Long-context language model agents typically process lengthy documents sequentially, reading one chunk at a time while maintaining a compact memory state. According to a new paper posted on Hugging Face, this coupling of document traversal to reasoning depth makes such agents sensitive to where evidence is placed and ties inference latency linearly to document length.
The proposed method, PARSER, decouples reading from reasoning. A bank of lightweight subagents, each bound to a single chunk, reads the entire document in parallel. Meanwhile, a lead agent reasons in depth through iterative scatter-gather rounds: in 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.
All learnable behavior is concentrated in the lead agent, which is optimized with reinforcement learning, while the subagents remain frozen off-the-shelf models. This design keeps the system modular and avoids retraining the reading components.
On multi-hop question answering 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 paper reports that accuracy stays nearly lossless from 7K to 896K tokens and that inference is substantially faster than memory-based approaches.