Daniel Chen

A runner knows the route. Ours learned it with the professionals.

meta: my chat with Claude got too long while drafting this critique of the RLM paper. Claude couldn't fit the full conversation in context. so it grepped the local transcript file and pulled in relevant sections. context as external variable, examined and retrieved programmatically... wait, my Claude is already doing RLM? the paper (@a1zhang, @lateinteraction). the core problem is real: models need clean separation between the context they're reasoning over and the intermediate results of exploring that context. tool outputs and sub-call results shouldn't pollute the window you're thinking in. context rot from accumulated junk is a genuine failure mode. but this divide-and-conquer is already happening at the harness level and useful patterns are being RLed into models. plan mode → external checklist → Ralph Wiggum loops working through tasks one at a time with fresh context. subagents returning distilled results so junk never hits the parent window. context-driven file exploration (check length, grep structure, selectively read)... do the above well and each sub-task gets a focused window with mostly relevant context. this is where RLM's recursive approach actually costs you — every sub-call is a fresh prefill with no KV cache sharing, plus scaffolding overhead. when context is mostly relevant and fits in window, a warm cache with full cross-context attention wins outright. the training contribution is clean RL env design: the model can't read long snippets from the prompt, forcing it to learn selective exploration and recursive decomposition. but existing coding tools already impose the same constraint — Claude Code's read tool rejects files over ~25k tokens. models are already learning context decomposition because their harness tooling forces it when being RLed. for frontier models, the path forward is better divide-and-conquer, better tool use for external context — transcripts, persisted state files, disk artifacts — and better RL for learning when to decompose. not a new paradigm. All these are already underway. some RLM patterns are already there, as the opening makes clear.