PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Hanwen Liu, Yuanfu Sun, Qiaoyu Tan

arXiv:2609.39154v1cs.CLcs.AI

TL;DR

Deep research agents must synthesize evidence across many sources, but existing DAG-based systems commit to a task-level plan before execution. DAGent grows its DAG from completed-node confidence and uncertainty signals and uses its recorded topology in DAGRPO. Across three benchmarks, it reports gains over strong baselines, including 5.3 / 5.8 / 2.0 points at Qwen3-235B-A22B.

  • Problem

    Existing DAG-based agents commit to a task-level plan before execution and repair the graph only after failures or missing evidence emerge.

    Source: §1 · p. 2

  • Method

    DAGent incrementally grows its DAG using completed-node feedback, and DAGRPO adds topology-conditioned credit and structural-compliance regularization.

    Source: §1 · p. 2

  • Results

    Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent leads the strongest baseline by 5.3 / 5.8 / 2.0 points at Qwen3-235B-A22B.

    Source: §4.2 · p. 7

  • Takeaways & Limitations

    DAGent's advantage spans three benchmarks and multiple backbone settings, including four open-source backbones and GPT-5 at 327K context.

    Source: §1 · p. 2

    The paper does not establish whether DAGRPO gains persist at larger training scales, longer schedules, or full fine-tuning; its RL experiments use Qwen3-8B with LoRA for 21 update steps over three seeds.

    Source: F Limitations · p. 26

Abstract

from arXiv · show

Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent

1 Introduction

Deep research can make intermediate findings redirect investigation, while Plan-then-Patch agents commit to a full plan before execution. DAGent instead grows a DAG incrementally, uses its recorded topology for structural RL signals, and reports broad benchmark advantages.

Motivation

  • Plan-then-Patch commits to a task-covering plan before any node executes, then patches it as findings emerge, a poor fit when evidence can redirect research.The authors associate this pattern with degraded accuracy and wasted compute as task complexity grows.Source: §1 · p. 2

Approach

  • DAGent grows its task DAG one batch at a time, using completed-node feedback to decide what to plan next, while preserving compact summaries and full traces for downstream recall.New nodes are executed in parallel by ReAct Executors.Source: §1 · p. 2
  • DAGRPO adds topology-conditioned credit for Executor rollouts and structural-compliance regularization for Orchestrator plans, using the recorded DAG structure.The append-only graph makes a node's execution context match its ancestry in the final graph.Source: §1 · p. 2

Empirical advantage

  • DAGent consistently outperforms strong training-free baselines across three benchmarks and backbone scales, with results replicating across four open-source backbones and extending to GPT-5 at 327K context.The evaluation covers BrowseComp-Plus, GAIA, and xbench-DeepSearch.Source: §1 · pp. 2–3
  • At the same compute budget, DAGRPO further improves Qwen3-8B DAGent over its role-separated GRPO counterpart.The paper also reports that DAGRPO's structural signals add gains over an outcome-only GRPO baseline.Source: §1 · p. 3

2 Related Work

Prior deep-research agents either follow linear context strategies or commit to a structured plan before execution; DAGent instead expands a DAG using feedback from completed nodes. Related RL approaches add structure at several granularities, while DAGRPO uses topology-conditioned credit that relies on an append-only graph.

2.1 Deep Research Agents

  • DAGent differs from prior deep-research agents by growing its DAG batch by batch under confidence and uncertainty signals from completed nodes.The related work contrasts this with linear-trajectory methods and structured plans committed before execution.Source: §2.1 · p. 3

2.2 Structure-Aware RL-Based Optimization for LLM Agents

  • DAGRPO uses answer-feeding-chain credit that is uniquely determined by the recorded trajectory only when the graph is append-only.Plan-then-Patch can delete or rewire nodes, so a node's final ancestry may differ from what synthesis actually read.Source: §2.2 · p. 3

3 Method

DAGent grows research plans incrementally from completed-node evidence, then uses the recorded DAG to manage context propagation and shape role-specific reinforcement-learning signals.

