acceptodds
Withdrawn from ICLR 2027

ELDER: LEARNING COMPACT EVIDENCE FOR LONG- DOCUMENT VISUAL REASONING

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-document visual reasoning is constrained by visual-token budgets, as processing full pages allocates substantial input capacity to content unrelated to the question. To address this problem, we introduce Evidence Learning for Document-Element Reasoning (ELDER), a dual-role framework that reduces visual-token input for answer generation while preserving the internal layout of selected evidence. A shared vision-language model selects complete document elements and answers using only their original crops. Supervised fine-tuning uses families of verified evidence sets to initialize both roles, teaching multiple supporting combinations for the same question instead of a single canonical selection. We further formulate Minimal Reliable Evidence Group Relative Policy Optimization (MRE-GRPO) to compare alternative evidence sets by the mean quality of multiple answers generated from each set and its visual-token cost. This set-level comparison can favor complementary combinations and useful corroboration when their answer-quality gains justify the added tokens, and discourage extra content that does not improve sampled answers. A separate answer objective improves reasoning under each fixed set, coupling evidence compression with learning to use retained content. Across five benchmarks, ELDER achieves an average score of 67.6, exceeding the strongest external method with results on all five by 2.3 score points. Compared with answer-only reinforcement learning, ELDER reduces mean answer-input visual tokens by 43.2%, with a 0.3-point decrease in average score.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Document Visual Reasoning

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

This paper studies evidence selection and answer-input compression for long-document visual question answering. It introduces ELDER, a dual-role framework built on a shared vision-language model. An Evidence Extractor selects complete, pre-parsed document elements, and an Evidence-only Generator answers using only their original image crops. Teacher-based support checks and a reader that does not receive reference answers construct families of acceptable evidence sets for supervised initialization. The proposed MRE-GRPO uses nested sampling to evaluate evidence sets through mean answer scores and visual-token costs, while separately optimizing answers conditioned on fixed evidence. Across five benchmarks, the paper reports an average score of 67.6. Relative to dual-role RL without the visual-token cost penalty, ELDER reduces mean Generator input visual tokens by 43.2%, with a 0.3-point decrease in average score. Role-loss ablations, full-document training controls, and frozen-checkpoint combinations further examine the contributions of selection and answering.

Critical strengths

S1. The method addresses a practically relevant quality–input-cost tradeoff with a coherent training design. Selecting complete document elements, preserving their internal layouts, and restricting answering to the selected crops is a reasonable approach to reducing irrelevant visual context in long documents. MRE-GRPO distinguishes quality–cost comparisons across evidence sets from answer optimization under fixed evidence, and explicitly handles invalid selections, empty sets, and all-zero feedback. This provides a clear, operational approach to jointly learning evidence selection and evidence use. [Sections 3.3–3.5 and Appendix A.3] S2. The internal controls provide meaningful evidence for the overall framework’s empirical value. Beyond the five-benchmark results and role-loss ablations, Table 3 includes full-document controls matched by data progress or training GPU-hours, as well as combinations of frozen Extractor and Generator checkpoints. Full ELDER exceeds the strongest full-document control in Table 3 by 4.0 points on average. With either role fixed, replacing the other with its RL checkpoint still improves scores. These experiments support contributions from both evidence selection and evidence-conditioned answering, and strengthen the case beyond simply applying additional training. [Tables 1–3 and Section 4.4]

