acceptodds
Under review as a conference paper at ICLR 2027

Perceiving Before Cognizing: Agentic Visual Enhancement for Robust Document Parsing

Abstract

Modern document parsing models perform well on clean images but remain vulnerable to real-world visual degradations, motivating input enhancement to improve robustness across parsers. Yet document-dependent enhancement effects and potentially erroneous tool choices call for adaptive policies that can reassess intermediate results and explore alternative paths. We propose Perceiving Before Cognizing, which defines visual enhancement before parsing as an independently learnable interactive decision-making task, using parsing utility to guide optimization and improve the accessibility of task-relevant visual evidence. We introduce the AgenticDoc framework to support the reuse of this enhancement capability across parsers. To support policy learning for this task, we introduce Action-to-Token On-Policy Distillation (A2T-OPD), which bridges the gap between action-level teacher supervision and token-level generative policy learning. AgenticDoc improves all 13 evaluated parsers on Wild-OmniDocBench by 7.15 on average, including an average gain of 7.44 across nine parsers not used for training feedback. Overall performance on OmniDocBench v1.6 documents is preserved, with an average score change of +0.41. On PureDocBench, average gains across ten parsers reach 2.48 across the three scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MultiModal Agent, Document Recognition

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XwbCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AgenticDoc, an agent-based framework that enhances document images before passing them to a frozen document parser. The agent selects enhancement tools and parameters, observes intermediate results, and can revisit historical image states before deciding to stop. Training combines online image degradation, parsing-based feedback, and Action-to-Token On-Policy Distillation (A2T-OPD). A2T-OPD converts a search-derived distribution over executable actions into conditional token targets through a semantic action trie and is jointly optimized with a critic-free PPO objective.

The method is evaluated on Wild-OmniDocBench, OmniDocBench v1.6, and PureDocBench. On Wild-OmniDocBench, the authors report an average improvement of 7.15 points across 13 parsers, including 7.44 points across nine parsers not used for training feedback, while largely maintaining average performance on clean OmniDocBench inputs.

Critical strengths
  1. Consistent transfer across downstream parsers. The strongest empirical result is that AgenticDoc improves all nine parsers not used for training feedback on Wild-OmniDocBench. These include different parsing architectures and general-purpose VLMs. This breadth provides meaningful evidence that the learned enhancement policy transfers beyond the specific models supplying its training signal.

  2. A practically useful formulation of document enhancement. Optimizing an input-side policy against downstream parsing quality allows existing parsers to benefit without updating their parameters. Evaluating both degraded and clean documents is also valuable, since unnecessary enhancement can damage otherwise readable inputs. Together, these aspects make the framework relevant to document-processing pipelines facing variable capture conditions.

Critical weaknesses
  1. The novelty and contribution of the proposed components are not sufficiently established. The related-work discussion omits closely relevant approaches, including RL-Restore (CVPR 2018), which learns sequential restoration-tool selection; AgenticIR (ICLR 2025), which incorporates reflection, rollback, and replanning; and AdaptiveISP (NeurIPS 2024), which optimizes adaptive image-processing pipelines using downstream task performance. These precedents do not negate the paper’s document-specific contributions, but they make a clearer positioning and a relevant adapted baseline important.

    The ablations also leave the main contributions insufficiently isolated. Table 3 does not directly evaluate intermediate visual feedback, historical-state revisiting, or adaptive degradation sampling. The action-order analysis in Appendix B does not establish the necessity of online replanning or rollback. In Table 4, adding A2T-OPD introduces search-derived supervision as well as a new distillation objective, while Search-SFT differs in its supervision targets and training-state distribution. Consequently, these comparisons do not isolate the value of the action-to-token formulation. A controlled comparison using identical student-visited states, candidate actions, and search-derived teacher distributions would be more informative. Furthermore, the teacher probability-preservation result should be distinguished from any claim about equivalence between the local training loss and an action-level KL objective.

  2. Incomplete methodological details and limited reporting weaken reproducibility and the interpretation of the results. Important details remain unspecified, including the training search procedure and evaluation budget, complete action-parameter discretization, the dewarping model, training-data sources and annotation provenance, and several RL hyperparameters. The critic-free PPO description also lacks an explicit return-computation algorithm, particularly for historical-state revisiting and discarded branches. Additionally, the main text describes filtering for images that both parsers recognize correctly, whereas Appendix D.1 uses agreement between their outputs; agreement alone does not establish correctness.

