acceptodds
Under review as a conference paper at ICLR 2027

Think in Rows and Columns: Decoupled Latent Visual Reasoning for Multimodal Tables

Abstract

Multimodal table reasoning requires Large Vision Language Models to jointly recognize fine-grained cell content and reason over structural relations across rows, columns, and headers. Existing methods either generate answers directly or externalize intermediate computation through textual traces, programs, and image operations. While latent visual reasoning offers a more efficient alternative, current methods lack explicit organization according to table structure. To this end, we propose **Decoupled Latent Visual Reasoning (DLVR)**, which organizes continuous latent computation into progressive row, column, and subtable stages while separating structural organization from semantic content. DLVR is trained through decoupled latent supervised fine-tuning followed by **Latent Intervention GRPO**, which employs a **Latent Contribution Reward** to promote the predictive contribution of generated latent components. To supervise this process, we further construct **TableLVR-41K**, a corpus with structure-guided reasoning routes and stage-specific visual and content annotations. Experiments with Qwen2.5-VL and Qwen3-VL across seven benchmarks demonstrate consistent improvements over generic latent reasoning, and explicit table reasoning methods. Ablation studies further validate the complementary contributions of structure-content decoupling, latent supervision objectives, and intervention-based optimization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal table reasoning, latent visual reasoning, large vision-language models

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_NjyKOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Decoupled Latent Visual Reasoning (DLVR) for multimodal table reasoning. DLVR organizes latent computation into progressive row, column, and subtable stages, while separating structural and content representations within each stage. The model is trained using a two-stage framework consisting of decoupled latent supervised fine-tuning followed by Latent Intervention GRPO with a Latent Contribution Reward. The authors also construct TableLVR-41K to provide stage-specific supervision. Experiments across seven benchmarks and two LVLM backbones show improvements over direct answering, explicit reasoning methods, and generic latent visual reasoning baselines, together with substantially reduced inference time and generation length.

Critical strengths
  • DLVR provides an effective and computationally efficient alternative to explicit multimodal reasoning. Across two LVLM backbones, the method achieves strong performance on a broad set of table reasoning benchmarks while requiring substantially fewer generated tokens and lower inference time than textual CoT and explicit image-based reasoning methods. This combination of reasoning performance and inference efficiency makes the approach particularly relevant for practical deployment.

  • The methodological design is well aligned with the structural characteristics of table reasoning. Rather than directly applying generic latent visual reasoning or GRPO, the paper introduces row, column, and subtable latent stages, structure-content decoupling, stage-specific latent supervision, and an intervention-based reward that encourages the generated latent states to contribute to answer prediction. The matched baselines and component ablations provide empirical support for these table-specific design choices.

Critical weaknesses

None

Questions for authors
  • Why is fusion reconstruction performed only on the average-pooled stage representation, rather than preserving a finer-grained correspondence across the KK latent states? Would token-level or position-wise reconstruction provide additional supervision for the relational structure encoded within each stage?

  • Could the authors clarify how PoolK\mathrm{Pool}_K is implemented for the visual and content targets in Eq. (3)? In particular, how is correspondence established between the resulting KK visual states and KK content states before computing Sm=Vm−P(Tm)S_m = V_m - P(T_m)?

3 comments
AuthorsOpenReview ↗

Response to Reviewer NjyK

Thank you for reviewing our paper and for the insightful comments. Our point-to-point responses to your comments are given below. We have clarified the implementation of Pool⁡K\operatorname{Pool}_K and the index-based correspondence between visual and content targets in Section 3.2.2. We have also added a position-wise reconstruction comparison to the ablation studies in Section 5 and Appendix C. All revisions to the manuscript are highlighted in blue for ease of reference. We hope these revisions address your questions, and we welcome any further suggestions.


Q1:Why is fusion reconstruction performed only on the average-pooled stage representation, rather than preserving a finer-grained correspondence across the KK latent states? Would token-level or position-wise reconstruction provide additional supervision for the relational structure encoded within each stage?

