PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents
Kun Li, Zexuan Qiu and Tianhua Zhang with Irwin King and Helen Meng at the Chinese University of Hong Kong present PARSER, which gives every document chunk its own frozen subagent and puts all the learned behavior in a lead agent that queries them in scatter-gather rounds.
Ask this paper
The coupling being broken: Sequential memory agents read chunks one after another while maintaining a compact state, which ties reasoning depth to document traversal and makes both accuracy sensitive to evidence placement and latency linear in document length.
Parallel reading, iterative reasoning: A bank of lightweight subagents each bound to one chunk read the whole document at once, while the lead agent broadcasts a query, aggregates returned evidence, and forms a deeper follow-up query conditioned on what it has found.
Only the lead agent is trained: Learnable behavior is concentrated in the lead agent, optimized with reinforcement learning; the subagents stay frozen off-the-shelf models, which keeps the trainable surface small.
Gains grow with context length: 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 by 12.0 points at 896K. At 9B it surpasses DeepSeek-V4-Pro by 6.3 points.
Robust where sequential methods swing: Controlled experiments show stability under perturbations to evidence position, order and distance, conditions that cause large accuracy swings in sequential methods, with inference latency reduced by up to 11x.
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.