PARSER freezes the readers and trains only the lead agent, gaining 12 points at 896K tokens of context

arXiv (2609.06702)·medium signalXBlueskyLinkedInCopy link

arXiv 2609.06702 splits long-context work into a bank of lightweight frozen subagents, each bound to a single chunk and reading in parallel, and one lead agent that runs iterative scatter-gather rounds, broadcasting a query, aggregating evidence and forming a deeper follow-up. All learnable behavior concentrates in the lead agent, which is trained with RL while the subagents stay off-the-shelf. With a 4B backbone it beats the strongest sequential memory baseline by 5.7 points on average across 7K to 896K token contexts, and by 12.0 points at 896K.