CFMS: A Coarse-to-Fine Multimodal Synthesis Framework for Enhanced Tabular Reasoning

Qixian Huang; Hongqiang Lin; Tong Fu; Yingsen Wang; Zhenghui Fu; Qirui Wang; Yiding Sun; Dongxu Zhang

cs.AIApr 13, 2026

CFMS: A Coarse-to-Fine Multimodal Synthesis Framework for Enhanced Tabular Reasoning

Authors: Qixian Huang, Hongqiang Lin, Tong Fu, Yingsen Wang, Zhenghui Fu, Qirui Wang, Yiding Sun, Dongxu Zhang

Abstract

Reasoning over tabular data is a crucial capability for tasks like question answering and fact verification, as it requires models to comprehend both free-form questions and semi-structured tables. However, while methods like Chain-of-Thought (CoT) introduce reasoning chains, purely symbolic methodes are inherently limited by their blindness to holistic visual patterns. To address this, we propose the Coarse-to-Fine Multimodal Synthesis framework (CFMS), a novel two-stage paradigm that hierarchically decouples high-level visual perception from granular symbolic reasoning. In the Coarse Stage, CFMS leverages the Multimodal Large Language Models (MLLMs) to perform a one-time synthesis of a multi-perspective knowledge tuple. This tuple subsequently serves as a dynamic reasoning map to guide the fine stage, where a symbolic engine executes a targeted and efficient sequence of iterative operations over the table. Extensive experiments on the WikiTQ and TabFact benchmarks demonstrate that CFMS achieves competitive accuracy. The framework exhibits particular robustness when handling large tables and when instantiated with smaller backbone models, validating its effectiveness and generalizability.

Figures & tables

Fig. 1: An illustration of CFMS’s advantage in holistic reasoning. (a) A symbolic-only method must resort to inefficient, brute-force calculations to identify volatility. (b) Our CFMS framework uses the Coarse Stage to synthesize a knowledge tuple $K_{coarse}$ , including a visual perception $R_{visual}$ of a generated chart, which immediately identifies the holistic pattern and guides the Fine Stage to a correct and efficient answer.

Fig. 2: An overview of the CFMS framework. The process begins with the Coarse Stage, where the initial table $T_{0}$ is analyzed from multiple perspectives (visual, semantic, statistical) to synthesize a textual knowledge tuple $K_{coarse}$ . Subsequently, in the Fine Stage, an LLM planner leverages this holistic context to guide a step-by-step iterative reasoning process, executing a sequence of precise table operations to derive the final answer.

Methods	GPT 3.5		LLaMA 2
Methods	TabFact	WikiTQ	TabFact	WikiTQ
End-to-End QA	70.45	51.84	44.86	23.90
Few-Shot QA	71.54	52.56	62.01	35.52
Chain-of-Thought	65.37	53.48	60.52	36.05
Chain-of-Table	80.20	59.94	67.24	42.61
Text-to-SQL	64.71	52.90	64.03	36.14

TABLE I: Table understanding results on TabFact and WikiTQ datasets across GPT 3.5 and LLaMA 2 models. We report accuracy (%).

Methods	Table Size
Methods	Small	Medium	Large
Binder	56.54	26.13	6.41
Dater	62.50	42.34	34.62
CFMS (ours)	69.26	53.47	46.92

TABLE II: Performance of Related Methods on small, medium, and large tables from WikiTQ. We also report accuracy (%).

Methods	Total # of queries	# of queries in each step
Binder	50	Generate Program: 50
Dater	100	Decompose Table: 40; Generate Cloze: 20; Generate SQL: 20; Query: 20
CFMS (ours)	$\leq$ 15	Coarse Stage: 4; Fine Stage: $\leq$ 10; Final Query: 1

TABLE III: Number of queries generated for a single question in the proposed CFMS on the WikiTQ dataset.

Fig. 3: Performance of the proposed CFMS on WikiTQ for questions that require an operation chain of varying lengths.

