Trampoline-AI/predict-rlm

Production focused Self-harnessed LM runtime (RLM) that allows the LM to call its sub-lm with DSPy signatures. Define your inputs, outputs, and tools — the model handles its own control flow. Get fully interpretable trajectories and performance that scales directly with model improvements. Without context rot.

What it solves

Many LLM workflows are too complex for a single prompt and too adaptive for a fixed chain of steps. predict-rlm provides a runtime for these complex workflows, allowing models to explore, reason, and adapt their path to a solution rather than following a rigid, pre-defined sequence.

How it works

Based on the Recursive Language Models (RLM) paper, the system allows a model to write and execute Python code within a sandboxed REPL. The user defines the inputs, outputs, tools, and an operating procedure (via signatures). The model then iteratively writes code to inspect files, manage state, call focused sub-models for specific tasks (like image analysis or data extraction), and use tools until it can produce the final typed output.

Who it’s for

It is designed for developers building applications that require deep investigation and adaptive reasoning, such as:

  • Codebase analysis and incident investigation.
  • Document review, redaction, and extraction.
  • Log analysis and evidence gathering.
  • Financial modeling and spreadsheet workflows.
  • Audits and compliance reviews.

Highlights

  • Sandboxed Execution: Executes model-generated Python in a secure REPL (supporting JSPI/Deno/Pyodide or Docker Sandboxes).
  • Context Management: Avoids "context rot" by keeping large inputs as file paths and metadata, only inspecting them as needed.
  • ** uma Focused Sub-calls**: Uses predict() to spin up typed DSPy signatures for narrow perception or extraction tasks, including multimodal support.
  • Structured Traces: Records every iteration, tool call, and sub-call for debugging and optimization via tools like GEPA.
  • Skill System: Allows bundling of domain instructions, PyPI packages, and tools into reusable "skills" (e.g., built-in spreadsheet and PDF skills).

Related

  • Project
    alexzhang13/rlm

    An inference engine and training environment for Recursive Language Models (RLMs) that enable LLMs to handle near-infinite context by programmatically decomposing tasks via recursive sub-calls in a REPL environment.

  • Project
    grishahq/recursive-llm

    A Python implementation of Recursive Language Models that allows LLMs to analyze large contexts by interacting with them via a restricted Python REPL instead of placing the full text in the prompt.

  • Project
    avbiswas/fast-rlm

    A minimal implementation of Recursive Language Models (RLMs) that allows LLMs to handle arbitrarily long prompts by programmatically exploring data via a Python REPL and recursive sub-agents.

  • Project
    IBM/prompt-declaration-language

    A declarative YAML-based language for creating reliable, composable LLM prompts and integrating them into software systems with support for control flow and code execution.

  • Project
    chr15m/runprompt

    A command-line tool to execute .prompt files, allowing users to treat LLM prompts as versionable artifacts and automate AI workflows via the shell.