PARSER's parallel subagents let a 9B beat DeepSeek-V4-Pro

A 9-billion-parameter open model beats DeepSeek-V4-Pro by 6.3 points on multi-hop QA. The design that gets it there, described in a new arXiv paper, is a bank of lightweight subagents that each read one chunk of a document in parallel while a single lead agent reasons across the returned evidence.

The authors call the architecture PARSER, short for 'Read in Parallel, Reason in Depth,' and pitch it against sequential memory agents that scan a long document chunk-by-chunk. Sequential memory agents, they write, 'process long documents by reading chunks one after another while maintaining a compact memory state, coupling document traversal to reasoning depth.' The consequence, in their words: it 'ties inference latency linearly to document length.'

PARSER breaks the coupling by fanning out. Frozen off-the-shelf subagents each hold one chunk; the lead agent, the only part trained with reinforcement learning, '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.' On contexts stretching from 7K to 896K tokens, a 4B version beats the strongest sequential baseline by 5.7 points on average and by 12.0 at 896K tokens. Scale the backbone to 9B and the paper reports it 'surpasses DeepSeek-V4-Pro by 6.3 points.'

Latency drops up to 11x, and controlled tests find the system 'robust to perturbations in evidence position, order, and distance, conditions that cause large accuracy swings in sequential methods.'

Unnamed in the abstract: the sequential baseline being outperformed, the specific QA benchmarks used, and the compute cost of running many subagents in parallel against one long call.