RLMs: processing long prompts through a REPL

What happened and what to do

Published on October 15, 2025, the description presents Recursive Language Models (RLMs), an inference strategy that lets language models decompose and interact recursively with very long prompts through a REPL. It reports results on the OOLONG and BrowseComp-Plus benchmarks but gives no scores. The approach is presented as a possible alternative to explicit retrieval for tasks involving long contexts. To verify the details, consult the public publication under the title “Recursive Language Models processam prompts longos por meio de um REPL” and check its stated scope and reported results.

A company can assess whether its workflows truly require loading large amounts of context or whether a recursive approach in a controlled environment is a better fit. Implementation could include a prototype, execution and access limits, tests using representative tasks and data, and measures for quality, latency, and cost. The choice between a REPL, retrieval, and other approaches should be validated against the company’s operating context.

How the consultancy can help

Wendelmaques can diagnose long-context use cases, define security and evaluation criteria, and propose a scoped implementation of recursive inference or alternatives such as retrieval. Scope can include a prototype, integration, monitoring, and an operating plan on the client’s infrastructure.

Next step

Send a brief description of the task that needs to handle long context to receive a proposal covering diagnosis and implementation scope.