acceptodds
Under review as a conference paper at ICLR 2027

DocAgentic-R: Learning Utility-Aware Retrieval for Document Deep Research

Abstract

Training retrievers with downstream utility has improved agentic search, but existing frameworks assume homogeneous text passages. We study the same idea when candidates are text-carried document units that differ in modality and granularity. method trains a BGE-M3 retriever with local relevance (lr), a prefix-freeze suffix-replan answer-contribution score (gac), optional chain and localization labels, and multi-positive contrastive learning. On document-disjoint BlueprintSynth-Eval (2,430 questions; single runs), the full stack raises full-chain@20 from 0.588 to 0.674 and MRR from 0.713 to 0.800, above frozen Qwen3-Embedding-4B at 0.605. lr-only training falls below the backbone; gac is the signal that restores it. On the same pool, simpler positive rules reach 0.678–0.681, and text-only positives fall to 0.657. External transfer is about one point: MMDocIR Page R@5 moves from 0.705 to 0.710 and Layout R@5 stays below the backbone; the best LongDocURL setting gains 1.7 R@1 points and 1.0 LLM-judged point. Figure-unit coverage does not rise. Utility supervision helps synthetic multi-hop chain recovery. It does not, by itself, beat simpler utility labels or a native visual encoder.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Document Deep Research, Utility-Aware Retrieval, Multimodal Document Retrieval, Agentic Retrieval, Evidence Chain

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jdPcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to train a utility-aware retriever for document deep research, placing evidence carried by different information carriers such as text, tables, and figures, as well as evidence at different granularities such as chunk, page, and summary, into a unified retrieval pool. The method builds on BGE-M3 and uses four types of supervision signals—LR, GAC, ECC, and LOC—to select positive samples and perform multi-positive contrastive learning. Among these, tables and figures participate in retrieval mainly through text descriptions generated by a VLM. On BlueprintSynth-Eval, the full method improves Full-chain@20 from 0.588 to 0.674, but simple positive-sample rules also achieve comparable or higher results, and the gains on external retrieval and answering tasks are small.

Critical strengths

The problem is practically meaningful. Document retrieval needs to handle different information carriers, different granularities, and citation locations; local relevance alone may not reflect how much evidence helps the final answer. Constructing the supervision signal in an offline stage and using only the trained retriever at deployment is also a practically valuable design.

Critical weaknesses
  1. I am still not sure whether the existing experiments clearly demonstrate the advantage of the proposed mixed retrieval design. The full method achieves Full-chain@20 of 0.674, while simple positive-sample rules on the same mixed pool reach 0.678–0.681. External gains are also small, for example MMDocIR Page R@5 improves from 0.705 to 0.710. These comparisons do not answer whether unified retrieval is more effective than retrieving separately by information carrier or granularity and then fusing the results. The coverage of existing benchmarks and baselines is also insufficient to show under which conditions this design helps.

  2. It is still not clear to me why the four signals are combined in this way. The paper gives some priority and gating rules, but does not sufficiently explain the basis for choosing these rules, or how conflicts between signals are handled concretely. For example, chain units with GAC≤0 are demoted to weak positives, while the training section also says that at most three positives passing the GAC gate are selected. How these weak positives ultimately participate in training remains unclear, making the positive-sample selection process difficult to reproduce.

Questions for authors
  1. Under the same encoder, corpus, and context budget, how does unified mixed retrieval compare with "retrieve separately, then fuse"? Which types of questions benefit the most?
  2. How are the signal priorities and gating rules determined? How do weak positive samples with GAC ≤ 0 participate in the training loss?
  3. In the external benchmarks, which subsets directly require evidence retrieval across information carriers and across granularities? Can comparative results for the relevant retrieval strategies be reported on one such subset?
Reviewer_QpqQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DocAgentic-R trains a BGE-M3 retriever over heterogeneous document units represented through text and VLM-generated captions. It combines local relevance, answer contribution measured through evidence removal and trajectory replanning, and optional evidence-chain and localization supervision. Evaluation covers a document-disjoint synthetic chain-retrieval benchmark and three external benchmarks; improvements are strongest on synthetic chain recovery, while external gains are modest and simpler supervision rules match or outperform the complete method.

Critical strengths
  • A practically relevant training target. The paper directly optimizes first-stage retrieval over document units with different modalities and granularities. Keeping the utility teachers offline preserves a straightforward deployment architecture without an online teacher or reranker.

  • Informative reporting of limitations. The authors explicitly report that simpler positive-selection rules outperform the complete stack, external gains are small, and figure coverage does not improve. These findings provide useful guidance about which additional supervision mechanisms currently lack demonstrated value.

Critical weaknesses
  • The evaluation does not clearly establish the magnitude of the training benefit. Table 2 reports a gain from 0.588 to 0.674, whereas Table 7 reports 0.665 for frozen BGE-M3 and 0.674 for the trained model on the same named synthetic benchmark, without explaining the protocol difference. Moreover, Appendix G selects checkpoints and hyperparameters using 150 MMDocIR Eval and 50 LongDocURL questions that remain included in the reported test metrics, weakening the interpretation of the small, single-run transfer gains.