3.1 Evaluate-then-Grow Incremental Planning

  • At each iteration, the Orchestrator evaluates completed nodes, refines uncertain or missing evidence (up to three attempts), and adds dependency-linked batches, parallelizing independent nodes and isolating final synthesis in an answer batch.Source: §3.1 · pp. 4–5
  • Nodes receive direct dependencies’ compact QueryDocs by default, while RECALLTOOL can retrieve targeted evidence from a dependency’s full transcript when needed.Source: §3.1 · p. 5

3.2 DAGRPO: DAG-Conditioned Reinforcement Learning

  • DAGRPO uses recorded DAG topology to give full outcome reward to Executor rollouts feeding final synthesis and attenuate off-chain rollouts; this closure is a structural proxy, not causal attribution.Source: §3.2 · pp. 5–6
  • Orchestrator plans incur a structural compliance penalty for violations including missing node statuses, omitted refinements, invalid dependencies, and a non-singleton answer batch.Source: §3.2 · p. 6

4 Experiments

Across three deep-research benchmarks, DAGent leads training-free baselines across Qwen3 scales and replicates across open-source backbones and GPT-5. Ablations favor evidence-conditioned incremental planning, while same-architecture comparisons report higher accuracy with smaller task footprints than Plan-then-Patch.

4.1 Experimental Setup

  • Evaluation spans three benchmarks with shared retrieval backends, five agent families, and Qwen3-8B, Qwen3-32B, and Qwen3-235B-A22B backbones.FlowSearch is included at Qwen3-32B; Qwen3-8B training-based results compare DAGRPO with same-budget outcome-only GRPO.Source: §4.1 · p. 7

4.2 Main Results (RQ1, RQ2)

  • At Qwen3-235B-A22B, DAGent leads the strongest non-DAGent training-free baseline by 5.3 / 5.8 / 2.0 points across BrowseComp-Plus / GAIA / xbench-DeepSearch.At Qwen3-32B, it reaches 47.3 / 55.3 / 65.0, 4.0 / 3.8 / 2.0 points above FlowSearch on those benchmarks.Source: §4.2 · p. 7, Table 1 · p. 8
  • DAGRPO improves Qwen3-8B DAGent over a same-budget outcome-only GRPO baseline by 3.0 average points.Across BrowseComp-Plus / GAIA / xbench-DeepSearch, DAGRPO scores 49.6 / 53.4 / 65.7, exceeding that baseline by 3.6 / 2.6 / 2.7 points.Source: §4.2 · pp. 7–8
  • DAGent's lead replicates across four mid-size open-source backbones and at GPT-5 327K, where margins are 5.7 / 2.5 / 2.0 points across the three benchmarks.The GPT-5 comparison uses GAIA's full 165-task validation set with multimodal tools.Source: Figure 3 · p. 7, §4.2 · p. 8, Figure 4 · p. 8

4.3 Ablation Study (RQ3)

  • Within the same architecture, Plan-then-Patch costs 5.3 / 4.8 / 4.0 Pass@1 points versus Evaluate-then-Grow across the three benchmarks.A single upfront plan costs another 8.7 / 7.8 / 6.0 points; context-component drops rank Selective Propagation, QueryDoc, then InteractionTranscript.Source: §4.3 · p. 9, Table 2 · p. 9
  • Both DAGRPO structural signals have positive Pass@1 differences, and the off-chain credit coefficient α peaks at 0.5 on every benchmark.The ablation rows use one seed, while DAGRPO is a three-seed mean; the smaller per-benchmark differences are within seed noise.Source: §4.3 · p. 9, Table 3 · p. 9

4.4 Efficiency Analysis (RQ4)

  • Plan-then-Patch uses 40 / 29 / 18% more tokens and 28 / 21 / 21% more execution steps, yet trails DAGent by 5.3 / 4.8 / 4.0 Pass@1 points.For Qwen3-32B training-free DAGent, mean per-task footprints are 42.7 / 31.2 / 26.8 steps, 67.6 / 36.8 / 32.9 tool calls, and 1.20M / 0.66M / 0.44M total input-plus-output tokens.Source: §4.4 · p. 10, Figure 5 · p. 10

5 Conclusion

DAGent replaces upfront task decomposition and reactive graph repair with incremental Evaluate-then-Grow planning, and DAGRPO uses the resulting topology for structural reinforcement-learning signals. The appendix explains how this planning differs from Plan-then-Patch and how DAGRPO’s credit and compliance signals are designed.

