CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation

Yunfan Yang; Cuiling Lan; Jitao Sang; Yan Lu

cs.AIApr 13, 2026

CSPO: Alleviating Reward Ambiguity for Structured Table-to-LaTeX Generation

Authors: Yunfan Yang, Cuiling Lan, Jitao Sang, Yan Lu

Organizations: Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, Beijing Jiaotong University · Microsoft Research Asia · Peng Cheng Laboratory

Abstract

Tables contain rich structured information, yet when stored as images their contents remain "locked" within pixels. Converting table images into LaTeX code enables faithful digitization and reuse, but current multimodal large language models (MLLMs) often fail to preserve structural, style, or content fidelity. Conventional post-training with reinforcement learning (RL) typically relies on a single aggregated reward, leading to reward ambiguity that conflates multiple behavioral aspects and hinders effective optimization. We propose Component-Specific Policy Optimization (CSPO), an RL framework that disentangles optimization across LaTeX tables components-structure, style, and content. In particular, CSPO assigns component-specific rewards and backpropagates each signal only through the tokens relevant to its component, alleviating reward ambiguity and enabling targeted component-wise optimization. To comprehensively assess performance, we introduce a set of hierarchical evaluation metrics. Extensive experiments demonstrate the effectiveness of CSPO, underscoring the importance of component-specific optimization for reliable structured generation. Our code is available at https://github.com/microsoft/CSPO.

Figures & tables

Figure 1: Motivating example of reward ambiguity in table image-to-LaTeX generation. denotes a positive reward, while denotes a penalty. (a) Given a table image, multiple LaTeX sequences are generated with varied errors in structure, content, and style (errors marked with ✗ ). Using a single global reward leads to reward ambiguity , where (a1) an incorrect structure is erroneously reinforced, (a2) correct content is unfairly penalized, and (a2) and (a3) receive identical rewards despite differing component fidelity. (b) Our method mitigates this by decomposing the LaTeX code into functional components and assigning component-specific rewards for targeted optimization, thereby alleviating reward ambiguity .

Figure 2: Overview of Component-Specific Policy Optimization (CSPO) . Particularly, CSPO decomposes each generated code sequence into functional components ( e . g ., structure, cell appearance, caption, package inclusion, alignment, and line style) using a LaTeX parser. It conducts component-specific rewarding by assessing each component’s fidelity (with a strong LLM as the judge), performs component-specific credit assignment and optimization , effectively mitigating reward ambiguity in table image-to-LaTeX generation.

Figure 3: Illustration of proposed CSPO algorithm. Each component-specific advantage $A_{c}^{(g)}$ ( $c\in\tilde{\mathcal{C}}$ ) is calculated, and credited exclusively to its relevant tokens in rollout $g$ .

Model	TEDS ↑	OF ↑	S ↑	C ↑	Y ↑	$Y_{line}$ ↑	$Y_{align}$ ↑	$Y_{cell}$ ↑	R ↑
GPT-4o ( OpenAI, 2025a )	72.3	6.8	65.0	73.1	9.4	23.0	54.0	72.5	99.9
Gemini-2.5 Flash ( Comanici et al., 2025 )	66.2	11.1	78.8	79.2	12.9	33.5	47.2	81.3	99.3
Qwen2.5-VL-72B ( Bai et al., 2025 )	75.9	10.1	79.6	76.2	12.6	32.1	56.7	69.9	99.7
Nougat ( Blecher et al., 2023 )	67.9	21.2	68.7	63.3	25.8	37.1	71.9	70.1	99.1
Qwen2.5-VL-3B	66.0	3.1	51.8	59.8	4.5	18.9	34.0	61.7	93.7
Qwen2.5-VL-3B-VSGRPO ∗ ( Ling et al., 2025 )	87.8	40.4	77.5	86.6	47.6	70.9	71.4	91.9	99.7