    The reported gains lack uncertainty estimates or repeated-training results, which matters particularly for the 0.86-point improvement over Search-SFT and the 0.68-point improvement over A2T-OPD alone. Reporting primarily overall scores also obscures which parsing components and degradation conditions benefit. Although average enhancement latency is provided, the measurement conditions, end-to-end parsing overhead, and search-related training costs are insufficiently documented. Training on 1,000 source images demonstrates limited source-data requirements, but does not by itself establish computational efficiency relative to parser training on larger datasets.

Questions for authors
  1. Search-teacher construction: Could you provide pseudocode specifying candidate actions and parameter values, search expansion and pruning, and the evaluation budget used to construct QtQ_t? Does the beam width of 3 mentioned in Appendix B also apply to training? How many parser evaluations are required per student-visited state on average?

  2. PPO update weights: Could you provide the exact process-reward definition and return-computation algorithm, including stopping, invalid calls, historical-state revisiting, and discarded branches? Is the policy-update weight simply the return divided by a running RMS, without a baseline? Please also specify any discounting and how the running RMS is maintained.

  3. Attribution of the A2T-OPD gain: Do you have a comparison that fixes the student-visited states, search results, and soft teacher distributions while changing only the supervision method—for example, comparing trie-based distillation with teacher-probability-weighted sequence cross-entropy over complete actions? Such evidence would help distinguish the contribution of action-to-token distillation from that of search supervision.

  4. Evidence for interactive correction: How does the full policy compare with a policy that plans from the initial image only, or one that observes intermediate images but cannot revisit earlier states? What fraction of evaluation trajectories actually revisit historical images, and what benefit does this provide? These results would directly inform my assessment of the paper’s central motivation.

4 comments
AuthorsOpenReview ↗

Rebuttal [1/2]

Thank you for your careful review and constructive suggestions. We respond in two parts: this part addresses the two weaknesses, and Part 2 answers the four specific questions.

W1 – Novelty and contribution

Thank you for highlighting the need for clearer positioning and component-level evidence. We have expanded Section 2.3 to discuss RL-Restore, AgenticIR, AdaptiveISP, and DocRes. The completed DocRes/AdaptiveISP comparisons, settings, and visual analysis are presented in General Response R3. AgenticDoc exceeds these baselines by 11.46 and 10.42 points, respectively, on the four-parser Wild-OmniDocBench average, while largely preserving the clean-document average.

We have also completed the adaptive-sampling and teacher-search controls:

SettingPaddleOCR-VL-1.6MinerU2.5-ProUnlimited-OCROvisOCR2Avg
Without adaptive degradation sampling92.4391.9187.8691.2890.87
Single-step teacher instead of multi-step teacher92.1891.7487.2691.0690.56
Full AgenticDoc92.7092.0888.2791.6791.18

These are Overall scores (%) on Wild-OmniDocBench. Removing adaptive sampling reduces the average by 0.31 points; replacing the multi-step teacher with a single-step teacher reduces it by 0.62 points. Both changes reduce scores on all four parsers. The sampling control retains online degradation generation, and the single-step teacher retains candidate-action evaluation. The latter therefore tests the added value of considering subsequent enhancement steps.

The A2T-OPD attribution question is discussed in Q3, and the completed revisiting comparison in Q4.

W2 – Reproducibility and reporting

We appreciate your detailed requests. Appendix A.3 now provides teacher-construction pseudocode. During training, cached search results are queried first; missing results are computed online. Multi-level caching reuses intermediate images and parsing results.

The training source collection contains 1,000 images after filtering: 300 each from DocLayNet, PubLayNet, and DocStruct4M, plus 100 web-document images. We use only their images, without the original annotations. Training pseudo-references are obtained from MinerU2.5-Pro and PaddleOCR-VL-1.6, retaining pages with NED < 0.02. Agreement is a filtering proxy; shared errors can remain. These details are clarified in Section 3.2 and Appendix D.1.

Our dewarping tool is based on DewarpNet, adapted and trained by our team. We will publicly release the complete training dataset and code, including the trained dewarping model.

To address the variability of the original +0.86 and +0.68 differences, we completed repeated training with four seeds under the same evaluation protocol:

SeedA2T-OPD onlySearch-SFTPPO + A2T-OPD
4290.5090.3291.18
202690.6990.2190.92
1790.3890.3391.27
373990.0990.2191.32
Mean ± sample SD90.42 ± 0.2590.27 ± 0.0791.17 ± 0.18

Scores are four-parser averages on Wild-OmniDocBench. The joint method leads in every run: its advantage ranges from 0.71 to 1.11 points over Search-SFT and 0.23 to 1.23 over A2T-OPD only. These results show that the observed advantages persist across the evaluated training seeds.