5 Conclusion

  • DAGent consistently outperforms strong open-source baselines across backbone scales and vendors, leads released FlowSearch in a controlled comparison, and uses lower per-task cost than Plan-then-Patch counterparts.DAGRPO adds gains over a same-budget outcome-only GRPO baseline.Source: §5 · p. 10

A DAGent Algorithm

  • The workflow iteratively evaluates the current graph, adds new nodes and dependency edges, executes new nodes, and returns the answer field from the answer node.The appendix’s algorithm listing identifies the task, Orchestrator, Executor, and iteration limit as inputs.Source: §B.1 · p. 17

B.1 Evaluate-then-Grow versus Plan-then-Patch

  • Evaluate-then-Grow creates each node only after its dependencies execute and are evaluated; its append-only graph records the workflow that actually ran.Unlike Plan-then-Patch, it does not create a complete task plan before execution or delete and rewire existing nodes.Source: §B.1 · p. 17

B.2 DAGRPO

  • DAGRPO attenuates off-chain Executor credit by α; α = 0 removes that credit and falls below outcome-only GRPO, while α = 0.5 performs best on every benchmark.Off-chain rollouts may contain useful evidence not integrated into the final synthesis.Source: §B.2 · p. 17
  • The answer-inclusive closure S(τ) is a reachability-based proxy for contribution, not causal attribution, and is read from the recorded graph without re-execution.The appendix describes safeguards against dependency inflation, including the accuracy cost of adding irrelevant QueryDocs to downstream contexts.Source: §B.2 · p. 17
  • DAGRPO’s structural signals assign topology-conditioned credit to Executor rollouts and impose structural compliance regularization on Orchestrator plans.The recorded graph supports these signals because it preserves the dependencies and execution context of nodes.Source: §B.2 · pp. 17–18
  • DAGRPO’s trajectory-level structural penalty regulates evidence flow: it requires node statuses, re-examination of weak nodes, evidence access for final synthesis, and dependencies on already evaluated nodes.The penalty scale keeps a correct plan with up to three violating turns above a clean failure; the cap bounds the penalty for heavily violating plans.Source: §B.2 · p. 18

B.3 Relation to Concurrent Credit-Assignment Methods

DAGRPO assigns credit using whether a rollout feeds the answer node, rather than only its agent role. The recorded, append-only task graph lets this reachability signal be read without rerunning downstream work.

B.3 Relation to Concurrent Credit-Assignment Methods

  • Unlike Graph-GRPO’s topology optimization over fixed agents, DAGent creates task-decomposition nodes during execution; unlike M-GRPO’s role-based credit, DAGRPO also considers whether a rollout feeds the answer.Counterfactual removal as in CCPO would require re-executing descendants that already consumed the removed node’s QueryDoc, whereas DAGRPO reads reachability from the recorded graph.Source: §B.3 · p. 18
  • Append-only growth keeps a node’s final-graph ancestry aligned with the context it saw when it ran, unlike systems that can later delete or rewire executed nodes.This makes the topology-based signal well defined without re-execution.Source: §B.3 · p. 18

C.1 Benchmarks, Tool Stack, and Judge

The appendix documents benchmark tooling, framework and baseline settings, judge calibration, and supplementary training-free comparisons. It also describes a multimodal extension for full-set GAIA evaluation and details the Plan-then-Patch comparison protocol.