Table 1: Performance comparisons on TableTex-test (with fine-grained metrics evaluated by GPT-4o). We report metrics across hierarchical dimensions. Global metrics : TEDS, Overall Fidelity (OF), Compilation Rate (R); Fine-grained metrics : (i) Structure Fidelity (S), (ii) Content Fidelity (C); (iii) Style Fidelity(Y): Line Style (

Y_{\text{line}}

), Alignment (

Y_{\text{align}}

), Cell Style (

Y_{\text{cell}}

). Higher scores (↑) indicate better performance. Note that all evaluation metrics, initially defined in the

[0,1]

range ( e . g ., scoring and correctness measures), are presented as percentages (%) in all the tables for clarity. ∗ denotes that, for fair comparison, we reimplement the reward design of Ling et al. Ling et al. (2025) within our codebase and dataset.

	DocGenome-table-1k					Table2LaTeX-test-simple
Method	TEDS ↑	OF ↑	S ↑	C ↑	Y ↑	TEDS ↑	OF ↑	S ↑	C ↑	Y ↑
GPT-4o ( OpenAI, 2025a )	60.2	5.1	55.0	41.2	8.4	70.6	3.4	60.9	66.1	4.6
Gemini-2.5 Flash ( Comanici et al., 2025 )	60.5	8.0	64.9	57.1	12.5	71.5	10.5	72.0	78.4	12.9
Qwen2.5-VL-72B ( Bai et al., 2025 )	60.3	5.0	61.2	53.9	8.2	74.7	7.5	64.9	66.5	11.7
Nougat ( Blecher et al., 2023 )	30.8	0.8	27.7	4.4	8.8	27.0	0.4	20.6	2.8	5.7
Qwen2.5-VL-3B	50.2	1.1	43.5	37.0	1.7	63.5	2.0	39.1	44.2	3.0

Table 2: Generalization performance on DocGenome-table-1k and Table2LaTeX-test-simple (with fine-grained metrics evaluated by GPT-4o). ∗ denotes that, for fair comparison, we reimplement the reward design of Ling et al. Ling et al. (2025) within our codebase and dataset.

Model	TEDS ↑	OF ↑	S ↑	C ↑	Y ↑
Base	66.0	3.1	51.8	59.8	4.5
SFT	87.1	39.8	75.8	86.1	47.3
GRPO	87.7	42.0	74.7	86.7	50.3
CSPO w/ Comp. Sum	87.7	42.2	76.3	87.0	50.9
CSPO w/o Global	87.7	44.7	77.9	86.3	51.2
CSPO (Ours)	87.9	45.2	77.6	87.0	53.6

Table 3: Ablation studies on 3B models evaluated on TableTex-test .

LLM Evaluators/Judge	Method	Overall Fidelity	Structure Fidelity	Content Fidelity	Style Fidelity
GPT-4o (by default)	SFT	39.8	75.8	86.1	47.3
	GRPO	42.0	74.7	86.7	50.3
	CSPO (Ours)	45.2	77.6	87.0	53.6
Qwen3-Next-80b	SFT	37.6	72.0	88.4	47.2
	GRPO	39.7	71.6	88.2	49.9
	CSPO (Ours)	43.3	75.5	88.5	52.5

Table 4: Performance comparisons among 3B SFT, GRPO and our CSPO models, with fine-grained metrics evaluated by using different LLM judges.

Figure 4: A typical example comparing GRPO and CSPO of 7B models, showing CSPO mitigates structure and line style errors .

Figure 5: A typical example comparing GRPO and CSPO of 3B models, showing CSPO mitigates cell style errors .

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Description
$\mathbf{x}$	Input table image
$\mathbf{y}$	Generated LaTeX sequence
$\mathbf{y}^{*}$	Reference LaTeX sequence
$\pi_{\theta}$	Policy model with parameters $\theta$
$T$	Length of LaTeX sequence
$G$	Number of rollouts per group

Appendix

Table 5: Notation summary for CSPO framework.