R1: We thank the reviewer for this question. We use average-pooled fusion reconstruction to recover stage-level visual evidence while allowing complementary information to be distributed across latent states. This design has three motivations:

(1) Reduced reliance on strict positional correspondence. Visual and content targets are pooled independently, and matching indices do not necessarily correspond to the same cells or visual regions. Stage-level reconstruction reduces reliance on this approximate correspondence rather than imposing an additional position-wise matching constraint.

(2) Flexible information allocation across latent states. The KK latent states represent sequential computation rather than KK predefined visual regions. Average-pooled reconstruction allows these states to collectively represent the evidence within each stage, without requiring every fused state to reproduce a prescribed visual target.

(3) Complementarity with content supervision. The content loss already provides position-wise supervision for content states. The fusion loss instead constrains the combined structural and content representation at the stage level, complementing content alignment without imposing the same granularity on both objectives.

Empirical comparison. To further investigate this design choice, we conduct an additional experiment replacing average-pooled fusion reconstruction with position-wise reconstruction, while keeping other settings unchanged. As shown in Table 1, average-pooled reconstruction achieves a higher average score, with particularly notable improvements on InfoTabs and PubHealthTab. These results suggest that imposing position-wise reconstruction does not necessarily provide more effective supervision, supporting our choice of stage-level reconstruction to preserve flexibility in latent information allocation.

Table 1: Comparison of average-pooled and position-wise fusion reconstruction.

ReconstructionWTQTAT-QAHiTabAIT-QATabFactInfoTabsPubHealthTabAvg.
Average-pooled59.9436.4043.1589.0478.3572.3076.8865.15
Position-wise59.2536.9243.1588.0678.3169.8370.1363.66

Q2: Could the authors clarify how PoolKPool_K is implemented for the visual and content targets in Eq. (3)? In particular, how is correspondence established between the resulting KK visual states and KK content states before computing Sm=Vm−P(Tm)S_m=V_m-P(T_m)?

Thank you for pointing out this ambiguity. We implement Pool⁡K\operatorname{Pool}_K using one-dimensional adaptive average pooling along the token dimension (i.e., torch.nn.functional.adaptive_avg_pool1d), producing KK representations from each input sequence. For visual targets, we preserve the visual encoder’s output token order. For content targets, we remove padding and preserve the token order of the serialized cell content. The two sequences are pooled independently.

The correspondence is therefore index-based rather than an explicit semantic alignment between cells and visual patches. Specifically, we compute $ S_{m,k} = V_{m,k}-P (T_{m,k} ) $ for $ k=1,...,K $, where the learnable projection PP maps content representations into the visual feature space but does not explicitly establish token-level alignment. Accordingly, SmS_m serves as an approximate structural proxy during teacher forcing, rather than a guaranteed content-free structural representation.

Action taken: We have clarified the implementation of Pool⁡K\operatorname{Pool}_K and the index-based correspondence between visual and content targets in Section 3.2.2, and added the position-wise fusion reconstruction comparison and an analysis of the flexibility provided by average-pooled reconstruction in Section 5 and Appendix C (Tables 2 and 9).

Thanks for the detailed clarifications. My questions have been adequately addressed, and I will maintain my positive rating.

AuthorsOpenReview ↗

Response to Reviewer NjyK

Thank you for your kind follow-up and for confirming that your questions have been adequately addressed. We sincerely appreciate your continued positive assessment of our work.

We are particularly grateful for your recognition of the following aspects:

  • Reasoning effectiveness and inference efficiency: DLVR achieves strong performance across diverse table reasoning benchmarks while substantially reducing generated tokens and inference time compared with explicit reasoning methods.
  • Table-specific methodological design: Our row, column, and subtable latent stages, structure-content decoupling, stage-specific supervision, and intervention-based reward are tailored to the structural characteristics of table reasoning, with matched baselines and ablations supporting their effectiveness.

We are especially encouraged by your recognition of the practical efficiency and methodological contributions of DLVR, which are central motivations of this work.

Thank you again for your thoughtful feedback, constructive suggestions, and support.