C.1 Benchmarks, Tool Stack, and Judge

  • Within each benchmark, methods share the same retrieval backend; BrowseComp-Plus uses local dense retrieval, while the other two benchmarks use live Google search.Pass@1 uses GPT-4o-mini, with GPT-4.1 as a tie-breaker on borderline cases.Source: §C.1 · p. 18, Table 4 · p. 18

C.2 DAGent Framework Parameters

  • Each Orchestrator turn and Executor sub-task has an 8,192-token prompt cap and a 32,768-token response cap, within the per-context budget shared by multi-context baselines.A separate 5,400-second session timeout bounds whole tasks; none of the evaluated tasks reaches it.Source: §C.2 · p. 18, Table 5 · p. 19

C.3 Baseline Configurations and Caps

  • FlowSearch uses its released question-answering preset with the evaluation backbone, enabled Coordinator, and shared search tools; its official budgets are retained.The Plan-then-Patch variant reuses DAGent’s components and per-node refine budget, changing only the Orchestrator’s cross-iteration planning behavior.Source: §C.3 · p. 19, §D.2 · p. 21

C.4 RL Training Hyperparameters

  • RL training uses LoRA on all linear layers with asynchronous vLLM rollouts, while unlisted settings inherit VeRL defaults.The appendix lists the RL training and DAGRPO hyperparameters in Table 7.Source: §C.4 · p. 19, Table 7 · p. 20

C.5 Human Agreement Study on the LLM Judge

  • The human-majority comparison yields aggregate Cohen’s κ = 0.95 and 97.3% raw agreement with the LLM judge; agreement is slightly lower on Chinese-language xbench-DeepSearch.Three annotators judged 150 answer tuples sampled across the three benchmarks.Source: §C.5 · p. 20, Table 8 · p. 20

D Additional Training-Free Results

  • The appendix collects supplementary training-free results for the paper’s cross-backbone, frontier-scale, and efficiency analyses.It supplements results discussed in §§4.2–4.4.Source: D Additional Training-Free Results · p. 20

D.1 Cross-Backbone and Frontier-Scale Details

  • For full-validation GAIA at GPT-5 327K, DAGent adds image, document, and audio inspectors to textualize attachments while leaving its other components unchanged.The extension covers 62 multimodal questions in the 165-task validation set.Source: §D.1 · pp. 20–21

D.2 Plan-then-Patch Variant Details

  • The Plan-then-Patch protocol commits the complete DAG, including its answer node, before execution, then executes ready batches and permits capped refine or add operations.Patch turns are capped at 30, with mean counts of 4.2 / 1.8 / 1.4 on BrowseComp-Plus / GAIA / xbench-DeepSearch; patch nodes are connected to the answer node automatically.Source: §D.2 · pp. 21–22

D.3 Extended Efficiency Comparison

In a same-backbone, same-retrieval comparison, DAGent is the most accurate and cheapest among systems with separate per-node contexts. Its Plan-then-Patch counterpart costs more on every reported per-task measure, while lower-cost Flash-Searcher is less accurate.

D.3 Extended Efficiency Comparison

  • DAGent is the most accurate and cheapest on every per-task cost column among systems with separate per-node contexts; Flash-Searcher costs less but scores 8.6 / 10.6 / 7.0 points lower.No workflow in Table 10 matches DAGent’s accuracy at any cost level.Source: §D.3 · pp. 22–23, Table 10 · p. 23
  • The Plan-then-Patch variant makes 35 / 27 / 21% more external tool calls and takes 35 / 27 / 23% more wall-clock time than DAGent.Its external calls are 82.3 / 42.0 / 35.4 against DAGent’s 60.8 / 33.1 / 29.2; times are 731.9 / 584.6 / 466.8 s against 541.6 / 461.8 / 380.9 s.Source: §D.3 · p. 22, Table 10 · p. 23
  • FlowSearch makes 26 / 6 / 8% more external tool calls and takes 32 / 23 / 28% more time than DAGent, while scoring 4.0 / 3.8 / 2.0 points lower.The comparison uses the same Qwen3-32B backbone and retrieval backends; reported calls include model-driven workflow calls, while external calls count search and open_page only.Source: §D.3 · p. 22
  • Across the three benchmarks, DAGent averages 55.9 Pass@1, compared with 52.6 for FlowSearch, 51.2 for Plan-then-Patch, and 47.1 for Flash-Searcher.The Plan-then-Patch variant retains DAGent’s component stack and falls between the two released Plan-then-Patch systems in average accuracy.Source: §D.3 · p. 23

