A team at the Chinese University of Hong Kong has published a paper on PARSER, a system that splits long-document question answering into two separate problems: reading every chunk of a document in parallel, and reasoning across those chunks in depth. The result is a 4-billion-parameter model that outperforms much larger sequential baselines on multi-hop QA tasks, with the advantage growing as documents get longer.
Why Sequential Agents Struggle With Long Documents
The standard approach to long-document QA is a sequential memory agent. It reads chunks one after another, maintaining a compact memory state that compresses what it has seen. The problem is that reasoning depth and document traversal are coupled. To reason deeply about evidence buried at the end of a 900,000-token document, the agent must first process everything before it. That makes accuracy sensitive to where evidence appears in the document and ties inference latency linearly to document length.
If the answer to a multi-hop question sits in the first chunk and the last chunk, a sequential agent has to read the entire document before it can connect the two pieces. Evidence placed early gets compressed into the memory state by the time the agent reaches the relevant later chunks. Order matters too: the same facts presented in a different sequence can produce different answers because the memory state evolves differently.
Parallel Reading, Iterative Reasoning
PARSER breaks this coupling. A bank of lightweight subagents, each frozen and bound to a single chunk, reads the entire document at once in parallel. A lead agent then reasons across the evidence through iterative scatter-gather rounds. At each round, the lead agent broadcasts a query to all subagents, collects the returned evidence, and formulates a deeper follow-up query conditioned on what it has found.
The key architectural decision is that only the lead agent is trained. It is optimized with reinforcement learning. The subagents remain frozen off-the-shelf models. This keeps the trainable surface small: one model, one set of parameters, one optimization target. The subagents do not need to learn anything about the task. They just need to be able to read a chunk and return relevant evidence for whatever query the lead agent sends them.
The scatter-gather mechanism means the lead agent can iteratively refine its understanding without re-reading the entire document. The first round might pull surface-level matches from every chunk. The second round, informed by what the first round found, can ask more specific questions. The third round can synthesize across the evidence gathered so far. Each round narrows the search without requiring the subagents to coordinate with each other.
The Numbers at Scale
Tested on multi-hop QA with contexts ranging from 7,000 to 896,000 tokens, a 4B-parameter PARSER outperforms the strongest sequential memory baseline by 5.7 points on average. At 896,000 tokens, the gap widens to 12.0 points. Scaling to a 9B backbone, PARSER surpasses DeepSeek-V4-Pro by 6.3 points.
The 12-point advantage at 896K tokens is the result that matters most. It shows that the parallel-read advantage compounds with document length. At shorter contexts, sequential agents perform adequately because they can traverse the document quickly enough. At 896K tokens, the sequential agent is processing hundreds of chunks in order, and the compression of early evidence into memory becomes a binding constraint. PARSER's parallel approach avoids that constraint entirely.
Robustness Under Perturbation
Controlled experiments tested what happens when you change where evidence appears in the document, what order chunks are presented in, and how far apart related evidence is. Sequential methods show large accuracy swings under these conditions. PARSER stays stable.
That stability is a direct consequence of the architecture. Because every subagent reads its chunk independently and the lead agent queries all of them, the position of evidence within the document does not affect what the subagents return. A fact in chunk 1 and a fact in chunk 500 are equally accessible to the lead agent on the first scatter-gather round. Sequential agents, which process chunks in order, are inherently sensitive to these factors.
Inference latency drops by up to 11x compared to sequential baselines. The parallel reading eliminates the linear dependence on document length. Whether the document is 10,000 or 900,000 tokens, the subagents all read simultaneously. The bottleneck shifts from document traversal to the number of scatter-gather rounds, which is a function of query complexity, not document length.
What This Means for Long-Context Applications
The practical implication is that long-document QA no longer requires choosing between accuracy and latency. Sequential agents force that tradeoff: you can read fast but lose depth, or read deeply but accept linear latency. PARSER reads everything at once and reasons iteratively, giving you both.
For teams building applications that process long documents, legal contracts, research papers, codebases, or financial reports, the architecture suggests a pattern. Freeze a set of reader models on document chunks. Train a single lead agent to query them. Let the lead agent handle the reasoning. The reader models do not need to be large or task-specific. They just need to extract relevant evidence for whatever the lead agent asks.
The small trainable surface is also a practical advantage. Reinforcement learning on a single lead agent is far cheaper than training or fine-tuning a full long-context model. Teams that want to adapt PARSER to a specific domain can train the lead agent on domain-specific queries without touching the reader subagents. The frozen subagents can be updated independently as better base models become available.
The paper is available at arXiv under identifier 2609.06702, authored by Kun Li, Zexuan Qiu, Tianhua Zhang, Irwin King, and Helen Meng at the Chinese University of Hong Kong.