Reproduce “Recursive Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alex L. Zhang, Tim Kraska, Omar Khattab
Recursive Language Models (RLMs) keep a long prompt in a program environment, letting a root model inspect, transform, and slice it before calling language models on selected pieces. On 32K-token OOLONG-Pairs, GPT-5 with depth-1 RLM scores 58.0 F1, versus 0.1 directly and 28.4 with CodeAct. This adds cost and does not guarantee performance on all tasks.
Key insights
- Symbolically externalizing the user prompt and interacting with it via a REPL provides an effectively unbounded input, output, and semantic horizon for LLMs.
- Programmatic recursive sub-calling is crucial for information-dense tasks requiring complex semantic transformations and aggregations, enabling multi-step reasoning over large inputs.
- RLMs can generate unbounded outputs by constructing results in REPL variables, circumventing the underlying LLM's inherent output length constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.