D.4 Behavioral Statistics of the Evaluate-then-Grow Loop

Orchestrator evaluations make uncertain and negative evidence a routine input to Evaluate-then-Grow planning, while recall remains a fallback. The reported statistics also connect DAGRPO’s structural compliance regularization with accuracy and format-warning behavior.

D.4 Behavioral Statistics of the Evaluate-then-Grow Loop

  • Uncertain and Not Found account for 54.3% of executed BrowseComp-Plus nodes and about a third on GAIA and xbench-DeepSearch; 22.0% / 13.6% / 11.0% of tasks use recall.Figure 6 summarizes evaluation outcomes over all executed nodes and first-batch nodes, alongside recall usage.Source: §D.4 · p. 23, Figure 6 · p. 23
  • RECALLTOOL averages 0.33 / 0.21 / 0.18 calls per task on BrowseComp-Plus / GAIA / xbench-DeepSearch; removing InteractionTranscript costs 4.0 / 2.9 / 2.0 Pass@1 points.Given the usage rates, the paper estimates the tool changes outcomes for roughly one in five tasks that call it.Source: Figure 6 · p. 23, §D.4 · p. 24
  • First-batch nodes are Success / Uncertain / Not Found in 61.9% / 34.9% / 3.2% of cases, and refinements turn 44.4% of weak nodes into Success.Across all executed nodes, Uncertain / Not Found rates are 41.6% / 12.7% on BrowseComp-Plus, 26.4% / 5.2% on GAIA, and 28.8% / 4.8% on xbench-DeepSearch.Source: §D.4 · p. 24
  • Removing DAGRPO’s structural compliance regularization costs 1.6 / 1.0 / 0.7 Pass@1 points, while tasks with structural violations score 14.7 points lower than tasks without them.The authors caution that task difficulty may contribute to the score association, so it is not a causal estimate.Source: §D.4 · p. 24

E.1 Per-Seed Training Results

Across three training seeds, DAGent’s RL gains are stable, and no seed reverses the DAGRPO-versus-GRPO ranking on any benchmark.

E.1 Per-Seed Training Results

  • Each of the six Qwen3-8B RL variants was run with three independent training seeds: 42, 123, and 777.All other hyperparameters were held fixed; each seed affected both dataloader shuffle and LoRA initialization.Source: §E.1 · p. 24
  • Standard deviations for DAGent’s two RL rows remain below 1.6 absolute Pass@1 points across all three benchmarks, with no seed reversing the DAGRPO-versus-GRPO ranking.Figure 7 reports per-seed Pass@1 for the six Qwen3-8B RL variants on BrowseComp-Plus, GAIA, and xbench-DeepSearch.Source: §E.1 · p. 24, Figure 7 · p. 24
  • Figure 7 breaks out Pass@1 (%) by seed for BrowseComp-Plus, GAIA, and xbench-DeepSearch.The three seeds are 42 / 123 / 777.Source: Figure 7 · p. 24

E.2 DAGRPO Training Dynamics

DAGRPO’s topology-conditioned credit and structural compliance regularization exhibit their intended training-time behaviors in the Qwen3-8B LoRA runs.

