RLM paper: GPT-5-mini-based RLM doubles GPT-5 accuracy on OOLONG, handles 10M+ tokens
gordic_aleksa · x · 2026-09-20
The Recursive Language Models (RLM) paper (paper, code) proposes an inference strategy where LMs treat their input context as a variable in a Python REPL, recursively decomposing and interacting with it to handle essentially unbounded context and mitigate context rot.
Key results:
- On OOLONG, the hardest long-context benchmark they found, an RLM using GPT-5-mini more than doubles GPT-5's correct answers at lower average cost per query
- On a new long-context Deep Research task built from BrowseComp-Plus, RLMs beat ReAct + test-time indexing/retrieval
- Surprisingly, performance doesn't degrade at 10M+ tokens
The authors argue RLMs explicitly trained to reason recursively may be the next milestone in inference-time scaling after CoT and ReAct-style agents.
Related event: Recursive Language Models Let LMs Call Themselves on Long Contexts(2 posts)→
More from Research
- Self-Rewarding LLMs isn't news: RLAIF has long been standard practice — burny_tech · 2026-09-20
- Long-standing Catalan constant irrationality proof posted, claimed to be LLM-assisted — burny_tech · 2026-09-20
- Brain Runs on 20 Watts: Can Neuromorphic Computing Make AI Less Power-Hungry? — burny_tech · 2026-09-20
- Amid the AI math proof debate, a curated set on proofs across generations — RexDouglass · 2026-09-20
- Four LLMs play Doom: Jev leads with 5.63 mean kills but 15x higher latency — shniydder · 2026-09-20
- Odyssey-3: One Pretrained World Model Adapts to Different Robot Arms With Hours of Demos — ChongZzZhang · 2026-09-20