Enhancement-action counts exclude parse; AgenticDoc runtimes include policy inference, explicit parse generation when present, tool execution, and final-image output, but exclude downstream parsing.

For inference cost, General Response R1 and Section 4.2.2 report enhancement-plus-parsing measurements: an average 0.54 s/image increase on clean OmniDocBench v1.6 and 12.10 s/image reduction on Wild, across the four parsers. Parser measurements use serial processing on a single A100-80GB GPU.

The 1,000-image figure describes training-data requirements. It does not establish lower training compute than large-scale parser training; teacher-search costs are discussed separately in Q1.

AuthorsOpenReview ↗

Rebuttal [2/2]

Q1 – Search-teacher construction

Thank you for requesting these details. We have added teacher-construction pseudocode to Appendix A.3, specifying the budget-conditioned returns and teacher probabilities.

During training, we query the cache for the current image and remaining budget, computing missing results online. Search uses depth 3 and beam width 3, evaluating up to six first-level candidates and retaining the best three nodes per level. Each retained node has at most five successors because consecutive identical actions are excluded.

A full-budget query therefore evaluates at most 6 + 3×5 + 3×5 = 36 candidate images. Across four feedback parsers, these require at most 144 parser–image evaluations before cache reuse. These are upper bounds, not measured averages.

The six calls are denoising, tone adjustment, dewarping, USM sharpening, 1.5× resizing plus USM, and Laplacian sharpening. Each acts on its search-node image; parameters are in the supplied code.

The teacher uses the best explored parsing-score improvement within budget, minus 0.05 per enhancement call. Any evaluated intermediate image can be an endpoint.

Q2 – PPO update weights

We are happy to clarify the reward and return assignment. A successful call receives the parsing-score difference between its output and selected input image, minus costs. The selected input may be historical. The PPO enhancement cost is 0.0001, invalid-call penalty 0.20, and tool-failure penalty 0.02, separate from the teacher-search cost of 0.05. Immediate parse reward is zero.

Returns are undiscounted, accumulated backward along the final image's dependency path. Discarded-branch actions retain only their own immediate negative feedback; nonnegative off-path rewards and unscorable transitions are excluded from PPO return-based updates.

The update weight is the assigned return divided by running RMS, without baseline or mean subtraction. Let v be the global mean squared nonzero assigned returns, aggregated across workers using squared-return sums and counts. The first nonempty batch initializes m=v; later batches update m ← 0.99m + 0.01v. The weight is R / √max(m, 10⁻⁸), shared by generated tokens in the assistant turn; environment observations are excluded.

Q3 – Attribution of the A2T-OPD gain

We appreciate this suggestion. The key challenge is mapping action-level preferences to the few tokens that determine tool and parameter choices. Full-sequence supervision also optimizes shared syntax and fixed fields that do not distinguish candidate actions, potentially diluting the effective gradient signal for these decisions. Teacher-probability-weighted sequence CE still distributes supervision across the complete call.

Before A2T-OPD, we explored complete-action distillation using forward, reverse, and mixed KL objectives. These preliminary implementations did not yield effective policy learning. This experience motivated the decision-token formulation in Section 3.3 and Figure 3.

A2T-OPD explicitly concentrates alignment where candidate actions diverge: teacher mass is conditioned at each trie branch, student probabilities are normalized over the same candidate tokens, and all candidate tokens at student-visited decision positions receive supervision. Shared and fixed tokens remain in context without direct distillation loss. The proposition establishes teacher-distribution factorization, without claiming that the visited-token loss equals action-level KL.

Table 5 reports a +0.86-point advantage over Search-SFT under the same SFT initialization and optimization-step budget, with consistent advantages across four seeds in W2. A2T-OPD addresses the underlying supervision bottleneck by directing the soft teacher’s guidance to the token decisions that determine the enhancement action.

Q4 – Evidence for interactive correction

Thank you for suggesting a direct recovery test. General Response R2 and Appendix E, Table 10 report the completed comparison. Revisiting occurs on 14/1,350 pages (1.04%), always returning to the original image after the first enhancement.

For each case, we retain the path before revisiting and exhaustively evaluate all legal continuations that process only the latest image within the same remaining budget, including immediate parse. Against the best such result, AgenticDoc achieves 12 wins, 2 ties, and a +2.12pp four-parser mean gain across these 14 cases.

