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

Hanwen Liu; Yuanfu Sun; Qiaoyu Tan

cs.CLSep 30, 2026

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

Authors: Hanwen Liu, Yuanfu Sun, Qiaoyu Tan

Organizations: New York University · New York University Shanghai

Abstract

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

Figures & tables

Figure 1 : Comparison of Plan-then-Patch (top) and Evaluate-then-Grow (bottom) paradigms.

Figure 2 : Overview of DAGent. The top panel shows Evaluate-then-Grow planning, where the Orchestrator evaluates each executed layer and expands the DAG accordingly. Nodes marked as Uncertain or Not Found trigger refine nodes for further verification, while child nodes receive the required context from their parents until an answer node is produced. The bottom-left panel shows four representative planning cases (a–d), marked at the corresponding nodes of the top panel: a node that succeeded and was expanded, a node that was uncertain and refined, a successfully refined node, and a node stopped after its refine budget was exhausted. The bottom-right panel summarizes DAG-structured policy optimization, where Orchestrator and Executor rewards are computed separately and optimized jointly, as detailed in § 3.2 .

Backbone	Max #Token	Agent Paradigm	BrowseComp-Plus				GAIA				xbench-DS
Backbone	Max #Token	Agent Paradigm	Easy	Med.	Hard	Avg.	L1	L2	L3	Avg.	Avg.
Training-free
Qwen3-8B	32K	ReAct Agent	46.0	14.0	0.0	20.0	38.5	23.1	25.0	29.1	39.0
	109K	ReAct Agent	64.0	18.0	2.0	28.0	46.2	26.9	25.0	34.0	44.0
	32K $\times$ N	Summary Agent	76.0	20.0	2.0	32.7	51.3	26.9	16.7	35.0	54.0
	32K $\times$ N	Fold Agent	78.0	24.0	2.0	34.7	48.7	28.8	16.7	35.0	52.0

Table 1 : Main results on BrowseComp-Plus, GAIA, and xbench-DeepSearch: Pass@1 (%), best result per backbone in bold . Training-free rows use a single greedy inference pass; training-based rows report mean

\pm

std (sample standard deviation) over three training seeds on the benchmark averages. FlowSearch is evaluated with the authors’ released implementation at Qwen3-32B only (§ 4.1 ).

Figure 3 : Cross-backbone DAGent Pass@1 (%); per-baseline breakdown in Appendix D.1 .

Figure 4 : Performance comparison of DAGent and state-of-the-art agent systems on BrowseComp-Plus, GAIA, and xbench-DeepSearch. Pass@1 (%) is reported.

Method	BrowseComp-Plus		GAIA		xbench-DS		Overall
Method	Avg.	$\Delta$	Avg.	$\Delta$	Avg.	$\Delta$	Avg.	$\Delta$
DAGent (Full)	47.3	—	55.3	—	65.0	—	55.9	—
Plan-then-Patch Variant	42.0	-5.3	50.5	-4.8	61.0	-4.0	51.2	-4.7
w/o Evaluate-then-Grow Planning	33.3	-14.0	42.7	-12.6	55.0	-10.0	43.7	-12.2
w/o Selective Propagation	34.7	-12.6	44.7	-10.6	57.0	-8.0	45.5	-10.4
w/o QueryDoc	39.3	-8.0	47.6	-7.7	60.0	-5.0	49.0	-6.9

Table 2 : Component ablation with Qwen3-32B in training-free mode: Pass@1 (%) per benchmark, Overall is the unweighted mean, and

\Delta

is the change from DAGent (Full). All rows use a single greedy inference pass, as in the training-free rows of Table 1 .

