DAGent Grows Research Plans From Evidence Confidence

DAGent grows deep-research agent plans from evidence confidence and beats open baselines on three benchmarks.

DAGent is a DAG-based multi-agent framework for deep research that expands its plan one batch at a time, conditioning each growth step on confidence and uncertainty from completed nodes rather than locking in a full plan up front. A hierarchical context layer shares compact QueryDocs while keeping full traces for recall, and DAGRPO assigns topology-conditioned credit during reinforcement learning of executor rollouts. At the Qwen3-235B-A22B scale, both reports say it leads the strongest open-source baseline by 5.3, 5.8, and 2.0 points on BrowseComp-Plus, GAIA, and xbench-DeepSearch. That lead is reported across four open backbones and GPT-5 at 327K context. At Qwen3-8B, DAGRPO adds 3.0 average Pass@1 over same-budget outcome-only GRPO with lower token, tool-call, and step cost. The Hugging Face and arXiv write-ups agree on these figures and do not state conflicting results.

• DAGent is a DAG multi-agent framework that grows a deep-research plan one batch at a time from node confidence and uncertainty instead of committing to a full plan first.
• A hierarchical context layer shares compact QueryDocs while retaining full traces, and DAGRPO adds topology-conditioned reinforcement-learning credit to executor rollouts.
• At Qwen3-235B-A22B, DAGent beats the strongest open-source baseline by 5.3, 5.8, and 2.0 points on BrowseComp-Plus, GAIA, and xbench-DeepSearch.
• The lead holds across four open backbones and GPT-5 at 327K context.
• At Qwen3-8B, DAGRPO adds 3.0 average Pass@1 over same-budget outcome-only GRPO while using fewer tokens, tool calls, and steps.
• Hugging Face daily papers dated the item 2026-09-29; the arXiv cs.AI/cs.LG/cs.CL listing is dated 2026-09-30, with matching results.