Research map / Multi-agent systems and games / LLM agents

DAGent: Evaluate-then-Grow Planning for Deep Research Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DAGent, a DAG-based deep research agent that grows its task graph incrementally with evidence-conditioned evaluation (Evaluate-then-Grow), plus DAGRPO, a GRPO variant with topology-conditioned credit and structural compliance regularization.

Problem

Deep research agents with DAG-based decomposition commit to a full plan before execution and only patch it after failures (Plan-then-Patch), which is brittle when evidence is weakest and wastes compute; RL for such agents also ignores DAG structure.

Approach

Grow the task DAG one batch at a time, conditioning each expansion on per-node QueryDoc status (Success/Uncertain/Not Found), with hierarchical context (compact QueryDocs by default, full transcripts recallable), and train with DAGRPO which assigns executor credit by whether rollouts feed the answer node and penalizes structurally invalid orchestrator plans.

Key results

• At Qwen3-235B-A22B, DAGent beats the strongest open-source baseline by 5.3 / 5.8 / 2.0 points on BrowseComp-Plus / GAIA / xbench-DeepSearch.
• DAGRPO improves Qwen3-8B DAGent over the same-budget GRPO baseline by 3.0 average Pass@1 points (49.6 vs 46.6 averaged across three seeds per the reported means).
• Replacing Evaluate-then-Grow with Plan-then-Patch costs 5.3 / 4.8 / 4.0 points on the three benchmarks at Qwen3-32B.

Surprising findings

• Removing all off-chain executor credit (alpha=0) drops performance below the outcome-only GRPO baseline, showing off-chain rollouts carry learning value.

Limitations

• Closure-based credit is a structural proxy for contribution, not a causal attribution.
• Many ablation differences are within seed noise since most ablation rows use a single seed.
• FlowSearch evaluated only at Qwen3-32B and GPT-5 comparison relies on numbers reproduced from other papers.

Experimental setup

Datasets: BrowseComp-Plus

Benchmarks: BrowseComp-Plus GAIA xbench-DeepSearch

Metrics: Pass@1

Baselines: ReAct Agent Summary Agent Fold Agent Flash-Searcher FlowSearch GRPO-DAGent

Builds on: GRPO DAPO ReAct FlowSearch Flash-Searcher

Closest papers by meaning

• From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents (NeurIPS 2026 · similarity 0.88)
• MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (NeurIPS 2026 · similarity 0.87)
• ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence (NeurIPS 2026 · similarity 0.86)
• WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research (ICLR 2026 · similarity 0.86)
• FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.85)
• InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.85)