This demonstrates recovery at observed revisiting points. It is a targeted comparison rather than a separately trained initial-image-only or no-revisiting policy ablation.

We hope these clarifications and additional results address your concerns and provide a clearer understanding of our contributions. If any questions or concerns remain, we would appreciate the opportunity to address them and would be happy to respond promptly.

Thank you for the additional experiments and clarifications. The multi-seed results strengthen confidence in the reported gains, and the training procedure is now clearer. However, my main methodological concerns remain.

The path-filtered, RMS-scaled returns retain negative feedback from discarded branches while excluding nonnegative feedback. This heuristic’s relationship to the full interaction objective remains unclear, and its benefits over simpler credit-assignment schemes are not established. My concern is not the absence of a critic itself, but the insufficient justification of these choices.

Likewise, qualitative reports of unsuccessful complete-action distillation do not replace controlled comparisons using the same teacher distributions, student-visited states, and training budget. The results support the combined system’s effectiveness but do not isolate A2T-OPD’s contribution. I therefore maintain my original rating.

Thank you for taking the time to read our earlier response and follow up. We appreciate your questions about credit assignment and distillation, which helped us identify where further evidence would be most useful. We have now completed two controlled comparisons on Wild-OmniDocBench to address these points.

1. PPO credit assignment for discarded branches

To address your question about credit assignment, we compare three schemes on Wild-OmniDocBench using the same SFT initialization, A2T-OPD setup, enhancement budget, and training budget.

  • Chronological reward-to-go sums rewards from each step to the end of the interaction. Each reward is the change in the current image’s score minus applicable costs and penalties. If a later revisit makes the net interaction return positive, a harmful first enhancement can receive a positive return even when its output is discarded.
  • Final-path returns w/o off-path feedback accumulates returns along the image lineage that produced the final output. Successful enhancements whose outputs are discarded receive zero PPO weight.
  • Final-path returns w/ off-path feedback (ours) uses the same final-path rule and gives a discarded enhancement its immediate reward when that reward is negative.
Credit assignmentPaddleOCR-VL-1.6MinerU2.5-ProUnlimited-OCROvisOCR2Avg.Revisited pagesFirst-enhancement dropsEnhancements/page
Chronological reward-to-go92.3891.9487.7591.2990.8465108/1,331 (8.11%)2.12
Final-path returns w/o off-path feedback92.6192.0088.2591.5191.092752/1,280 (4.06%)1.77
Final-path returns w/ off-path feedback (ours)92.7092.0888.2791.6791.181439/1,280 (3.05%)1.67

We count a first-enhancement drop when the first successful enhancement lowers the average score relative to the original image. The denominator counts pages with a scorable first enhancement.

The chronological baseline assigns credit using the net return of the full interaction. A later revisit can therefore give positive credit to an earlier enhancement whose output was discarded. Assigning recovery gains along the final image lineage improves the mean parsing score from 90.84 to 91.09, reduces the first-enhancement drop rate from 8.11% to 4.06%, and lowers enhancement use from 2.12 to 1.77 calls per page. These results are consistent with crediting the actions that produced the recovered image.

The last two rows show what off-path feedback contributes. The extra feedback gives PPO a local signal to discourage a harmful discarded enhancement even when the interaction ends with a net gain. In this comparison, adding off-path feedback reduces first-enhancement drops from 52 to 39 (−25%), revisited pages from 27 to 14, and average enhancement calls by 0.10, with comparable final parsing scores (91.09 and 91.18). The main benefit we observe is fewer early degradations and fewer enhancement calls at similar final parsing quality.

2. Contribution of action-to-token distillation

To isolate the contribution of action-to-token distillation, we use the original A2T-OPD-only run from Table 5 as the reference. We train the teacher-weighted full-action CE baseline from the same SFT checkpoint by replaying the exact sequence of states visited during that run, including intermediate states. At each state, CE uses the recorded candidate actions, search results, and soft teacher probabilities to weight the cross-entropy over complete action sequences. Both runs use the same training budget, with PPO disabled.

Supervision methodPaddleOCR-VL-1.6MinerU2.5-ProUnlimited-OCROvisOCR2Avg.
Teacher-weighted full-action CE91.8291.4287.2290.6290.27
A2T-OPD92.0791.8287.4090.7090.50

On Wild-OmniDocBench, A2T-OPD raises the mean parsing score from 90.27 to 90.50 (+0.23 points), with gains on all four parsers (+0.08 to +0.40 points). With training states and teacher targets held fixed, these gains support the contribution of the action-to-token supervision rule.