Variant	TEDS ↑	OF ↑	S ↑	C ↑	Y ↑	$Y_{line}$ ↑	$Y_{align}$ ↑	$Y_{cell}$ ↑	R ↑
GRPO (no components)	87.7	42.0	74.7	86.7	50.3	74.9	71.5	92.0	99.6
CSPO w/o Content Reward	87.9	45.2	77.4	86.9	53.4	76.0	74.4	92.5	99.7
CSPO w/o Structure Reward	87.8	43.3	77.0	86.0	51.8	73.3	74.5	92.4	99.6
CSPO w/o Style Reward	87.7	43.1	79.0	87.5	49.6	72.3	72.7	92.4	99.6
CSPO (Ours)	87.9	45.2	77.6	87.0	53.6	76.5	74.2	92.1	99.6

Appendix

Table 6: Ablation study on reward components on 3B models.

Variant	TEDS ↑	OF ↑	S ↑	C ↑	Y ↑	$Y_{line}$ ↑	$Y_{align}$ ↑	$Y_{cell}$ ↑	R ↑
GRPO ( $w_{global}=10,w_{comp}=0$ )	87.7	42.0	74.7	86.7	50.3	74.9	71.5	92.0	99.6
CSPO ( $w_{global}=7,w_{comp}=3$ )	88.0	43.6	76.4	85.9	52.1	74.4	74.1	91.9	99.8
CSPO ( $w_{global}=5,w_{comp}=5$ )	87.9	44.8	77.1	86.9	53.1	76.0	73.8	92.7	99.7
CSPO ( $w_{global}=3,w_{comp}=7$ )	87.9	45.2	77.6	87.0	53.6	76.5	74.2	92.1	99.6
CSPO ( $w_{global}=0,w_{comp}=10$ )	87.7	44.7	77.9	86.3	51.2	73.7	73.4	92.1	99.8

Appendix

Table 7: Ablation on reward weights for different components on 3B models.

Method	TEDS	OF	Structure	Content	Style
GRPO	87.7	42.0	74.7	86.7	50.3
CSPO (Graded 0-3)	88.0	44.4	77.3	86.7	51.8
CSPO (Binary 0/1)	87.9	45.2	77.6	87.0	53.6

Appendix

Table 8: Performance comparison of CSPO using binary (0 or 1) and graded (0 to 3) reward on 3B models.

LLM Evaluators/Judge	Method	Overall Fidelity	Structure Fidelity	Content Fidelity	Style Fidelity
GPT-4o (by default)	SFT	50.7	79.7	90.3	59.0
	GRPO	50.6	80.7	89.7	58.9
	CSPO (Ours)	53.0	81.4	90.0	60.6
Qwen3-Next-80b	SFT	47.1	77.1	91.5	57.5
	GRPO	49.7	77.6	91.5	59.8
	CSPO (Ours)	52.0	79.5	92.2	61.8

Appendix

Table 9: Performance comparisons among 7B SFT, GRPO and our CSPO models, with fine-grained metrics evaluated by using different LLM judges.

Evaluator	Precision	Recall	F1	Accuracy
GPT-4o	89.8	86.9	88.3	89.8
Qwen3-Next-80b	88.8	85.6	87.2	88.8
DeepSeek-v3.2	92.4	76.9	84.0	87.0
GPT-5.2	97.5	70.7	82.0	86.2

Appendix

Table 10: Agreement between LLM judges with human annotations.

Method	OF	Structure	Content	Style
GRPO	42.0	74.7	86.7	50.3
CSPO (Reported)	45.2	77.6	87.0	53.6
CSPO (Mean ×8)	44.9	77.7	86.8	52.9
CSPO (Var ×8)	0.3	0.1	0.1	0.4

Appendix

Table 11: Stability of GPT-4o evaluator in testing (8 independent runs) evaluated on 3B models.

Figure 6: Prompt for LLM-based rewarding for components of package dependencies, caption correctness, structural organization, cell appearance, column alignment, and rule placement (vline, hline).