Reviewer_38wNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Decoupled Latent Visual Reasoning (DLVR) for reasoning over table images. DLVR organizes the latent trajectory into row, column, and subtable stages, and each stage has separate structure and content latents. Training has two phases. The first is decoupled latent supervised fine-tuning. The second is Latent Intervention Group Relative Policy Optimization (LI-GRPO), which pushes the model to actually use its latents through a Latent Contribution Reward: the drop in answer log-likelihood when one latent block per stage is replaced by its control-token embedding. The authors also release TableLVR-41K, 41.6K question-answer pairs that GPT-5.4-mini generated from the training splits of five table benchmarks, with stage-wise evidence annotations. Across seven benchmarks and two backbones (Qwen2.5-VL-7B and Qwen3-VL-8B), DLVR outperforms general and table-tuned large vision-language models, explicit reasoning methods, and other latent reasoning baselines.

Critical strengths

The method is well motivated, and adapting latent visual reasoning to the row/column structure of tables is an interesting direction. Competitive accuracy is obtained at lower inference cost than explicit table reasoning. I like the idea of using a reward based on how much the latents actually contribute, though counterfactual intervention, rather than only whether the answer is correct. Overall, the paper is easy to follow, and the figures explain the pipeline clearly.

Critical weaknesses
  1. Narrow scope for ICLR. The method is built around tables: row, column and subtable stages are fixed in the architecture, and training needs stage-wise annotations derived from the underlying structured table. The paper does not show or discuss whether the approach carries over to other structured visual inputs or to general visual reasoning, which limits its interest to the broader ICLR audience.

  2. Limited novelty. The idea of structure-content decoupling for tables has been explored in DiSCo [Zhu et al, 2026] (used as baseline) and the latent contribution reward seems incremental over LUT [Kang et al, 2026]

  3. Small gains and weak baselines. DLVR improves only slightly over LVR-Random, which uses randomly sampled regions as latent targets. This suggests the table-specific structure contributes little. The only closed-source baseline, GPT-4o-mini, has poor scores (e.g., below chance on TabFact), which points to an evaluation or answer-parsing issue. The paper should add stronger proprietary models (e.g., recent versions of Claude, GPT, Gemini) as reference baselines

Questions for authors

The paper has poor results for GPT-4o-mini. Can you provide results for recent versions of proprietary models such as Claude/GPT/Gemini?

Can you make it more clear how the novelty claims contrast with prior work?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 38wN (1/3)

Thank you very much for your thoughtful response and for taking the time to review our rebuttal. Our responses below clarify the scope of DLVR and its methodological distinctions from prior work. We have clarified the main contributions of DLVR and its distinctions from prior table reasoning and latent visual reasoning methods in Section 2, strengthened the empirical evaluation by adding GPT-6 Luna to Table 1 and reporting three-seed comparisons with LVR-Random in Section 5 and Appendix C. All revisions to the manuscript are highlighted in blue for ease of reference. We hope these additions and clarifications address your concerns. Please feel free to let us know if you have any additional concerns or questions.


W1: Narrow scope for ICLR.

R1: We thank the reviewer for raising this concern. We would like to clarify the broader relevance of our work and the generalizability of DLVR.

(1) Relevance to ICLR. Table understanding and reasoning have been actively studied at ICLR, including Chain-of-Table (ICLR 2024), CABINET (ICLR 2024), TnT (ICLR 2025), TABLET (ICLR 2026), and TaTToo (ICLR 2026). Several of these works explicitly leverage table-specific structural properties to design specialized representations or reasoning mechanisms. Similarly, our work uses tables as a challenging setting to investigate a broader representation-learning question: how to organize, supervise, and optimize continuous latent reasoning over structured visual evidence.

(2) Generalizability of the framework. While the row, column, and subtable stages are tailored to table structures, the underlying methodology is not inherently restricted to tables. Specifically, DLVR introduces structure-content decoupling within intermediate latent blocks, jointly optimizes them through content alignment, fusion reconstruction, and orthogonality regularization, and employs intervention-based rewards to encourage their contribution to answer prediction. These mechanisms could potentially extend to other structured visual inputs by adapting the reasoning stages to their structural properties.