Candidate calls often share syntax and fixed fields, so many tokens covered by full-action CE carry little information about which action the teacher prefers. At each trie branch, A2T-OPD converts the teacher probability mass of actions compatible with the current prefix into a soft target over valid next tokens. This gives the student a direct comparison between the choices that determine the tool or its parameters, even when the complete calls share many tokens or differ in length.

We hope these comparisons address the questions you raised. Thank you again for taking the time to consider our response. We would be grateful for any further feedback.

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

This paper proposes AgenticDoc, which employs a learnable visual enhancement agent prior to document parsing. The agent selects enhancement tools, tool parameters, and determines whether to process the current image or revert to the historical image state based on the current image and interaction history; ultimately deciding to cease enhancement and pass the image to the downstream parser. The experiment covers 13 parsers and three benchmarks. The paper reports significant gains in real-world degradation scenarios, transfer to parsers not involved in training feedback, and overall results that are close to not compromising the performance of clean documents.

Critical strengths

1.The motivation is interesting. Rather than retraining every downstream parser, the paper turns enhancement into a reusable pre-processing policy. Having the agent decide whether to enhance at all, which operation to use, and when to stop is clearly more reasonable than applying one fixed filter to every image, and the controls back this up: random selection actually hurts (−0.77), the best fixed operator (Dewarp) gains only 1.61, while the trained agent gains 6.50.

2.The results generalize beyond the models used in training. Gains hold across all 13 parsers on Wild-OmniDocBench, including the 9 unseen ones (+7.44 on average), and across parser families (pipeline experts, end-to-end experts, general VLMs). The fact that clean-document scores are not sacrificed (+0.41 on OmniDocBench v1.6) is a meaningful point in favor.

Critical weaknesses

1.The mechanisms the paper emphasizes are not validated individually. The core design of the paper also includes historical image backtracking, dynamic degenerate sampling, and online search. It is not yet clear from the current experiments whether each of these components brings its own benefits.

2.Optimizing parse scores invites overfitting to the feedback parsers or metrics. The unseen-parser results mitigate this concern considerably, but I would still like to see the distribution of per-parser gains and an analysis of which visual changes actually drive cross-model benefit.

3.The averages hide some regressions. Some models on PureDocBench exhibit negative gains in the clean or digitally degraded subsets, and not every parser on Wild-OmniDocBench and the clean set shows the same magnitude of improvement.

4.Efficiency and training cost are under-reported. The paper provides the average number of augmentation operations and processing time for each image, but it does not fully break down the additional costs incurred by agent inference, image tool execution, and different operation paths. The training phase also includes online search and feedback from multiple parsers, and the search overhead is also worth reporting.

Questions for authors

See Weakness

1 comment

Thank you for your constructive comments and positive assessment of the motivation and transfer results. We address your four concerns below.

Q1 – Individual contributions of the proposed mechanisms

Thank you for suggesting controls that separate the roles of these mechanisms. We have completed ablations of adaptive degradation sampling and multi-step teacher search. The completed ablations on Wild-OmniDocBench are reported below.

SettingPaddleOCR-VL-1.6MinerU2.5-ProUnlimited-OCROvisOCR2Avg
Without adaptive degradation sampling92.4391.9187.8691.2890.87
Single-step teacher instead of multi-step teacher92.1891.7487.2691.0690.56
Full AgenticDoc92.7092.0888.2791.6791.18

Both components improve all four parsers in these comparisons. The single-step teacher retains candidate-action evaluation; this control measures the additional value of considering subsequent enhancement steps.

For historical-image revisiting, General Response R2 and Appendix E, Table 10 report the completed comparison: on the 14 revisiting pages, AgenticDoc achieves 12 wins, 2 ties, and a mean paired gain of 2.12pp over the best result obtainable by retaining the pre-revisit path and continuing only from the latest image within the same remaining budget.

Q2 – Transfer, gain distributions, and visual changes

We appreciate your request for a closer examination of transfer and visual changes. Across the 13 parsers on Wild-OmniDocBench, gains have a median of +6.07 points and a range of +2.46 to +15.93. The mean gains are +6.50 for the four feedback parsers and +7.44 for the nine parsers not used for training feedback. These results support transfer beyond the specific feedback models. If you intended a finer-grained distribution, such as document-level gains within each parser, we would appreciate that clarification.