Figure 7: Prompt for LLM-based fine-grained fidelity evaluation in terms of content, structure, and style (line style, alignment, and cell style).

Figure 8: Training curves for 3B-GRPO and 3B-CSPO (ours). All curves are smoothed using a moving average (MA) for better visualization.

Figure 9: A typical example comparing GRPO and CSPO of 3B models, showing CSPO mitigates content errors .

Figure 10: A typical example comparing GRPO and CSPO of 3B models, showing CSPO mitigates structure , and content (in table caption) errors.

Figure 11: A typical example comparing GRPO and CSPO of 3B models, showing CSPO mitigates line style errors .

Figure 12: A typical example comparing GRPO and CSPO of 3B models, showing CSPO mitigates alignment errors .

Figure 13: Failure case for GRPO and CSPO of 3B models. Both models’ generations exhibit structure errors (marked by red boxes) on this complex table, where \multicolumn{} is used in groudtruth code but is ignored in the generated code. In addition, GRPO generation further has alignment errors (center aligned rather than left aligned as groudtruth).

Reward Decomposition

See all 11 papers

StrucTab: A Structured Optimization Framework for Table Parsing
Jun 29, 2026Gengluo Li, Shangpin Peng, Chengquan Zhang +10Table Structure Recognition VLM Reasoning
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, China · LLM Department, Tencent, China +3
TABVERSE: Benchmarking Cross-Format Table Understanding in LLMs and VLMs
Jun 8, 2026Momina Ahsan, Sarfraz Ahmad, Ming Shan Hee +2Document Understanding Table QA
1Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) · 2Singapore University of Technology and Design (SUTD)
InterTab: Interleaved Visual-Structure Alignment for Multi-Modal Table Reasoning
Sep 26, 2026Hanqian Li, Sirui Huang, Chen Ling +12Table QA Multimodal CoT Reasoning
The Hong Kong University of Science and Technology (Guangzhou) · Ant Group · Hong Kong Polytechnic University +4

Jun 29, 2026J/K move · Enter open · S save

StrucTab: A Structured Optimization Framework for Table Parsing

Gengluo Li, Shangpin Peng, Chengquan Zhang, Binghong Wu+9

Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, China · LLM Department, Tencent, China+3

Table Structure Recognition VLM Reasoning RL for VLMs

Open paper AskPDF arXiv

Table parsing aims to convert table images into structured, machine-readable representations, a task requiring the joint perception of complex spatial layouts and textual content. While recent vision-language models (VLMs) enable end-to-end parsing, they typically rely on direct supervision of the final output, thereby bypassing the explicit intermediate reasoning that is crucial for understanding complex table structures. Furthermore, attempts to optimize these models using reinforcement learning (RL) are often hindered by unstable or ambiguous reward designs, limiting potential performance gains. To address these limitations, we propose StrucTab, a table parsing model learned through intermediate structural supervision and reward decomposition. At the modeling level, by decomposing the parsing process into human-inspired subtasks, such as row-column counting and merged-cell analysis, StrucTab progressively unifies them through a sequential reasoning strategy. At the optimization level, we introduce Uni-TabRL, a unified RL framework that leverages decomposed rewards (validity, structure, and content) to provide stable and informative optimization signals. Finally, at the evaluation level, we present TableVerse-5K, a large-scale, challenging benchmark encompassing diverse, real-world table scenarios. Extensive experiments demonstrate the state-of-the-art performance of StrucTab across all evaluated public benchmarks and significant improvements on TableVerse-5K, validating the effectiveness of explicit structural modeling and decomposed reward optimization. Code and benchmark are publicly available at https://github.com/VirtualLUOUCAS/StrucTab.

Abstract

Reward Decomposition

Explore similar work

StrucTab: A Structured Optimization Framework for Table Parsing

TABVERSE: Benchmarking Cross-Format Table Understanding in LLMs and VLMs

InterTab: Interleaved Visual-Structure Alignment for Multi-Modal Table Reasoning