The idea in one paragraph.
Treat a long input as external data: inspect and split it with code, and make recursive model calls over selected parts.
THE MACHINERY
Prompt as a REPL variable
The input is an addressable object, not tokens in the window.
Symbolic recursion
Code inside the REPL invokes the model over programmatically built slices, inside loops.
Metadata-only root context
Only constant-size metadata about the prompt and stdout enters the root's history.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Processes inputs more than an order of magnitude beyond model context window limits, and outperforms vanilla frontier LLMs even on shorter prompts.
Inspect a source excerpt
RLMs can successfully process inputs more than an order of magnitude beyond model context window limits and, even for shorter…Read the passage in context on arXiv ↗
On GPT-5, beats compaction by a median 26%, CodeAct with sub-calls by 130%, and Claude Code by 13% across four long-context tasks at comparable cost.
RLM-Qwen3-8B, post-trained on 1,000 samples from unrelated domains, beats base Qwen3-8B by a median of 28% and approaches vanilla GPT-5 on three long-context tasks.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Root iterations are still bounded by the root context window divided by the per-turn trim.
- Cost and latency scale with the number of sub-calls, which the model controls and can misjudge.
- Requires a sandboxed execution environment - a substantial harness dependency.
Why it belongs in this atlas.
The paper names the three design flaws that make an ordinary sub-agent scaffold weak: the prompt goes into the context window, the answer is verbalised rather than assembled in a variable, and recursion is verbalised rather than programmatic. Those three contrasts are the best teaching device in the corpus.