The qualitative cases in Appendix F illustrate clearer character strokes after sharpening, improved visibility after tone adjustment, and corrected page geometry through dewarping and resizing, alongside gains in the four-parser average score. Attributing cross-parser gains to specific visual changes is more challenging: each operation can affect multiple visual properties, successive operations can interact, and different parsers may respond differently to the same change. Tool-selection statistics alone are therefore insufficient to disentangle these effects. We are exploring more targeted analyses and will make every effort to provide further evidence and update the discussion with any additional findings.

Q3 – Regressions hidden by averages

Thank you for drawing attention to the individual regressions. Our clean-document preservation statement refers to aggregate performance. Tables 1 and 3 report the individual changes, including negative gains.

On clean OmniDocBench v1.6, seven parsers improve, one is unchanged, and five decline, with a median change of +0.01. On PureDocBench, three, two, and one parsers decline in the clean, digitally degraded, and real-world degraded subsets, respectively. These results show that the policy does not benefit every parser–condition combination, even when the aggregate change is positive.

Q4 – Inference efficiency and training/search cost

We are happy to clarify the inference cost and training-time search budget. General Response R1 distinguishes enhancement-action counts from AgenticDoc runtime and reports the completed enhancement-plus-parsing measurements. Across the four parsers, the net processing-time change is +0.54 s per image on clean OmniDocBench v1.6 and −12.10 on Wild-OmniDocBench. Measurement settings and repetition analysis are provided in Section 4.2.2.

For teacher construction during training, we use a maximum search depth of 3 and a beam width of 3. Search evaluates up to six first-level candidates and retains at most three nodes per level. Each retained node has at most five successors because consecutive repetitions of the same action are excluded. A teacher query therefore evaluates at most 6 + 3 × 5 + 3 × 5 = 36 candidate images before cache reuse.

These counts are upper bounds per teacher query before cache reuse. Multi-level caching reuses intermediate results and reduces redundant computation.

We hope these clarifications and additional results address your concerns and provide a clearer understanding of our contributions. If any questions or concerns remain, we would appreciate the opportunity to address them and would be happy to respond promptly.

Reviewer_1WCBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AgenticDoc, an agent-based framework for image enhancement prior to document parsing. The agent selects enhancement tools and parameters based on the current image and the image-action history, determines when to stop, and passes the final image to a frozen parser. During training, the framework generates samples through online degradation, optimizes the enhancement policy using parsing feedback, and introduces A2T-OPD to convert the action preferences obtained from the search into conditional token supervision, combining distillation with PPO.

The paper compares parsing performance before and after enhancement on Wild-OmniDocBench, OmniDocBench v1.6, and PureDocBench. All 13 parsers on Wild-OmniDocBench showed improvements; nine of these parsers were not used to provide training feedback. Overall performance on OmniDocBench v1.6 remained largely unchanged, and average gains were achieved under clean, digitally degraded, and real-world degraded conditions on PureDocBench; however, some combinations of models and conditions still experienced performance declines.

Critical strengths
  1. The paper decouples the image enhancement policy from downstream parsers, utilizing parsing feedback to learn enhancement decisions while keeping the parsers frozen. This design allows the enhancement capabilities to be reused across different parsers without the need to retrain each parser, offering greater flexibility.

  2. In experiments, the enhancement policy delivered clear gains for degraded documents while generally maintaining the parsing performance on clean documents. As shown in Table 1, all 13 parsers achieved improvements on Wild-OmniDocBench, with an average gain of 7.15 points. The policy comparison in Table 3 demonstrates that the trained agent outperforms the tested fixed-action and GPT-5.4 policies, supporting the value of learning enhancement decisions specifically for parsing tasks. In Table 4, joint training also outperforms PPO alone and Search-SFT, providing further support for improving policy training by leveraging action preferences obtained through search.

Critical weaknesses
  1. The paper does not sufficiently discuss existing image enhancement and restoration work, and its incremental contributions are not clear enough. The paper does not discuss related work such as DocRes and AgenticIR. Tables 1 and 2 primarily compare the parsing results before and after enhancement; although Table 3 includes fixed operations and the GPT-5.4 policy, it does not directly compare the method with existing enhancement or restoration methods. Therefore, while the current results support the finding that the proposed method outperforms the selected baselines, its advantages over existing enhancement methods remain unverified. If the authors supplement the paper with direct comparisons against existing enhancement or restoration methods and can demonstrate the additional benefits of the proposed method, I will reconsider my current rating.