$\alpha$	Regularization	Configuration	BrowseComp-Plus		GAIA		xbench-DS		Overall
$\alpha$	Regularization	Configuration	Avg.	$\Delta$	Avg.	$\Delta$	Avg.	$\Delta$	Avg.	$\Delta$
0.00	on	—	44.0	-5.6	49.5	-3.9	61.0	-4.7	51.5	-4.7
0.25	on	—	48.7	-0.9	52.4	-1.0	64.0	-1.7	55.0	-1.2
0.50	on	DAGRPO	49.6	—	53.4	—	65.7	—	56.2	—
0.75	on	—	48.0	-1.6	52.4	-1.0	64.0	-1.7	54.8	-1.4
1.00	on	w/o topology credit	47.3	-2.3	51.5	-1.9	64.0	-1.7	54.3	-2.0

Table 3 : DAGRPO component ablation and

\alpha

sensitivity on Qwen3-8B: Pass@1 (%) per benchmark, Overall is the unweighted mean, and

\Delta

is the change from DAGRPO; the Overall

\Delta

is the mean of the three per-benchmark

\Delta

values and can differ by 0.1 from the difference of the rounded Overall values. The DAGRPO and GRPO-baseline rows are the three-seed means of Table 1 ; all other rows are single-seed runs with the same training budget.

Figure 5 : DAGent per-task efficiency on Qwen3-32B training-free. (a) Tool calls vs. execution steps with 2.2

\sigma

confidence ellipses, and per-task tool calls per step. (b) DAGent (Full) vs. the Plan-then-Patch variant on per-task footprint metrics.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

	BrowseComp-Plus	GAIA	xbench-DeepSearch
Tasks	150	103 (text-only validation)	100 (2505 release)
Difficulty splits	50 easy / 50 medium / 50 hard	39 L1 / 52 L2 / 12 L3	none
Language	English	English	Chinese
Search	Local dense retrieval	Google search via Serper	Google search via Serper
Page access	Verified corpus	Jina extraction	Jina extraction
Judge prompt	Benchmark rubric	Equivalence prompt	Benchmark rubric (Chinese)

Appendix

Table 4 : Benchmarks and tool stack; all methods on a benchmark share its backend and judge.

Parameter	Scope	Value
Prompt length	per context	$8{,}192$ tokens
Response length, including tool outputs	per context	$32{,}768$ tokens
Orchestrator iterations $K$	per task	$30$
Session timeout	per task	$5{,}400$ s
Executor turns	per sub-task	$200$
Executor timeout, evaluation / training	per sub-task	$600$ s / $1{,}800$ s

Appendix

Table 5 : DAGent framework parameters, grouped by scope. “Per context” denotes one Orchestrator turn or one Executor sub-task; “per task” denotes one benchmark query.

Workflow	Setting	Cap
ReAct Agent	Context window	$32$ K or $109$ K tokens
Summary Agent / Fold Agent	Sessions per task	$10$
Flash-Searcher	Action steps per task	$40$
Flash-Searcher	Summary interval	$4$ (release default $8$ )
FlowSearch	Main iterations per task	$5$
	Planner iterations / nodes	$2$ / $7$

Appendix

Table 6 : Budget caps of the baseline workflows (the DAGent caps are listed in Table 5 ).

Parameter	Value
Hardware	$2\times$ H200, tensor parallel $2$
LoRA rank / alpha	$64$ / $64$ , all linear layers
Optimizer, learning rate	AdamW, $1\times 10^{-5}$
Tasks per batch, rollouts per task $G$	$32$ , $8$
PPO mini-batch (tasks / trajectories)	$16$ / $128$
Asymmetric clip $\epsilon_{\ell}$ / $\epsilon_{h}$	$0.20$ / $0.28$

Appendix

Table 7 : RL training and DAGRPO hyperparameters.

Subset	N	Fleiss’ $\kappa$ (3H)	Cohen’s $\kappa$ (H vs J)	Raw %
BrowseComp-Plus	50	0.98	0.96	98.0
GAIA	50	0.98	0.96	98.0
xbench-DeepSearch	50	0.96	0.92	96.0
Aggregate	150	0.97	0.95	97.3

Appendix

Table 8 : Human agreement on the LLM judge across

150

(question, predicted answer, reference answer) tuples. Fleiss’

\kappa

is computed on the three-human ensemble; Cohen’s

\kappa

is between the human-majority label and the LLM judge; raw % is between the same pair.

