajectories and performance that scales with base model advancements. This system is designed for technical users aiming for efficient LM utilization, cost reduction, and enhanced capabilities, even with smaller models.

How It Works Based on MIT CSAIL's Recursive Language Models (RLMs), this runtime allows a root LM to programmatically call sub-LMs using DSPy signatures, managed via a REPL. This architecture prevents context degradation by keeping the root LM within its operational range. RLMs provide symbolic reasoning, where single lines represent complex computations, and offer fully traceable execution paths for optimization. Performance, speed, and cost directly correlate with base LM improvements, ensuring scalability and efficiency.

Quick Start & Requirements Installation involves uv add predict-rlm or npx skills add Trampoline-AI/predict-rlm for compatible coding agents. It requires a Python environment and DSPy. Specific skills may necessitate additional packages like openpyxl and pandas. No specific hardware prerequisites (e.g., GPU) are mandated in the documentation.

Highlighted Details

• Supports multimodal inputs (images, documents, audio, video) via native provider APIs.
• Features asynchronous tool calling within a WASM sandbox.
• Includes prompt-optimized skills and tools, bundled with necessary dependencies.
• Provides simple file I/O using a File type (S3 support planned).
• Enables structured, type-safe sub-LM calls with Pydantic and DSPy signatures.
• Offers pre-built skills like pdf and spreadsheet for common workflows.

Maintenance & Community Developed by Trampoline AI. The provided README snippet lacks specific details on maintainers, community channels (Discord, Slack), roadmaps, or sponsorships.

Licensing & Compatibility The licensing terms for predict-rlm are not specified in the provided README content. This omission is a significant caveat for potential adopters, particularly concerning commercial use or integration into closed-source projects.

Limitations & Caveats Support for S3 cloud file storage is noted as "coming soon." The absence of a stated license is a critical adoption blocker requiring clarification before use in production or commercial environments.