  • The distinctive supervision machinery has limited supporting evidence. Simpler same-pool rules achieve 0.678–0.681 versus 0.674 for the complete method, so the additional replanning, chain, and localization machinery has no demonstrated advantage. The text-only comparison also discards 295 of 850 training groups, confounding the contribution of non-text positives with training-set size; consequently, the experiments do not yet isolate a sufficiently strong methodological or mechanistic contribution.

Questions for authors
  • Clarification: What changes between the frozen BGE-M3 evaluations in Tables 2 and 7—candidate index, preprocessing, query encoding, or another setting? Which frozen-versus-trained comparison uses fully identical evaluation inputs?

  • Assessment-critical evaluation: After excluding all calibration questions from the reported test sets, do the external improvements persist across training seeds and under paired confidence intervals?

  • Clarification: Sections 4.3–4.4 describe voting across two LLM families, whereas Appendix C explicitly states that only one Qwen-family judge is used. Which procedure produced the reported results, and what fraction of GAC labels comes from replanning versus the fallback procedures?

  • Clarification: Under Equation 5, an unannotated unit cannot increase the intersection with the gold chain, so its coverage contribution is zero. How does ECC provide the claimed complementarity bonus to unannotated units, and does the implementation use a different definition?

Reviewer_Z73nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DocAgentic-R, a retriever training framework for document deep research. It combines four supervision signals: local relevance (LR), global answer contribution (GAC), evidence-chain contribution (ECC), and localization correctness (LOC). GAC estimates a unit’s contribution by removing it from a retrieval trajectory, replanning the remaining steps, and measuring the resulting change in answer quality. The signals are used to select multiple positive units for contrastive training of a BGE-M3 retriever. On BlueprintSynth-Eval, DocAgentic-R improves full-chain@20 from 0.588 for frozen BGE-M3 to 0.674. However, simpler positive-selection rules achieve 0.678–0.681. Transfer to external benchmarks is modest and mixed: MMDocIR Page R@5 improves from 0.705 to 0.710, while Layout R@5 declines. The paper reports single-run results and notes that its synthetic evaluation shares a generator and distributional prior with training.

Critical strengths
  • The paper considers a relevant setting in which evidence units vary in modality and granularity, and their value may depend on the existing evidence chain.
  • GAC offers a concrete way to connect retrieval supervision to downstream answer quality.
  • The retriever is lightweight at deployment: the teacher models are used to construct supervision, not for online retrieval.
  • The paper reports negative as well as positive results, clearly identifying the limited external transfer and the lack of consistent gains from additional supervision signals.
Critical weaknesses
  • The strongest results are on a synthetic evaluation set whose training and test partitions share a generator and distributional prior. Generalization to realistic document research remains uncertain.
  • Simpler positive-selection rules outperform the full method on the main synthetic metric, making the added value of the full supervision pipeline unclear.
  • Improvements on external benchmarks are small and inconsistent; layout retrieval declines relative to the frozen BGE-M3 baseline.
  • GAC depends on multi-turn replanning and answer judging, which may be costly and sensitive to agent variability and fallback procedures.
  • The retriever does not encode page images directly. Figures and tables are represented by generated captions, limiting the evidence for gains in native visual-document retrieval.
  • The evaluation omits several relevant retrieval or document-agent baselines, including VRAG-RL, UniDoc-RL, and EviSRAG. Without comparison to these methods, it is difficult to assess DocAgentic-R’s standing relative to recent approaches.
Questions for authors
  • Why does the full supervision pipeline underperform simpler positive-selection rules on BlueprintSynth-Eval? Can the authors identify which parts of the decision process account for this gap?

  • How sensitive are GAC labels to the frozen agent, answer judge, and replanning procedure? Can the authors report agreement or stability across different agents or judges?

  • Can the authors provide multi-seed results or confidence intervals, particularly for the modest external-benchmark gains?

  • How much does BlueprintSynth-Eval performance depend on the shared generator and distributional prior? Can the authors evaluate with a different question generator or on a human-authored multi-hop set?

  • Given the decline in MMDocIR Layout R@5, what evidence supports the claim that the method improves retrieval of localized document evidence?

  • Can the authors quantify the end-to-end cost of constructing GAC supervision, including unsuccessful rollouts and retries, and relate it to the observed transfer gains?

-Could the authors compare DocAgentic-R with VRAG-RL[1], UniDoc-RL[2], and EviSRAG[3], or explain why these methods are not directly comparable to the proposed retriever-training setting?

[1] VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning [2] UniDoc-RL: Coarse-to-Fine Visual RAG with Hierarchical Actions and Dense Rewards [3] VisRAG2.0: Mitigating Visual Hallucinations via Evidence-Guided Multi-Image Reasoning in Visual Retrieval-Augmented Generation

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.