PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents
PARSER splits long-document reading across chunk-bound subagents, then lets one trained lead agent drive deeper evidence gathering.
The paper says this breaks the link between document length and sequential reasoning steps. In multi-hop QA over 7K to 896K-token contexts, a 4B PARSER model beats the strongest sequential memory baseline by 5.7 points on average and 12.0 points at 896K tokens. With a 9B backbone, it also tops DeepSeek-V4-Pro by 6.3 points. The authors report better resistance to evidence position, order, and distance changes, plus inference latency reductions up to 11x. HF Daily Papers' note
The paper says this breaks the link between document length and sequential reasoning steps. In multi-hop QA over 7K to 896K-token contexts, a 4B PARSER model beats the strongest sequential memory baseline by 5.7 points on average and 12.0 points at 896K tokens. With a 9B backbone, it also tops DeepSeek-V4-Pro by 6.3 points. The authors report better resistance to evidence position, order, and distance changes, plus inference latency reductions up to 11x. HF Daily Papers' note
score 5