Backbone	Vendor	Agent Paradigm	BrowseComp-Plus				GAIA				xbench-DS
Backbone	Vendor	Agent Paradigm	Easy	Med.	Hard	Avg.	L1	L2	L3	Avg.	Avg.
Qwen3-32B	Alibaba	ReAct Agent (32K)	50.0	16.0	2.0	22.7	46.2	28.8	8.3	33.0	54.0
		ReAct Agent (109K)	66.0	24.0	4.0	31.3	51.3	36.5	16.7	39.8	58.0
		Summary Agent	80.0	36.0	0.0	38.7	51.3	28.8	25.0	36.9	55.0
		Fold Agent	82.0	40.0	2.0	41.3	56.4	38.5	16.7	42.7	55.0
		Flash-Searcher	80.0	34.0	2.0	38.7	56.4	40.4	25.0	44.7	58.0

Appendix

Table 9 : Cross-backbone spot check with four mid-size open-source backbones; all training-free with the same tool stack as the main results; Pass@1 (%); best per backbone in bold . The Qwen3-32B block reproduces the corresponding rows of Table 1 .

Benchmark	Workflow	Pass@1	Tool calls	External calls	Steps	Calls / step	Tokens (M)	Time (s)	Nodes	Off-chain
BrowseComp-Plus	DAGent (Full)	47.3	67.6	60.8	42.7	1.58	1.20	541.6	11.8	0.25
	Plan-then-Patch variant	42.0	87.9	82.3	54.8	1.60	1.68	731.9	17.6	0.40
	Flash-Searcher	38.7	51.3	42.1	35.8	1.43	0.91	388.7	–	–
	FlowSearch	43.3	83.5	76.7	47.6	1.75	1.57	714.5	–	–
GAIA	DAGent (Full)	55.3	36.8	33.1	31.2	1.18	0.66	461.8	5.4	0.20
	Plan-then-Patch variant	50.5	45.3	42.0	37.6	1.20	0.85	584.6	7.3	0.30

Appendix

Table 10 : Per-task cost of four workflows with Qwen3-32B and the same retrieval backends. Pass@1 values are from Table 1 and Table 2 . Tool calls, external tool calls, execution steps, tokens, and time are per-task means; the calls / step column is the sum of tool calls divided by the sum of execution steps.

Figure 6 : Behavioral statistics of the Qwen3-32B training-free runs. (a) Distribution of the Orchestrator evaluation outcomes over all executed nodes per benchmark and over first-batch nodes pooled across the three benchmarks. (b) RecallTool usage: mean number of calls per task and share of tasks that make at least one call, per benchmark.

Figure 7 : Per-seed Pass@1 (%) for the six RL-trained variants in the Qwen3-8B block of Table 1 , broken out by training seed (42 / 123 / 777) on BrowseComp-Plus, GAIA, and xbench-DeepSearch.

Figure 8 : DAGRPO training dynamics on Qwen3-8B with LoRA. (a) BrowseComp-Plus validation Pass@1 over training steps, averaged over the three training seeds, so that the end points equal the Qwen3-8B means of Table 1 . (b) Per-step training-time reward gap

\bar{r}_{\mathrm{chain}}-\bar{r}_{\mathrm{off\text{-}chain}}

; the

\alpha=0.5

topology credit separates the two groups, while

\alpha=1.0

gives zero gap by construction. (c) Per-step Orchestrator format-warning rate from the structural validator. (d) Per-step on-chain ratio, the fraction of Executor nodes inside the answer-inclusive closure.

Benchmark	Setting	Pass@1	Tokens	Tool calls	Steps	Nodes	Off-chain
BrowseComp-Plus	DAGent (training-free)	40.0	1.14M	65.5	41.5	12.2	0.25
	+ RL (GRPO)	46.0	0.96M	59.6	36.5	11.3	0.21
	+ RL (DAGRPO)	49.6	0.90M	56.3	34.5	10.3	0.16
GAIA	DAGent (training-free)	46.6	0.62M	35.5	29.5	5.3	0.20
	+ RL (GRPO)	50.8	0.53M	33.0	26.8	4.6	0.17
	+ RL (DAGRPO)	53.4	0.51M	31.6	25.7	4.0	0.13

Appendix

Table 11 : Effect of RL training on the per-task efficiency of DAGent on Qwen3-8B; the RL rows are means across three independent training seeds. Pass@1 values are taken from the Qwen3-8B block of Table 1 . Tokens is total input plus output; tool calls aggregate search, open_page, recall, and Orchestrator plan invocations; nodes is the final graph size; off-chain is the fraction of Executor nodes outside the answer-inclusive closure, as in Table 10 .