Methods	TabFact	WikiTQ
Methods	Accuracy	Accuracy
CFMS	81.37	61.42
w/o $f\_add\_column()$	80.10 (1.27↓)	58.86 (2.56↓)
w/o $f\_select\_column()$	77.61 (3.76↓)	58.29 (3.13↓)
w/o $f\_select\_row()$	76.79 (4.58↓)	57.65 (3.77↓)
w/o $f\_group\_by()$	79.43 (1.94↓)	55.02 (6.40↓)

TABLE IV: Ablation study of the CFMS framework on TabFact and WikiTQ datasets. We report Accuracy (%).

Methods	TabFact	WikiTQ
Methods	Accuracy	Accuracy
CFMS (Full Model)	81.37	61.42
w/o Coarse Stage	73.55 (7.82↓)	55.18 (6.24↓)
w/o Fine Stage	75.96 (5.41↓)	56.44 (4.98↓)

TABLE V: Ablation study of the core components of our CFMS framework on TabFact and WikiTQ datasets.

Table QA

Table question answering studies how AI systems retrieve, interpret, and reason over structured data to answer natural-language questions.

See all 65 papers

TABQAWORLD: Optimizing Multimodal Reasoning for Multi-Turn Table Question Answering
Apr 3, 2026Tung Sum Thomas Kwok, Xinyu Wang, Xiaofeng Lin +7Table Reasoning Table QA
University of California, Los Angeles · SimpleWay · McGill University +2
Thinking with Tables: Enhancing Multi-Modal Tabular Understanding via Neuro-Symbolic Reasoning
Mar 25, 2026Kun-Yang Yu, Zhi Zhou, Shi-Yu Tian +6Multimodal Understanding Multimodal Large Language Models
State Key Laboratory of Novel Software Technology, Nanjing University, China · School of Artificial Intelligence, Nanjing University, China · School of Intelligence Science and Technology, Nanjing University, China
CRAFT: A Unified Counterfactual Reasoning Framework for Tabular Question Answering and Fact Verification
Jun 5, 2026Chenshuo Pan, Yu Zhao, Jie Zhang +7Table QA Claim Verification
Xingchen AGI Lab,China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd

Apr 3, 2026J/K move · Enter open · S save

TABQAWORLD: Optimizing Multimodal Reasoning for Multi-Turn Table Question Answering

Tung Sum Thomas Kwok, Xinyu Wang, Xiaofeng Lin, Peng Lu+6

University of California, Los Angeles · SimpleWay · McGill University+2

Table Reasoning Table QA LLM Reasoning Efficient Multimodal Inference Multimodal Reasoning

Open paper AskPDF arXiv

Multimodal reasoning has emerged as a powerful framework for enhancing reasoning capabilities of reasoning models. While multi-turn table reasoning methods have improved reasoning accuracy through tool use and reward modeling, they rely on fixed text serialization for table state readouts. This introduces representation errors in table encoding that significantly accumulate over multiple turns. Such accumulation is alleviated by tabular grounding methods in the expense of inference compute and cost, rendering real world deployment impractical. To address this, we introduce TABQAWORLD, a table reasoning framework that jointly optimizes tabular action through representation and estimation. For representation, TABQAWORLD employs an action-conditioned multimodal selection policy, which dynamically switches between visual and textual representations to maximize table state readout reliability. For estimation, TABQAWORLD optimizes stepwise reasoning trajectory through table metadata including dimension, data types and key values, safely planning trajectory and compressing low-complexity actions to reduce conversation turns and latency. Designed as a training-free framework, empirical evaluations show that TABQAWORLD achieves state-of-the-art performance with 4.87% accuracy improvements over baselines, with 5.42% accuracy gain and 33.35% inference latency reduction over static settings, establishing a new standard for reliable and efficient table reasoning.

Abstract

Table QA

Explore similar work

TABQAWORLD: Optimizing Multimodal Reasoning for Multi-Turn Table Question Answering

Thinking with Tables: Enhancing Multi-Modal Tabular Understanding via Neuro-Symbolic Reasoning

CRAFT: A Unified Counterfactual Reasoning Framework for Tabular Question Answering and Fact Verification