(3) Use of Stage-wise Annotations. We clarify that the stage-wise annotations derived from structured tables are used only for supervision during the initial SFT stage. LI-GRPO does not require these annotations, and inference operates directly on the original table image and question, without access to structured table representations, stage-wise annotations, or external image operations.

(4) Generalization Across LVLMs and Benchmarks. We evaluate DLVR across 7 benchmarks and 2 LVLM backbones. The consistent improvements across different benchmarks and backbones demonstrate the effectiveness and generalizability of DLVR.

AuthorsOpenReview ↗

Response to Reviewer 38wN (2/3)

W2: Limited novelty.

R2: We thank the reviewer for this comment. The primary contribution of DLVR lies in how structured table reasoning is organized, represented, and optimized within continuous latent space. We do not claim structure-content decoupling or latent utility optimization as standalone innovations; rather, we clarify below the specific advances over DiSCo and LUT.

(1) Decoupling within continuous latent reasoning. Separately modeling structure and content is a common practice in structured data understanding, as demonstrated in DCNE (TKDE 2023), LiLT (ACL 2022), DocLLM (ACL 2024), and ChainCoder (ICML 2023). Our novelty does not lie in introducing the general concept of structure-content decoupling. However, applying it to continuous latent reasoning is challenging because latent blocks lack explicit semantic boundaries. DLVR addresses this by organizing structural and content blocks across row, column, and subtable stages, and jointly optimizing them to learn and integrate complementary information during progressive reasoning.

(2) Difference from DiSCo. DiSCo focuses on explicit multimodal alignment and table understanding, disentangling structure and content through separate text-generation objectives. In contrast, DLVR addresses a largely unexplored challenge: how to effectively represent and utilize table structure during continuous latent reasoning, where intermediate states lack explicit semantic boundaries. To address this challenge, DLVR introduces stage-wise decoupled latent reasoning, organizing structural and content blocks across row, column, and subtable stages, with joint objectives to guide their complementary representation and integration. Thus, DLVR addresses the organization and utilization of structural information within latent reasoning, a challenge not explored by other explicit alignment and reasoning frameworks.

(3) Difference from LUT. LUT uses answer-likelihood differences to select teacher trajectories for distillation and attention-based attribution to reweight latent-step optimization. In contrast, DLVR performs targeted interventions on structural or content blocks at each reasoning stage, measures their contributions through changes in answer likelihood, and incorporates these signals as an additional reward in LI-GRPO. Thus, the key distinction lies in how latent utility is attributed to structurally defined intermediate blocks and incorporated into policy optimization. This encourages the model to learn structural and content representations that contribute more effectively to answer prediction, rather than merely optimizing overall reasoning outcomes.

(4) Empirical evidence. Our ablation results (Table 2) demonstrate that removing structure-content decoupling consistently degrades performance, particularly on WTQ and InfoTabs. Removing LCR also reduces performance across all five representative benchmarks, supporting its complementary contribution to the framework. Furthermore, DLVR outperforms both DiSCo-based direct answering and CoT baselines across all evaluated benchmarks, demonstrating the effectiveness of our structured latent reasoning framework.

Q2: Can you make it more clear how the novelty claims contrast with prior work?

R3: We thank the reviewer for this suggestion. We clarify our novelty claims by emphasizing three key contributions and their distinctions from prior work:

(1) Structure-guided progressive latent reasoning. Unlike conventional table reasoning methods that primarily rely on explicit textual reasoning or table operations, DLVR explores structured table reasoning within continuous latent space. Furthermore, unlike existing latent visual and textual reasoning methods that lack explicit modeling of table structure, DLVR organizes latent reasoning into row, column, and subtable stages, enabling progressive utilization and integration of table-specific structural information.

(2) Decoupled latent block learning. Unlike DiSCo, which separates structure and content through text-generation objectives during multimodal alignment, DLVR learns complementary structural and content blocks within the latent reasoning process itself, jointly optimized through content alignment, fusion reconstruction, and orthogonality regularization.

