acceptodds
Withdrawn from ICLR 2027

Mitigating Structural Hallucinations in Complex Table Recognition via Post-Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language models often struggle with complex tables containing merged cells, frequently hallucinating incorrect row and column spans and thereby producing structurally inconsistent table representations. In this work, this work investigates post-training strategies for improving span-level structural fidelity in table recognition. Starting from PaddleOCR-VL 1.6, this paper progressively applies supervised fine-tuning (SFT), direct preference optimization (DPO), and GSPO to reduce rowspan and colspan errors. Experiments on the datasets show consistent improvements in table structure recognition, with particularly strong gains on tables involving complex cell spanning patterns. The results demonstrate that post-training provides an effective way to mitigate span-related structural hallucinations in complex table recognition.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Table recognition, Document Parsing, VLM, Large language model

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aDbHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether post-training can improve the structural accuracy of table recognition, particularly for merged cells with complex row and column spans. Using PaddleOCR-VL 1.6 as the backbone, the paper proposes a three-stage pipeline consisting of supervised fine-tuning (SFT), direct preference optimization (DPO), and group relative policy optimization (GRPO). The method combines coverage-balanced training data, structurally filtered preference pairs, and rewards based on TEDS, TEDS-S, and exact merged-cell F1.

Critical strengths

The evaluation of exact merged-cell recovery is potentially useful.

Critical weaknesses

I have serious concerns about the integrity and preparation of this submission. The paper contains numerous basic errors that should have been caught by the authors:

  1. The OpenReview title does not match the PDF title.
  2. The submitted abstract claims GSPO, whereas the PDF describes GRPO.
  3. Multiple references contain incorrect author information or arXiv IDs. For example, the Dolphin reference (p. 11) points to an entirely unrelated paper.
  4. The mandatory AI Use Statement is missing.
  5. The reported improvements are based on test-set checkpoint selection, while several transfer datasets show regressions.

The number of elementary inconsistencies raises serious concerns that the manuscript may have been AI-generated or assembled without adequate human verification. Regardless of how it was produced, the submission appears not to have undergone even basic author-level fact-checking.

Questions for authors

The number of obvious errors in this paper is quite concerning. Several references are incorrect or point to unrelated papers. The submitted abstract describes GSPO, while the manuscript uses GRPO. The required AI Use Statement is also missing.

These are basic issues that should have been caught before submission. I strongly suspect that the manuscript was largely generated using AI without meaningful review by the authors.

Can the authors clarify how this paper was prepared, whether generative AI was used, and what steps were taken to verify the references, methodology, and experimental results before submission?

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

This paper proposes TabFit, a post-training strategy to mitigate the structural hallucination in table recognition. However, the PDF content does not match the paper title and abstract in OpenReview.

Critical strengths

None

Critical weaknesses
  1. The proposed TabFit utilizes existing SFT+DPO+GRPO training strategy for table recognition. The novelty is significantly limited.
  2. The PDF content does not match the paper title and abstract in OpenReview.
Questions for authors

No question for authors

Reviewer_KiH6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TABFIT (Table Fidelity Tuning), a three-stage post-training framework designed to improve structural table recognition in the compact vision-language model PaddleOCR-VL 1.6 without altering its architecture. The proposed method combines SFT, DPO, GRPO driven by a verifiable reward and addresses recurrent table parsing topology errors such as dropped spans, misaligned rows/columns, and unexpected repetitions. The evaluation shows competitive results of the proposed method.

Critical strengths

In general, the paper is organized, and easy to follow. Moreoever, the proposed systematic taxonomy of structural hallucinations and a well-designed three-stage post-training recipe seems to effectively bridge text transcription correctness with structural layout integrity.

Critical weaknesses
  • Too many acronyms being defined in inappropriate places. For instance, TEDS (Tree Edit Distance-based Similarity) is only defined in 3.1, while being used in Introduction. This really downgrades the readibility.
  • Also refer to some of the questions.
Questions for authors
  • Why the base model makes these errors in the first place?
  • Why are the specific values for training sets and sampling selected in this way? Could there be an ablation on this?
  • What is the reason behind 'results of the sweep' instead of 'held-out test set'?
  • What are some main causes for mixed results in the transfer?
  • Could there be some computational perspectives (advantages or disadvantages) to the proposed method?

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

Related papers

Discussion (0)

Sign in to comment.