E.2 DAGRPO Training Dynamics

  • Figure 8 reports training-time trajectories; its sampled-rollout reward gap, warning rate, and on-chain ratio differ in measurement setting from evaluation-time Table 11 values.The figure’s four panels show validation Pass@1, reward gap, format-warning rate, and on-chain ratio.Source: §E.2 · p. 25, Figure 8 · p. 25
  • At α = 0.5, topology credit separates on-chain and off-chain Executor rewards by construction, whereas the α = 1.0 baseline has zero gap.The reward-gap curves are computed on sampled training rollouts at each step.Source: §E.2 · p. 25, Figure 8 · p. 25
  • Structural compliance regularization lowers the Orchestrator format-warning rate faster than outcome-only GRPO during training.Figure 8 also tracks BrowseComp-Plus validation Pass@1 and the fraction of Executor nodes inside the answer-inclusive closure.Source: §E.2 · p. 25, Figure 8 · p. 25

E.3 Efficiency as a Byproduct of RL Training

RL training reduces DAGent’s per-task footprint, with DAGRPO showing further reductions than role-separated outcome-only GRPO despite no explicit efficiency design.

E.3 Efficiency as a Byproduct of RL Training

  • Neither topology-conditioned credit nor structural compliance regularization targets per-task footprint; the reductions are observed as a byproduct of accuracy-driven RL training.The reported reductions are averaged over three independent training seeds.Source: §E.3 · p. 25
  • DAGRPO reduces steps by 12 to 17%, tool calls by 10 to 14%, and tokens by 17 to 21% across BrowseComp-Plus / GAIA / xbench-DeepSearch.These reductions are larger than the role-separated outcome-only GRPO reductions of 8 to 12%, 6 to 9%, and 15 to 16%, respectively.Source: §E.3 · p. 25
  • DAGRPO reduces graph node count by 1.3 to 1.9 nodes per task and off-chain Executor ratio by 0.05 to 0.09; GRPO reductions are 0.7 to 0.9 nodes and 0.02 to 0.04 ratio.The additional DAGRPO reductions span steps, tool calls, and tokens, and are largest for the off-chain ratio.Source: §E.3 · p. 25

F Limitations

This section describes limitations tied to DAGent’s text-native design, retrieval quality, benchmark and RL scope, and documents its orchestration, execution, and transcript-recall prompts. It also outlines benchmark-specific LLM judging procedures, including GPT-4o-mini judging and GPT-4.1 tie-breaking for borderline cases.

F Limitations

  • DAGent converts attachments to text and cannot recover evidence its retriever misses, making results dependent on search, embedding, and extraction quality.Its answer-inclusive closure is a structural proxy rather than causal attribution, and the reported compliance–accuracy association is correlational. The benchmarks test closed-form answers, while RL uses Qwen3-8B with LoRA for 21 steps over three seeds; results at larger scales, longer schedules, or with full fine-tuning are unestablished.Source: F Limitations · pp. 25–26

G.1 Orchestrator System Prompt

  • The Orchestrator prompt requires evaluating completed nodes, planning executable dependency-safe batches, refining uncertain or missing results, and placing the answer node alone after candidate checks.It specifies distinct search strategies for refinement and requires a concrete, best-supported candidate before answering.Source: §G.1 · pp. 26–27

G.2 ReAct Executor System Prompt

  • The Executor prompt structures research as an iterative OODA loop, with source discovery, full-page review, checklist-based verification, and evidence-supported reporting.It also directs the agent to triangulate critical claims across independent sources and continue targeted research when checklist evidence is weak.Source: §G.2 · p. 29

G.3 Recall Extraction Prompt

  • The Recall prompt restricts extraction to the prior transcript and requires factual claims to carry transcript source identifiers.It calls for a concise summary and relevant supporting transcript excerpts.Source: §G.3 · p. 30

G.4 LLM Judge Prompts

  • BrowseComp-Plus follows its benchmark grader, GAIA uses an equivalence prompt instead of its original quasi-exact-match scoring, and xbench-DeepSearch uses its Chinese-language grader.All three are judged by GPT-4o-mini, with GPT-4.1 used as a tie-breaker on borderline cases.Source: §G.4 · p. 31

Related papers

Loading 2609.39154v1…