DISCO: Distributed Long Context Scaling with Grounding-Reasoning Disaggregation
The paper’s claim is that splitting long-context work between grounding “workers” and a reasoning “driver” can prevent context rot at million-token scale.
DISCO partitions the input across Worker LLMs that extract localized evidence, while a central Driver LLM plans the tasks and synthesizes the answer. The authors say the Driver is trained with GRPO to improve that orchestration. On RULER-QA at 1M tokens, they report 78.4% accuracy as standard baselines collapse. They also claim gains of up to 9.8 points on LongBench v2 and more than 80% lower inference cost versus full-context approaches. HF Daily Papers' note
DISCO partitions the input across Worker LLMs that extract localized evidence, while a central Driver LLM plans the tasks and synthesizes the answer. The authors say the Driver is trained with GRPO to improve that orchestration. On RULER-QA at 1M tokens, they report 78.4% accuracy as standard baselines collapse. They also claim gains of up to 9.8 points on LongBench v2 and more than 80% lower inference cost versus full-context approaches. HF Daily Papers' note
score 5