Recursive Language Models (RLMs) (alexzhang13.github.io)

Researchers introduced Recursive Language Models (RLMs), an inference strategy that lets an LLM decompose and recursively interact with arbitrarily large input context via an executable environment (they prototype a Python REPL). A “root” model call can peek at a context variable, run regex/grep-style narrowing, and spawn recursive sub-calls (to the same or smaller LMs) inside the REPL as if calling a function; when confident it returns FINAL(answer) or FINAL_VAR(var). The approach is designed to mitigate “context rot” — the empirically observed degradation in recall and reasoning as conversation/context length grows — without requiring individual model calls to absorb huge context windows. Technically, RLMs treat the prompt as programmatic data and offload partitioning, retrieval, and local reasoning into recursive model invocations, enabling adaptive chunking and learnable interaction policies (even RL-optimizable). In early benchmarks, an RLM built with GPT-5-mini beating GPT-5 by >33% on the hard OOLONG long-context split (128k–263k tokens) at similar API cost; ablating recursion drops performance ~10%. Preliminary BrowseComp-Plus experiments also show RLMs scale better across many documents where retrieval and single-call approaches struggle. Results are early but point to RLMs as a promising new axis of test-time scaling — a complementary path to larger windows or bigger models for long-context and multi-document reasoning.