PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents
Kun Li, Zexuan Qiu, Tianhua Zhang, Irwin King, Helen Meng
The Chinese University of Hong Kong
Published September 6, 2026 · on the feed September 10, 2026 · ▲ 25 · arXiv 2609.06702
The abstract
Sequential memory agents process long documents by reading chunks one after another while maintaining a compact memory state, coupling document traversal to reasoning depth. This coupling introduces sensitivity to evidence placement and ties inference latency linearly to document length. We introduce PARSER, which decouples reading from reasoning. A bank of lightweight subagents each bound to a single chunk read the entire document in parallel, while a lead agent reasons in depth through iterative scatter--gather rounds: at each round it broadcasts a query to all subagents, aggregates the returned evidence, and formulates a deeper follow-up query conditioned on what has been found so far. This decoupled design concentrates all learnable behavior in the lead agent, which is optimized with reinforcement learning, while the subagents remain frozen off-the-shelf models. On multi-hop QA with contexts ranging from 7K to 896K tokens, PARSER with a 4B backbone outperforms the strongest sequential memory baseline by 5.7 points on average and by 12.0 points at 896K tokens. Scaling to a 9B backbone, PARSER surpasses DeepSeek-V4-Pro by 6.3 points. Controlled experiments confirm that PARSER is robust to perturbations in evidence position, order, and distance, conditions that cause large accuracy swings in sequential methods, while reducing inference latency by up to 11x.
The Brief
2:26A short podcast summary of the paper. Written and voiced by AI.
Transcript
PARSER. Read in Parallel, Reason in Depth for Long-Context LLM Agents. By Kun Li and four others, from the Chinese University of Hong Kong. Published on the sixth of September. Its code repository is worked out: about one thousand five hundred source files, with a readme and a licence.
Sequential memory agents handle a long document by reading it chunk after chunk while carrying a compact memory state along. That ties traversal of the document to depth of reasoning, and the coupling has two costs. The agent becomes sensitive to where in the document the evidence happens to sit, and its inference time grows linearly with the length of the document. A reader that must go page by page cannot reason about the final page until it has been through every page before it, and whatever it chose to keep in memory along the way limits what it can connect later.
PARSER decouples reading from reasoning. The motto is in the title: read in parallel, reason in depth. The whole document is read at once by many readers, while a single agent does the thinking. How deeply the system reasons no longer depends on how long the document is.
A bank of lightweight subagents is each bound to a single chunk, so between them they read the entire document in parallel. A lead agent reasons through iterative scatter-gather rounds. In each round it broadcasts a query to every subagent, gathers the evidence they send back, and then writes a deeper follow-up query conditioned on everything found so far.
That division puts all the learnable behaviour in the lead agent, which is optimized with reinforcement learning. The subagents stay frozen, ordinary off-the-shelf models.
The testbed is multi-hop question answering, with contexts from seven thousand tokens up to eight hundred and ninety-six thousand. With a four billion parameter backbone, PARSER beats the strongest sequential memory baseline by five point seven points on average. At the longest context, the margin is twelve points. Scaled to a nine billion parameter backbone, PARSER surpasses DeepSeek-V4-Pro by six point three points.
Controlled experiments test robustness directly. PARSER holds up under perturbations to where the evidence is placed, in what order it appears and how far apart its pieces sit, conditions that cause large swings in accuracy for sequential methods. It also cuts inference latency by up to eleven times.
The two problems it set out to solve came from one coupling, and breaking that coupling answers both. Reading in parallel removes the dependence on document length and on where evidence sits. Keeping reasoning in one trained agent lets depth come from the number of rounds rather than from the order of pages.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a document too long to read in order without losing the thread.
You need a model to answer questions whose evidence is scattered across a document of hundreds of thousands of tokens. Even models with huge windows lose accuracy as the input grows.
The road ahead
- 1 · The problem ~6 min Long documents defeat both big context windows and chunk-by-chunk memory readers, because evidence gets diluted, dropped or stuck behind every earlier page.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Split reading from reasoning: many readers each scan one chunk at the same time, while one lead agent asks them questions, round after round.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Chunk the document, then loop: broadcast a query, read all chunks at once, gather evidence, write a deeper query, until the lead agent can answer.
- 6 · The experiment ~6 min You’ll guess: Reading in parallel instead of chunk by chunk, how much did PARSER cut inference latency at most?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.