  2. The paper repeatedly emphasizes preserving historical images, reverting to previous states, and exploring alternative paths; however, the current ablation studies do not compare the proposed method with adaptive policies that can only apply each operation to the image produced in the previous step. Tables 3 and 4 compare policy backends and training methods, respectively, but do not separately examine the role of selecting historical image states. The analysis in Appendix B indicates that the order of operations affects the results; however, the order can be adaptively selected during the step-by-step processing and does not necessarily require rollback. Therefore, the additional benefits of rollback to earlier image states relative to general adaptive enhancement have yet to be directly validated.

Questions for authors
  1. Section 3.2 and Appendix D describe the process of selecting and constructing the training data, but do not list specific sources. Please specify which datasets and versions the 1,000 source document images come from, as well as the number of images from each source and the training-validation split.

  2. On OmniDocBench v1.6, an average of 0.85 enhancement operations are still performed per image, yet the overall parsing performance remains largely unchanged. In addition to enhancement operations, the agent must also determine whether to continue processing and when to output "parse". Please specify which steps are included in the 1.13 seconds reported in the paper, and explain the increase in processing time relative to the original OCR pipeline after accounting for all agent calls and enhancement processing. For real-world input consisting of a mix of clean and degraded documents, how do the authors balance parsing gains against additional latency?

  3. Section 3.1.1 states that "parse" returns the current image, while the caption for Figure 2 describes the ability to submit a selected historical image. Please clarify the actual execution rule: when stopping, can only the most recently generated image be submitted, or can any historical image be selected directly?

1 comment

Thank you for your constructive comments and for recognizing the value of a reusable enhancement policy. We address your concerns below.

W1 – Related work and restoration baselines

Thank you for highlighting relevant restoration work and recommending direct comparisons with existing enhancement methods. Following your suggestions, we have expanded Section 2.3 to discuss DocRes, AdaptiveISP, RL-Restore, and AgenticIR, and completed direct comparisons with DocRes and AdaptiveISP. The settings, results, and visual analysis are presented in General Response R3. On Wild-OmniDocBench, AgenticDoc exceeds these two baselines by 11.46 and 10.42 points, respectively, averaged across the four parsers, while largely preserving clean-document performance.

W2 – Contribution of historical-image selection

We appreciate your suggestion to directly evaluate the contribution of historical-image revisiting. We have completed the targeted comparison described in General Response R2 and Appendix E, Table 10. For each of the 14 revisiting cases, we retain the enhancement path before revisiting and find the best result obtainable by continuing to process only the latest image within the same remaining budget, including immediate parse.

AgenticDoc achieves 12 wins and 2 ties, with a mean paired gain of 2.12pp on these 14 pages. This comparison provides evidence of recovery from unfavorable enhancement decisions under the same remaining budget.

Q1 – Training-data sources and evaluation protocol

We are happy to clarify the data sources, pseudo-reference construction, and evaluation protocol. The training collection contains 1,000 source images after agreement filtering: 300 each from DocLayNet (official 1.0.0 release), PubLayNet (original public release, IBM DAX 1.0.0), and DocStruct4M (official 2024 release accompanying mPLUG-DocOwl 1.5), and 100 from publicly accessible web documents, after agreement filtering. We use only the images, without their original annotations. As described in Section 3.2 and Appendix D.1, we construct OCR-derived pseudo-references and retain clean pages when MinerU2.5-Pro and PaddleOCR-VL-1.6 outputs have NED < 0.02. Agreement is a practical filtering criterion; shared errors may remain.

We construct a separate validation set containing 500 images after agreement filtering, using the same data sources and proportions as the training set. We apply compound degradations once to each clean image, producing a fixed set of 500 degraded validation inputs that are reused unchanged for all checkpoint comparisons. We select the checkpoint based on its performance on this validation set and evaluate it on Wild-OmniDocBench, OmniDocBench v1.6, and PureDocBench. Images from these three benchmarks are not included in the training or validation collections. We will publicly release the training and validation datasets and code to support reproducibility.

Q2 – Runtime measurement and overall processing overhead

Thank you for requesting a clearer account of the runtime measurements and overall processing overhead. As clarified in General Response R1, the reported 1.13 s includes all executed AgenticDoc policy-model calls, explicit parse generation when present, tool execution, and final-image output; downstream parsing is additional. Enhancement-action counts exclude parse.

The added enhancement-plus-parsing measurements show an average net increase of 0.54 s per image on clean OmniDocBench v1.6 and 12.10 s decrease per image on Wild-OmniDocBench, across the four parsers. For mixed workloads, the trade-off depends on the proportion of degraded documents and the downstream parser; these results quantify both the clean-input overhead and the benefits on degraded inputs. Details are provided in Section 4.2.2.