(3) Intervention-based latent optimization. Unlike LUT, which uses trajectory-level utility for teacher trajectory selection and attention-based attribution for latent-step optimization, DLVR evaluates the contribution of structurally defined latent blocks through targeted interventions and incorporates these signals as an additional reward in LI-GRPO.

Overall, our novelty lies in how structured reasoning is organized, represented, and optimized within continuous latent space, rather than in the individual concepts of structure-content decoupling or latent utility optimization.

AuthorsOpenReview ↗

Response to Reviewer 38wN (3/3)

W3&Q1: Small gains and weak baselines.

R4: We thank the reviewer for these valuable comments. To provide a more comprehensive comparison, we extend our evaluation of LVR-Random to three random seeds, reporting mean performance and standard deviation to account for experimental variability. We also include a stronger proprietary model, GPT-6 Luna, for evaluation.

Table 1: Comparison with LVR-Random and proprietary LVLMs across seven multimodal table reasoning benchmarks.

MethodWTQTAT-QAHiTabAIT-QATabFactInfoTabsPubHealthTabAvg.
GPT-6 Luna72.5436.2759.9041.8893.8580.9487.0267.49
LVR-Random57.60±0.4336.66±0.2642.07±0.3485.65±1.2077.49±0.5969.98±2.2477.24±2.1463.81±0.94
DLVR (Ours)59.81±0.1836.70±0.2742.64±0.5188.39±0.6078.28±0.0772.35±0.1276.31±1.1764.93±0.30

(1) Comparison with LVR-Random. Since LVR-Random relies on randomly sampled regions as latent targets, we repeat the experiments with three random seeds (42, 2026, and 26) and report mean ± standard deviation. As shown in Table 1, DLVR outperforms LVR-Random on six of seven benchmarks, with notable improvements on WTQ (+2.21), AIT-QA (+2.74), and InfoTabs (+2.37). Moreover, LVR-Random exhibits larger performance variations on several benchmarks, particularly InfoTabs and PubHealthTab, whereas DLVR achieves more stable results across seeds. Although LVR-Random remains competitive on PubHealthTab, these results suggest that explicitly modeling table-specific structural information provides more consistent benefits than randomly sampled latent targets.

(2) Evaluation with proprietary models. We clarify that the GPT-4o-mini results in our original comparison are adopted from previous work [1]. We retain the seven benchmarks shared with our evaluation and include these results as a reference baseline. Following the reviewer's suggestion, we further evaluate GPT-6 Luna, which achieves strong performance on WTQ (72.54), HiTab (59.90), TabFact (93.85), InfoTabs (80.94), and PubHealthTab (87.02).

Overall, these additional experiments provide further evidence of the benefits of explicit table-structure modeling over random latent targets and extend our evaluation to a stronger proprietary LVLM.

References:

[1] Yingjie Zhu, Xuefeng Bai, Kehai Chen, Yang Xiang, Youcheng Pan, Xiaoqiang Zhou, and Min Zhang. Decoupling skeleton and flesh: Efficient multimodal table reasoning with disentangled alignment and structure-aware guidance. ICML 2026.

Action taken: We have clarified the contributions of DLVR and its distinctions from prior methods in Section 2, added GPT-6 Luna to the main comparison in Table 1, and included three-seed comparisons with LVR-Random and different GRPO variants in Appendix C (Table 10), with corresponding discussion in Section 5.

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

DLVR organizes latent table reasoning into row, column, and subtable stages, with separate structure and content blocks. It learns these blocks using a new 41,567-example corpus and then applies LI-GRPO, whose reward measures how much replacing a latent block reduces the generated answer’s likelihood. Experiments use Qwen2.5-VL-7B and Qwen3-VL-8B on seven table benchmarks. The method improves average performance and uses shorter inference sequences than the tested explicit reasoning methods.

Critical strengths
  1. The table-specific organization is sensible, and the data-matched comparisons are valuable. Table 1 includes direct LoRA training and generic latent variants using TableLVR-41K, so the results go beyond comparing against an untuned backbone. Gains on AIT-QA and PubHealthTab also provide evidence of transfer beyond the training sources.

