acceptodds
Under review as a conference paper at ICLR 2027

TableKIE: A Benchmark and Learning Framework TeleTable for Key Information Extraction from Tables

Abstract

Key information extraction (KIE) from tables goes beyond recognizing textual content and associating values with entities. It requires synthesizing hierarchical row and column headers with contextual information into composite semantic keys that precisely characterize each value. Complex layouts and repeated field labels further challenge vision-language models in linking these keys to the correct values. We introduce TableKIE, a benchmark and learning framework for evaluating and improving table-centric KIE. The benchmark covers two complementary settings: closed-schema extraction, where target fields are predefined, and open-schema extraction, where models identify and organize key information explicitly present in a table without a fixed field inventory. Our primary evaluation focuses on field-value correctness and requested-field coverage. On the subset with verified structural annotations, we additionally evaluate row- and column-path agreement and value-box localization to assess the grounding behavior induced by auxiliary structural supervision.

Primary area:
datasets and benchmarks
Keywords:
Key Information Extraction, Table Understanding, Vision-Language Models, Benchmarking, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5XmqOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the problem of Key information extraction (KIE) from tables. It introduces a KIE benchmark which contains two setting. In the closed setting, both key and value are required to be correctly predicted. The paper further introduces a post-trained LLM for KIE

Critical strengths
  1. The paper is well presented and easy to understand.
  2. A lot of models are compared in the proposed benchmark and extensive experiments are conducted to demonstrate the effectiveness of their proposed training strategy.
  3. The benchmark is curated from real-world documents
Critical weaknesses
  1. There are benchmarks evaluating capacities very similar to the proposed one, like TableVQA-Bench, making the novelty of the benchmark limited.
  2. According to Table 2, SOTA LLMs like Kimi K3 and Grok 4.7 have already achieves extremely high performance, which makes me believe the benchmark is saturated and of low value for evaluation on KIE.
  3. It is hard to understand why extra evidence is not provided in SFT but required in GRPO training process.
Questions for authors
  1. What's the difference between TableVQA-Bench and the proposed one.
  2. Why GPT-5.6 Sol performs well in close setting but much worse in open setting.
  3. Since Kimi K3 and Grok 4.7 have already performs well in the benchmark, what's the remaining challenges in KIE.
Reviewer_DvLwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduced two contributions: (1) TableKIEBench, a comprehensive benchmark for table-centric key information extraction (2) A framework combining TeleTable and TableGRPO. TableKIEBench gathers multiple documents from various sources that contains tables and contributes to solving a challenge of extracting useful information from these table-structure visuals through table-specific supervised finetuning and a reinforcement learning approach. The proposed method is shown with competitive performance compared to other vlms.

Critical strengths

In general, the paper throws an interesting challenge to unlock improvements in extracting useful information from tables within documents. The structure of tables within documents are rather complex and the proposed method attempts address the challenge accordingly. In general, the paper is easy to follow and contribution of both a benchmark and a method can be considered as a strong contribution.

Critical weaknesses
  • My main concern for the paper is the focus of 'tables'. Phrases related to 'extraction of key information from tables' suggest that the paper is focusing on the modality of tables, but the paper is in fact focusing on information extraction for visual tables. This gives rather a over-extended scope. If it is to cover the concept of 'table understanding', it needs to provide an in-depth analysis comparing with table understaning models, where the tables after OCR have been stored as dataframes, for instance. In my opinion, the papers seems to be more related to parsing of visual tables from documents.
Questions for authors
  • Addressing the concern with weakness would be important for the tabular research community.
  • Some minor typos should be fixed.
  • In table 2, it is hard to see which is performing good on which problems. Possibly bold on the best performing ones could make the table more clear.
  • In tables two, would there be any error bounds?
  • I might have missed, but what are the examples of Close requests, Open requests?
Reviewer_E681OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses key information extraction (KIE) from tables. The paper introduces a new benchmark for table KIE: TableKIE with two settings. In the closed setting, the TableKIE formulation takes as input an image, a region, and an optional key; the function must predict the value or null. In the open setting, the key is removed from the input, and the function should predict a value-key pair. The paper also proposes an evaluation protocol that matches extraction and label data. In addition, the paper introduces TeleTable, a training approach for closed TableKIE. TeleTable comprises two stages: SFT and TableGRPO (with custom rewards)

The TableKIEBench dataset includes 268 images and 4,490 KVs from multiple domains in both Chinese and English. The experiment results on TableKIEBench show that existing LLMs (such as Kimi K3 or Qwen3.8-27B) excel on the benchmark in both settings. Smaller LLMs (0.8B-9B) struggle, and TeleTable using Qwen3.5-4B base outperforms other smaller LLMs. Ablation analysis shows that GRPO and its rewards are beneficial. Additional experiments show that TeleTable also outperforms small LLMs in general document KIE.

Critical strengths
  1. The paper introduces both a benchmark (dataset and evaluation) and a learning approach. An additional benchmark in Table KIE benefits other research in the area TeleTable method can serve as a baseline for training small LLMs.
  2. The paper provides a comprehensive analysis of TeleTable performance to confirm the contribution of each component, alternative optimization, and also the cost of each stage. While GRPO does not provide a large gain (Table 3) over SFT, its cost is relatively low.
Critical weaknesses

Although the new benchmark is a genuine contribution and the experimental setup is comprehensive, some details are missing, and a few points need clarification.

  1. The TableKIEBench is well-defined. However, the paper does not discuss or evaluate label quality. The paper post-processes the labels using Qwen3.8-27B, which may introduce bias. A human audit would be helpful. In addition, it is unclear whether the training set would be part of TableKIEBench, as Figure 2 does not include the training set in Table 1.
  2. The paper misses many details in their methodology, which hinders future reproducibility:
    1. The paper describes TeleTable as a closed-only training approach, but Table 2 shows its performance in the open setting as well. It is unclear how TeleTable is adapted to perform in the open setting.
    2. The prompts/instructions for other models are not documented in the appendix or the supplementary material.
    3. The evaluation of the general document KIE is an important addition, but the paper does not describe how to adapt TeleTable models to the Nanonets suite. Is there any additional fine-tuning? Is there any change in instructions?
    4. The cost and runtime will need a hardware specification.
  3. The motivation for this benchmark is not well situated by the literature. While the paper discusses previous work and related benchmarks, it does not discuss the position of thie benchmark in relationship with previous work.
  4. There is an odd, unexplained performance difference in GPT-5.6 Sol. In Table 2, Close F1 is the highest, but Open F1 is really low.
Questions for authors
  1. What is the total data that is going to be included in the TableKIEBench?
  2. Could you clarify the missing detail per Weakness 2?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.