Critical weaknesses

W1. Important experimental details are missing, limiting the verification and reproducibility of the central quality–cost results. The paper defines the visual budget BB, controller step size ηλ\eta_\lambda, calibrated target τ\tau, and fixed image preprocessing, but does not report the concrete values or settings needed to reproduce the experiments. Training sampling and evaluation decoding configurations are also insufficiently specified. These choices directly determine the executable selection space, visual-token counts, and the scale of the cost penalty. Table 4 reports question–evidence-set pairs without the actual numbers of unique questions, role-specific training records, or active RL questions, while the compute-matched controls in Table 3 lack numerical GPU-hour accounting. External baseline results are also not individually mapped to imported or reproduced sources and evaluation settings. These omissions make the central quality–cost comparisons and their fairness difficult to verify fully. [Section 4.1 and Appendices A.1–A.3, C–D] W2. The distinctive design choices and the support quality of the selected evidence are not sufficiently validated. Tables 2–3 establish benefits from the overall framework, role losses, and cost term, but do not directly isolate the gains from multiple-set supervision or nested set evaluation over simpler alternatives. Evaluation also relies primarily on answer scores and visual-token counts, without an independent assessment of whether the selected regions support the answers. Figure 6 illustrates the reference answer and evidence structure rather than reporting model predictions and selection traces. Although the authors explicitly acknowledge that answer correctness does not establish evidence sufficiency, this empirical gap still limits our understanding of what the model learns about evidence. Moreover, adaptive control improves the average score over fixed cost by approximately 0.3 points, without uncertainty estimates across repeated runs, leaving the necessity of the additional mechanism unclear. [Tables 2–3 and Appendices A.1, A.3, F–G]

Questions for authors

1.Q1. Experimental protocol and reproducibility Could you provide the complete configuration and data statistics, including BB, ηλ\eta_\lambda, the calibrated τ\tau, page and crop preprocessing, training sampling and evaluation decoding settings, the actual numbers of unique SFT/RL questions and documents, and the evaluation split and sample count for each benchmark? What are the numerical training GPU-hours for the methods in Table 3? Which results in Table 1 are imported versus reproduced, and are their evaluation settings comparable? This information would directly affect my confidence in the main comparisons. 2.Q2. Independent contribution of multiple-set supervision When source questions and training exposure are kept comparable, how does using only one verified evidence set per question compare with the current multiple-set supervision? A controlled comparison on a representative benchmark would help determine whether multiple-set supervision provides value beyond additional training exposure. 3.Q3. Stability of the adaptive controller’s gains Is the small improvement over fixed α=1\alpha=1 consistent across repeated runs? Could you provide repeated-run results for a representative setting, together with paired per-example uncertainty analysis and the cost-coefficient trajectory, to clarify the controller’s actual effect? This would affect my assessment of whether the additional control mechanism is necessary.

Reviewer_oR62OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ELDER, which employs a shared VLM to perform both document evidence selection and evidence-based answer generation. The system first parses document elements using MinerU; the Extractor selects indices of complete regions, after which the Generator answers solely based on the corresponding raw cropped images.

MRE-GRPO samples multiple evidence sets for each question, and then generates multiple answers for each set. The Extractor obtains rewards based on the average answer score and the cost of visual tokens; the Generator compares the quality of answers within the same evidence set. The method further incorporates an adaptive cost coefficient, a mechanism to disable cost punishment in the absence of positive feedback, and handling for invalid outputs.

Critical strengths
  1. The overall framework is supported by meaningful same-backbone controls. Beyond comparisons with the untrained backbone, Table 3 includes full-document SFT and SFT+GRPO baselines under reported data-exposure and training-compute matching.

  2. The frozen-checkpoint experiments provide evidence that both evidence selection and evidence-conditioned answering improve. Table 3 evaluates combinations of SFT and RL checkpoints while reusing the same selected crops across Generators for each Extractor. This analysis supports improvements in both capabilities rather than attributing the gains solely to a stronger answer generator.

Critical weaknesses
  1. Key experimental settings and dataset sizes are insufficiently specified, limiting reproducibility and the assessment of comparison fairness. The visual-token budget B and controller step size η_λ in Eqs. (2)–(4) are defined, but their actual values are not reported. These omissions reduce confidence in the reproducibility and fairness of the reported comparisons.

  2. The sources and evaluation protocols for external baselines are not sufficiently documented on a per-method basis. While these comparisons provide useful context, they do not establish that the observed gains are attributable to the proposed method.

  3. Several key ablation differences are small, but no means, standard deviations, or confidence intervals across repeated training runs are reported. It therefore remains unclear whether these differences are consistent across runs.