  2. The inference savings are concrete. Table 3 counts latent states and control tokens and reports measured runtime: the three-task average is 1.79 seconds for DLVR. This makes compact latent computation an attractive direction for table reasoning.

Critical weaknesses
  1. The central interpretation of the latent blocks is not established. The structural target is a residual of visual features after projecting cell text, and orthogonality alone does not identify table structure. Likewise, replacing blocks with control-token embeddings can hurt through an unusual input distribution. Figure 4 shows that LCR increases, but does not demonstrate that the blocks learn the intended row/column relations.

  2. The added value of LI-GRPO, especially LCR, is modest relative to the claims. In Table 8, removing LCR changes WTQ from 59.94 to 59.69 and AIT-QA from 89.04 to 88.65, while improving TAT-QA. No seed variation is reported for these comparisons. Also, the direct LoRA baseline does not receive a matched GRPO stage. I am not yet convinced how much of the gain requires the proposed intervention reward.

Questions for authors
  1. Assessment-relevant validation: Can the authors show that structure blocks capture row/header relations using layout changes that preserve cell text, or a matched latent replacement control? A drop under control-token replacement alone does not distinguish semantic contribution from sensitivity to corrupted inputs.

  2. Assessment-relevant comparison: How does direct TableLVR-41K-LoRA followed by the same correctness/format GRPO perform? Please also report seed variation for DLVR versus the no-LCR variant. These comparisons would materially affect my view of the second training stage.

6 comments
AuthorsOpenReview ↗

Response to Reviewer xmCa (1/2)

We sincerely thank you for your thoughtful and constructive feedback. Please find our point-by-point responses below. We have clarified the interpretation of the structural proxy in Section 3.2.2, added relation-preserving and counterfactual layout-transformation experiments in Section 5, and expanded the LI-GRPO comparisons with additional baselines and three-seed results in Section 5 and Appendix C. All revisions to the manuscript are highlighted in blue for ease of reference. We hope these revisions and responses address your concerns, and we welcome any further questions or suggestions.


W1&Q1: Interpretation and Validation of Structural Blocks.

R1: We thank the reviewer for these insightful comments. We clarify the interpretation of structural latent blocks and provide additional experiments to validate the model's ability to capture table-specific structural relations.

(1) Interpretation of Structural Blocks.

  • We agree that residual construction and orthogonality alone do not establish that structural blocks capture row/header relations. As table structure is inherently relational and lacks an explicit standalone target, we approximate structural information using the residual of visual features after projecting cell text, i.e., Sm=Vm−P(Tm)S_m=V_m-P(T_m). Beyond residual construction and orthogonality alone, DLVR jointly employs structure-guided stage supervision, content alignment, fusion reconstruction, and orthogonality regularization to encourage structural blocks to capture information complementary to cell content, including row/column relations.
  • Importantly, the residual is only an approximate structural proxy used during teacher forcing, rather than a direct regression target for the generated structural blocks. We have revised the manuscript to describe the structural blocks as encouraged to capture relational information complementary to cell content, rather than implying a guaranteed structure-content separation.

(2) Validation of Structural Information Learning. Following your suggestion, we conduct controlled layout-transformation experiments that preserve the collection of cell texts to validate whether the latent blocks capture row/header relations. Specifically, we first extract a subset of test samples and prompt GPT-6 Luna to construct two types of transformations, and then perform programmatic checks to verify table structure, text preservation, permutation consistency, and the validity of answer and evidence fields:

  • Relation-preserving (Positive): Rearrange complete rows or columns together with their headers, preserving header-cell associations and the original answer.
  • Relation-changing (Counterfactual, CF): Swap question-relevant values across rows or columns while keeping headers and row identifiers fixed, thereby changing their associations and updating the correct answer.

Table 1: Evaluation under layout transformations.

ModelPositive Original Acc.Positive Acc.Correct RetentionCF Original Acc.CF Acc.CF Change Rate
Qwen2.5-VL74.1773.1391.6574.2064.2473.85
Qwen3-VL73.6372.9192.3673.1565.2972.83

We compare accuracy on the original samples (Positive/CF Original Acc.) and their transformed counterparts (Positive/CF Acc.). Correct Retention measures the proportion of originally correct Positive pairs whose transformed predictions remain correct, while CF Change Rate measures the proportion of originally correct CF pairs whose predictions change after transformation, regardless of their correctness.