Q3 – Execution of the parsing action

We appreciate the opportunity to clarify which image the parse action submits. The parse action submits only the current image. Historical images can be selected as inputs to enhancement tools, and a successful tool output becomes the current image. We have corrected the Figure 2 and its caption to make this execution rule consistent with Section 3.1.1.

We hope these clarifications and additional results address your concerns and provide a clearer understanding of our contributions. If any questions or concerns remain, we would appreciate the opportunity to address them and would be happy to respond promptly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback. We have revised the manuscript and added the following evidence addressing shared concerns.

R1 – Processing time and inference overhead

We thank the reviewers for requesting clarification of the runtime measurements and overall processing overhead. The reported 0.85 enhancement actions per image on OmniDocBench v1.6 exclude parse. The 1.13 s AgenticDoc runtime includes all executed policy-model calls, including explicit parse generation when present, tool execution, and final-image output; downstream parsing is additional.

We have clarified this distinction and added overall processing-time results in Section 4.2.2. The +AD times combine mean AgenticDoc runtime with parsing time on enhanced images. We measure parser runtimes by processing images sequentially on a single NVIDIA A100-80GB GPU, using the official inference configurations (vLLM for OvisOCR2 and Transformers for the other three parsers).

We acknowledge the additional processing overhead on clean OmniDocBench v1.6: enhancement plus parsing takes 0.54 s more per image on average across the four parsers. On Wild-OmniDocBench, the real-world degraded setting we primarily target, the end-to-end processing time decreases by 12.10 s on average with AgenticDoc. Inspection of the parser outputs revealed fewer lengthy repetitive generations after enhancement, helping explain the shorter decoding time: repetition-flagged final outputs decrease from 222 to 95 (−57.21%). We have added this analysis in Section 4.2.2 and would be happy to include illustrative cases in the appendix if helpful.

R2 – Value of historical-image revisiting

We appreciate the suggestion to directly assess the contribution of historical-image revisiting. We have added a targeted comparison in Appendix E, Table 10.

Revisiting occurs on 14/1,350 Wild-OmniDocBench pages (1.04%). In all 14 cases, the agent returns to the original image after the first enhancement. For comparison, we retain the enhancement path immediately before revisiting. We then find the best result within the remaining budget by exhaustively evaluating all legal ways to continue processing only the latest image, including immediate parse.

Compared with the best such sequential continuation, AgenticDoc achieves 12 wins, 2 ties, with a mean paired gain of 2.12pp on these 14 pages, using the four-parser average score. This provides evidence that revisiting can recover from an unfavorable enhancement, although it is used infrequently.

R3 – Related work and restoration baselines

Following the reviewers’ suggestions, we have expanded the Related Work section (Section 2.3) to include citations and discussion of DocRes, AdaptiveISP, RL-Restore, and AgenticIR, and clarified how AgenticDoc differs from these methods.

For direct comparison, we evaluated DocRes and AdaptiveISP using their official checkpoints and processing pipelines on Wild-OmniDocBench and OmniDocBench v1.6, with identical inputs, full-resolution outputs, and the same four frozen parsers.

Enhancement methodWild-OmniDocBenchOmniDocBench v1.6
PaddleOCR-VL-1.6MinerU2.5-ProUnlimited-OCROvisOCR2PaddleOCR-VL-1.6MinerU2.5-ProUnlimited-OCROvisOCR2
Original87.3687.3376.1287.9196.3495.7594.0096.47
DocRes82.8582.2671.2582.5176.4576.1271.6376.72
AdaptiveISP83.0783.0872.9483.9792.2591.7290.7592.84
AgenticDoc92.7092.0888.2791.6796.2995.6393.7796.67

AgenticDoc outperforms both baselines on all four parsers across both benchmarks. Relative to the original inputs, the four-parser average changes on Wild are −4.96 points for DocRes, −3.92 for AdaptiveISP, and +6.50 for AgenticDoc.

The decline is especially pronounced for DocRes on clean OmniDocBench: its average falls by 20.41 points to 75.23, even below its Wild average (79.72). Inspection revealed cases where DocRes introduces geometric distortion and truncates originally clean pages. Such truncation causes irreversible content loss in the processed images and helps explain the parsing regression. We provide visual case studies in the supplementary material.

Together, these results support the value of enhancement guided by parsing utility and preservation of already-readable content. AgenticDoc improves degraded-document parsing while keeping its clean-document average close to Original (95.59 vs. 95.64 for four-parser average).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.