Explore similar work

May 15, 2026cs.CL

Argus: Evidence Assembly for Scalable Deep Research Agents

Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.

Zhen Zhang, Liangcai Su, Zhuo Chen +7

MiroMind AI

Deep Research

Ask PDF

Jun 5, 2026cs.AI

DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning

Deep Research (DR) has emerged as a new agentic paradigm to tackle complex, open-ended research tasks, demanding systems that can iteratively frame problems, acquire evidence, verify sources, and synthesize long-form reports. In practice, however, current DR systems are constrained by four interrelated limitations: long-horizon planning over an underspecified scope, the bottleneck of decomposing and scheduling such tasks within a single agent, hallucination risk in long-form synthesis, and limited process auditability. This technical report presents DuMate-DeepResearch, a multi-agent DR framework built on the Qianfan Agent Foundry. The framework decouples the Agent Core, which handles task understanding, planning, and scheduling, from an extensible Tool Ecosystem for retrieval, evidence acquisition, and report rendering, making every intermediate decision and tool invocation explicitly traceable. Building on this infrastructure, DuMate-DeepResearch further introduces three mechanisms: (i) a graph-based dynamic planning strategy expands the research roadmap coarse-to-fine and continuously revises it through reflection, re-planning, backtracking, and parallel branching; (ii) a recursive two-level execution design delegates each complex search sub-task to an inner Search Agent that runs its own planning loop, isolating noisy retrieval and stabilizing long-horizon execution; (iii) a rubric-based test-time optimization mechanism dynamically generates task-specific quality criteria and uses them as live reasoning scaffolds for evidence-grounded synthesis and adaptive stopping. Across two deep research benchmarks, DuMate-DeepResearch establishes new state-of-the-art results: the best overall score (58.03%) on DeepResearch Bench, and the best overall score (61.95%) on DeepResearch Bench II while ranking first in information recall and analysis.

Lingyong Yan, Can Xu, Yukun Zhao +13

DuMate Team, Baidu AI Cloud

Deep Research Model-Based Multi-Agent Systems Tool-Grounded Reasoning Recursive Models

Ask PDF

Aug 1, 2026cs.CL

Deep Research Pretraining via Predictive Navigation

Deep research agents are often trained on expensive, environment-grounded tool-use trajectories that require repeated retrieval, document inspection, and report evaluation. We introduce Deep Research Pretraining (DRP), an offline framework that derives predictive navigation supervision from naturally occurring evidence structures. Given a citation-bearing or hyperlinked passage, DRP constructs a proxy research objective, recovers linked evidence and graph-related alternatives, and converts them into search-open-write trajectories. This teaches models what to search for, which documents to inspect, and how to synthesize evidence, without a live retrieval environment or executed policy rollout. We instantiate DRP on scholarly citation graphs (DRP-Paper) and Wikipedia hyperlinks (DRP-Web), continually pretrain separate Qwen3-14B-Base models on 1B tokens, and fine-tune them on controlled fractions of 13K agent trajectories. Across five independently sampled subsets at each low-data budget, both variants consistently outperform matched no-DRP models on DeepResearch Bench. With one quarter of the SFT data, DRP-Web even surpasses a fixed no-DRP full-data checkpoint, with gains transferring to ResearchQA, WebWalkerQA, and SimpleQA. Starting from matched low-data SFT checkpoints, the DRP-Web advantage also persists through subsequent agentic RL. Source-matched and evidence-mismatch controls indicate that these improvements arise from evidence-conditioned navigation rather than domain exposure or agent-format imitation. DRP thus provides a promising complementary approach to trajectory-based agent training.

Jiang Zhou, Zhiyuan Fan, Xing Wu +3

Hunyuan Team, Tencent

Deep Research Pretraining Single Trajectory Citations

Ask PDF