This is a brilliant paper.
It's of the cleanest long-context agent designs I have seen in the past couple of months.
Sequential memory agents read chunks one after another while maintaining a compact memory state.
This behavior ties reasoning depth to document traversal and makes accuracy sensitive to where the evidence sits. It also makes latency grow linearly with document length.
PARSER decouples the two.
A bank of lightweight subagents, each bound to a single chunk, reads the whole document in parallel.
A lead agent reasons through iterative scatter-gather rounds, broadcasting a query to all subagents, aggregating the returned evidence, and forming a deeper follow-up query conditioned on what it has found.
All the learnable behavior is build into the lead agent, which is trained with RL. The subagents stay frozen off-the-shelf models.
On multi-hop QA from 7K to 896K tokens, a 4B PARSER beats the strongest sequential memory baseline by 5.7 points on average and 12.0 points at 896K. At 9B it passes DeepSeek-V4-Pro by 6.3 points.
Controlled experiments show it holds up under perturbations to evidence position, order and distance, which cause large accuracy swings in sequential methods, while cutting inference latency by up to 11x.