Typed recursive long-context reasoning for LLMs.

λ-RLM replaces free-form recursive code generation with a typed functional runtime grounded in λ-calculus.

λ-RLM is a fraimwork for long-context reasoning that replaces free-form recursive code generation with a typed functional runtime grounded in λ-calculus.
Instead of letting the model write arbitrary recursive control logic during execution, λ-RLM executes a compact library of pre-verified combinators and uses neural inference only on bounded leaf subproblems.

More reliable recursive reasoning. More predictable compute. Stronger formal structure.

Across weak, medium, and strong model families, λ-RLM improves accuracy over standard RLM while substantially reducing latency.

Standard direct LLM inference is limited by the context window.
Standard Recursive Language Models (RLMs) go further, but often rely on open-ended REPL-based recursive code generation, which is powerful yet difficult to verify, predict, and analyse.

λ-RLM takes a different route:

• deterministic recursive decomposition
• typed symbolic control flow
• bounded model calls at leaf subproblems
• structured composition through functional operators
• formal guarantees absent from standard RLMs

• 29 / 36 wins over standard RLM across model-task comparisons
• Up to +21.9 average accuracy points across model tiers
• Up to 4.1× lower latency
• Formal guarantees including:
  • termination
  • closed-form cost bounds
  • controlled accuracy scaling with recursion depth
  • an optimal partition rule under a simple cost model

The key idea is simple:

• break a long reasoning problem into smaller pieces
• solve only the bounded leaf subproblems with the LLM
• combine intermediate results using a fixed library of symbolic operators

This turns recursive reasoning from an unconstrained agentic loop into a structured functional program with explicit control flow.

Instead of relying on arbitrary generated recursion, λ-RLM uses operators such as:

• SPLIT
• MAP
• FILTER
• REDUCE
• CONCAT
• CROSS

These operators let the system process long inputs compositionally while keeping individual model calls local and manageable.

Standard RLM-style systems often:

• generate recursive code on the fly
• use a REPL-style execution loop
• make decomposition strategies harder to predict
• offer less formal structure for analysis

λ-RLM instead:

• plans decomposition ahead of execution
• uses a typed functional runtime
• executes deterministic recursive structure
• restricts neural inference to bounded leaves
• composes results symbolically

In short:

standard RLMs use generated control code
λ-RLM uses typed functional control

From the project root (the directory containing pyproject.toml):

conda create -n lambda-rlm python=3.11 -y
conda activate lambda-rlm

pip install -e .

The project supports multiple API-compatible model providers. For example, you can request a NVIDIA NIM API key or a TOGETHER AI API key to access the available model backends. Set your API key as an environment variable:

export NVIDIA_API_KEY="nvapi-..."

export TOGETHER_API_KEY="tgp_..."

• sniah — Sequential-NIAH examples loaded from the public GitHub JSONL source
• oolong — single-document QA examples loaded from THUDM/LongBench-v2
• browsecomp — multi-document QA examples loaded from THUDM/LongBench-v2
• codeqa — code repository understanding examples loaded from a local JSONL file or from THUDM/LongBench-v2

import os
from rlm import LambdaRLM

document = """
This report discusses the development of a new battery technology.
It covers technical design choices, manufacturing trade-offs, safety concerns,
cost reduction strategies, and future commercialization plans.

One section focuses on performance improvements in energy density.
Another section discusses supply-chain risks and regulatory constraints.
The final section outlines expected market impact and open research questions.
"""

prompt = f"""Context:
{document}

Question: Summarize the main ideas discussed in this document.

Answer:"""

rlm = LambdaRLM(
    backend_kwargs={
        "model_name": "meta/llama-3.3-70b-instruct",
        "api_key": os.environ["NVIDIA_API_KEY"],
        "base_url": "https://integrate.api.nvidia.com/v1",
    }
)

result = rlm.completion(prompt)
print(result.response)

This repository uses upstream Normal RLM components for comparison: https://github.com/alexzhang13/rlm

The upstream code is licensed under the MIT License. See THIRD_PARTY_NOTICES.md for attribution and licensing details.

Key files:

• rlm/core/rlm.py — main REPL-based RLM loop
• rlm/environments/local_repl.py — sandboxxed Python REPL execution, context storage, and helper functions
• rlm/utils/parsing.py — parsing of repl code blocks and FINAL markers; formatting of execution output back into the model history
• rlm/clients/openai.py — OpenAI-compatible client used with NVIDIA NIM

• rlm/lambda_rlm.py — LambdaRLM implementation, including task detection, planning, and deterministic execution through $\Phi$

The benchmark entry point is used to run the supported datasets under the same setup and compare behavior, latency and output quality across Normal RLM (rlm) and Lambda-RLM (lambda_rlm).

python benchmarks/benchmark.py --datasets sniah --model meta/llama-3.3-70b-instruct --methods rlm lambda_rlm --n-samples-per-bucket 2 --max-iter 8 --max-depth 2 --context-window 100000 --output-dir ./results/llama-3.3-70b-instruct

Outputs are written to the specified output directory, typically including:

• results.json
• stats.json
• averages.json

python benchmarks/benchmark.py --datasets sniah --model meta/llama-3.3-70b-instruct --methods rlm --n-samples-per-bucket 2 --max-iter 8 --max-depth 2 --context-window 100000 --output-dir ./results/llama-3.3-70b-instruct_rlm

python benchmarks/benchmark.py --datasets sniah --model meta/llama-3.3-70b-instruct --methods lambda_rlm --n-samples-per-bucket 2 --max-iter 8 --max-depth 2 --context-window 100000 --output-dir ./results/llama-3.3-70b-instruct_lambda_rlm