As shown in Table 1, both models achieve high correct retention rates under relation-preserving transformations, indicating that their correct predictions are largely maintained despite changes in visual layout. Under relation-changing transformations, the CF change rates reach 73.85% and 72.83%, respectively, demonstrating that the models respond to changes in row/header associations even when the cell texts remain unchanged. The corresponding CF accuracies (64.24% and 65.29%) further indicate that most transformed questions are answered correctly. These results provide complementary behavioral evidence that DLVR utilizes table-specific structural relations rather than relying solely on fixed visual positions or cell content.

Additionally, our training data are independently constructed, with rendering styles and overall table layouts substantially different from those in the evaluation benchmarks. This encourages the model to rely on underlying row/column relations rather than memorized visual patterns when reasoning over unseen tables.

AuthorsOpenReview ↗

Response to Reviewer xmCa (2/2)

W2&Q2: Contribution of LI-GRPO and Seed Variation.

R2: We thank the reviewer for this valuable suggestion. To disentangle the contribution of the proposed LCR, we conduct additional experiments with three random seeds (42, 2026, and 26), comparing direct TableLVR-41K-LoRA with standard GRPO, DLVR with standard GRPO (without LCR), and DLVR with LI-GRPO (with LCR). All variants use the same correctness and format rewards, with LCR additionally incorporated only in LI-GRPO. We report mean ± standard deviation on the five benchmarks used in our original ablation study.

Table 2: Performance comparison of different GRPO variants across five benchmarks.

MethodWTQHiTabAIT-QATabFactInfoTabs
LoRA + GRPO58.24 ± 0.4342.22 ± 0.0383.30 ± 1.1377.81 ± 0.4460.75 ± 0.45
DLVR + GRPO58.73 ± 0.8442.36 ± 0.4687.93 ± 1.0877.98 ± 0.0771.24 ± 0.60
DLVR + LI-GRPO59.81 ± 0.1842.64 ± 0.5188.39 ± 0.6078.28 ± 0.0772.35 ± 0.12

Matched GRPO comparison. DLVR + standard GRPO outperforms LoRA + GRPO on all five benchmark means, including +4.63 points on AIT-QA and +10.49 on InfoTabs. Thus, these gains neither require LCR nor arise solely from additional GRPO training.

Additional LCR contribution. Adding LCR consistently improves the three-seed mean performance across all five benchmarks, with gains +1.08 gains on WTQ and +1.11 on InfoTabs. These results support a consistent additional benefit from LCR beyond standard GRPO.

These results suggest that the primary gains come from the structured latent reasoning framework, while LCR provides additional improvements by explicitly evaluating the contribution of intermediate latent blocks and incorporating the resulting signals into policy optimization.

Action taken: We have clarified the interpretation of the structural proxy in Section 3.2.2, added relation-preserving and counterfactual layout-transformation experiments in Section 5 (Table 4), and expanded the LI-GRPO comparisons with additional baselines and three-seed results in Section 5 and Appendix C (Tables 2, 9, and 10).

AuthorsOpenReview ↗

Request for Feedback

Dear Reviewer xmCa,

I hope this message finds you well. Thank you very much for your valuable comments! We have carefully addressed each of your concerns in our rebuttal and hope that our clarifications and additional experiments are helpful.

In our previous responses, we specifically addressed your concerns regarding the interpretation and validation of structural latent blocks (W1 & Q1), including additional relation-preserving and counterfactual layout transformation experiments, as well as the contribution of LI-GRPO and LCR (W2 & Q2), supported by matched GRPO baselines and three-seed experimental results. We have also incorporated these revisions into the manuscript, with changes highlighted in blue.

We sincerely hope that our responses have adequately addressed your concerns and would be very grateful to hear your feedback. If any questions or concerns remain, we would be happy to discuss them further and provide additional clarifications.

Thank you again for your time and constructive suggestions!

Best regards,
All Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.