Questions for authors
  1. Could you provide the complete training and input-processing configurations? Please specify B, η_λ, the calibrated value of τ, image resolution, the Extractor’s context length, how inputs exceeding this limit are handled, and the decoding settings used during training and inference.

  2. Does the adaptive strategy still outperform fixed coefficients at matched token costs? Could you evaluate multiple fixed values of α, report the quality–cost curves, and repeat training for the most closely performing configurations?

Reviewer_NUwjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

ELDER addresses long-document visual question answering through a shared vision-language model with two roles: an Evidence Extractor selects indexed document elements, and an Evidence-only Generator answers from their original crops. Teacher-verified evidence sets initialize both roles through supervised fine-tuning. The proposed MRE-GRPO subsequently samples multiple evidence sets and multiple answers per set, optimizing selection through mean answer quality minus visual-token cost and optimizing generation through within-set answer comparisons. Experiments on five benchmarks show improvements over the listed external baselines. Compared with answer-only dual-role RL, ELDER reduces mean Generator input visual tokens by 43.2%, with a 0.3-point decrease in the average task score.

Critical strengths
  1. The paper provides meaningful experimental controls supporting the overall framework. Table 3 includes full-document baselines matched by training-data progress or cumulative GPU-hours, and ELDER outperforms the strongest reported full-document control across all five benchmarks. Frozen Extractor/Generator checkpoint combinations further show that improved selection benefits a fixed reader and improved reading benefits fixed evidence. These comparisons strengthen the attribution beyond simply comparing the trained pipeline with an untrained backbone.

  2. The optimization design explicitly accounts for several important practical issues, including invalid selections, insufficient-evidence abstention, visual-token costs, and the distinction between selection and answer-generation feedback. The authors also appropriately distinguish answer correctness from semantic evidence sufficiency and Generator input compression from total system efficiency. Appendix E reports end-to-end latency for both first processing and document reuse, acknowledging the substantial preprocessing overhead.

Critical weaknesses
  1. The incremental contribution over related evidence-selection methods and simpler training objectives is not sufficiently established.

The closest region-selection approaches are discussed but not experimentally compared. Doc-CoB [1] already uses layout-derived indexed boxes followed by focused answering, although it retains global context through visual masking. AgenticOCR [2] performs query-driven region extraction and provides compact evidence to downstream generation. These are particularly relevant comparisons for assessing the value of ELDER’s evidence interface. The broader discussion of learned document interaction is also incomplete: VRAG-RL [3] learns retrieval and fine-grained visual operations, while ALDEN [4] trains active navigation and evidence-gathering policies. Neither is discussed. These precedents do not invalidate ELDER’s specific set-level quality–cost objective, but the paper should more clearly establish its relationship to existing RL-based evidence acquisition and include a representative region-selection comparison under comparable settings.

The internal ablations establish the usefulness of joint selection and answering more clearly than the necessity of the distinctive training components. The paper does not isolate evidence-family supervision against a single accepted set per question under matched exposure and training budget. It also does not establish the benefit of nested sampling and its two grouping levels against a simpler trajectory-level objective with comparable rollout compute. Finally, the adaptive controller is compared only with a fixed coefficient of α = 1, rather than a tuned range of fixed coefficients. The full method improves the average score over this fixed-cost setting by only 0.3 points while using slightly more visual tokens. Without uncertainty estimates or a quality–cost frontier, this does not establish a robust advantage for adaptive control.

  1. The empirical characterization of evidence reliability and the information needed to reproduce the quality–cost tradeoff remain incomplete.

Answer agreement and Generator input size demonstrate a performance–compression tradeoff, but do not directly show whether selected sets preserve the facts and relations required to support their answers. The offline support checks are teacher pseudolabels, and online RL rewards answer quality rather than independently assessed support. The manuscript acknowledges this distinction, but does not quantify how often the learned selector retains complementary evidence, benefits from corroboration, or omits necessary context. A human-audited subset or controlled evidence-removal/substitution analysis would strengthen the central evidence-learning claim, particularly for cross-page and multi-element questions.

