Argus: Evidence Assembly Fixes Duplicate-Evidence Problem in Parallel Deep Research
arXiv·medium signal
Deep research agents that scale via parallel rollouts often duplicate evidence rather than finding complementary pieces, yielding diminishing returns while pushing context toward the model's limit. Argus proposes an evidence assembly framework that deduplicates and completes evidence across parallel trajectories. The approach addresses a key bottleneck: parallel rollouts are compute-expensive but produce redundant answers, and naive aggregation wastes the context budget on duplicates.