PARSER trains one agent, freezes the rest, gains 12 points

PARSER splits reading a long document across a bank of frozen subagents, then trains only the lead agent that directs them.

That split matters because retraining every part of a long-context system is what makes it expensive to scale. PARSER keeps all the learning in one coordinator and leaves the readers as off-the-shelf models. Adding more subagents doesn't mean retraining more of them.

Each subagent reads one chunk of the document in parallel and never changes. The lead agent runs scatter-gather rounds. It broadcasts a question to the subagents, gathers what comes back, and forms a deeper follow-up, repeating until it has enough evidence to answer. Reinforcement learning trains that lead agent's behavior; nothing else in the system does.

With a 4B-parameter backbone, PARSER beats the strongest sequential-memory baseline by 5.7 points on average across context lengths from 7,000 to 896,000 tokens. At the top end, 896,000 tokens, the gap grows to 12.0 points, according to the PARSER paper.

The gap widens at the longest context. Sequential memory systems tend to degrade as documents grow, since each new chunk has to be folded into whatever the system remembered so far. PARSER's readers don't carry that burden forward. They just answer what the lead agent asks. If that holds outside this benchmark, most long-context agent products are training their readers when they should be training the one agent that queries them.