DISCO partitions long inputs among Worker LLMs that perform localized evidence extraction in parallel. A central Driver LLM decomposes queries into atomic tasks, gathers the extracted evidence, and iteratively replans before producing an answer. The Driver is trained with GRPO to improve planning.
The system reports 78.4% accuracy on RULER-QA at 1 million tokens, gains of up to 9.8 points over full-context models on LongBench v2, and performance matching Gemini-3-Pro-Preview while reducing inference costs by over 80%. The interesting claim is that scaling grounding separately from reasoning can avoid the quality degradation seen when one model must search and reason over the entire context.