Despite the extensive appendix, several consequential settings remain unspecified: the numerical visual-token budget B, controller step size ηλ, calibrated target τ, exact page/crop image-processing settings, and rollout and deployment decoding configurations. The reported dataset statistics count question–evidence-set pairs, while the actual numbers of unique questions, role-specific SFT records, and admitted RL questions are not provided. These quantities affect both reproducibility and the interpretation of training-data scale.

The efficiency analysis is useful but incomplete. Figure 5 explicitly shows that first processing is substantially slower and that document reuse changes the comparison. However, the main token metric excludes Extractor input, even though this role processes indexed candidate pages and their region inventory. It remains unclear how Extractor inputs are bounded for very long documents. Actual training GPU-hours, inference hardware and batching, and token/latency breakdowns for both roles would make the reported tradeoffs and compute-matched controls independently assessable.

  1. The presentation makes the central method unnecessarily difficult to follow.

Section 3.2 compresses region proposal, evidence-set construction, support verification, answer validation, repair, and training-data export into a dense paragraph. Terms such as “base sets” and “verified pairs” appear without sufficient explanation. The main text should clearly distinguish the two reference-aware support checks from the answer-blind reader and explain how their outcomes determine Extractor SFT, Generator SFT, and question-level RL records.

The opening paragraph of Section 3.5 similarly introduces nested sampling, policy snapshots, set-level evaluation, and online supervision before clearly explaining the central distinction: comparing evidence sets to train selection versus comparing answers under fixed evidence to train generation. A small running example would make this substantially easier to understand.

Figure 3 adds further notation overhead. Its caption maps a_ij, y*, and R_i to different symbols used in the text, distinguishes its valid-branch notation from V(E), and introduces a maximization objective J and scores s despite the loss-minimization formulation in the main text. The figure should use the same notation and optimization convention as Eqs. 2–6, allowing the caption to explain the workflow rather than reconcile incompatible presentations.

Minor presentation issues:

  • At line 050, “element-level evidence selection” would describe the selection granularity more precisely than “evidence-level reasoning.” Evidence denotes a functional role and can exist at chunk, page, or element granularity.
  • Parenthetical citations are repeatedly rendered as narrative citations, producing constructions such as “GRPO Shao et al. (2024) uses …” at line 145. The manuscript should systematically distinguish narrative citations from parenthetical citations and correct the formatting.

[1] Mo et al. “Doc-CoB: Enhancing Multi-Modal Document Understanding with Visual Chain-of-Boxes Reasoning.” 2025. arXiv:2505.18603.

[2] Wang et al. “AgenticOCR: Parsing Only What You Need for Efficient Retrieval-Augmented Generation.” 2026. arXiv:2602.24134.

[3] Wang et al. “VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning.” NeurIPS, 2025. arXiv:2505.22019.

[4] Yang et al. “ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents.” 2025. arXiv:2510.25668.

Questions for authors
  1. Can the authors compare ELDER with a representative region-selection approach, such as an appropriately adapted Doc-CoB or AgenticOCR, under comparable backbone and input settings? Please also clarify its relationship to learned evidence-acquisition methods such as VRAG-RL and ALDEN.

  2. What are the independent contributions of evidence-family supervision and nested sampling? A matched-exposure single-set SFT control and a comparable-compute alternative to the nested grouping scheme would help establish their necessity.

  3. Does adaptive cost control improve the quality–cost frontier over tuned fixed coefficients? Please report τ, ηλ, the evolution of λ or α during training, and uncertainty estimates for the small differences between these variants.

  4. How frequently do selected sets provide complete support for correct answers? Can a human-audited subset or controlled evidence perturbation study distinguish supported reasoning from answers obtained despite missing evidence, especially for cross-page and multi-element questions?

  5. Please provide the numerical budget B, image-processing and decoding settings, unique-question and role-export counts, and the final active RL pool size. How are very long Extractor inputs handled? Please also report the actual GPU-hours underlying Table 3 and the inference hardware, batching, and stage-level timings underlying Figure 5.

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

Related papers

Discussion (0)

Sign in to comment.