Research map / Vision-language and multimodal / Multimodal LLMs

Decoupling Skeleton and Flesh: Efficient Multimodal Table Reasoning with Disentangled Alignment and Structure-aware Guidance

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DiSCo, a disentangled structure-content alignment of LVLMs on ~10K table images, and Table-GLS, a training-free global-to-local sub-table reasoning procedure, improving table understanding and reasoning on in-domain and OOD benchmarks.

Problem

LVLMs struggle to reason over table images with complex layouts; existing fixes need costly supervised/RL training or external tools, and textual alignment entangles structure with cell content.

Approach

Decouple table alignment into a content-free structure objective (anonymized HTML/Markdown/LaTeX) plus global and local content objectives tied to row/column coordinates, then apply a training-free three-stage global-to-local prompting pipeline that selects rows/columns, extracts a self-refined sub-table, and reasons over it.

Key results

• Qwen3-VL-8B with DiSCo and Table-GLS reaches 59.48 average on table reasoning vs 50.31 for direct answering Qwen3-VL-8B.
• Qwen3-VL-32B with DiSCo and Table-GLS reaches 71.84 average, up from 66.12 baseline; GPT-5.4-mini rises from 56.64 to 66.16 with Table-GLS.
• DiSCo with 10K images matches or beats Textual-All (97K) on many understanding tasks, e.g., Qwen3-VL-8B TCL 55.95 vs 41.42.

Surprising findings

• Table-GLS applied directly to the original Qwen3-VL-8B shows a significant drop on OOD benchmarks, so DiSCo alignment is needed for robust gains.
• Textual alignment with full table text can degrade general multimodal benchmarks (e.g., ScienceQA, HallusionBench) relative to the original model.

Limitations

• Relies on availability of structure-content decoupled table data; may fail on highly irregular tables.
• Multi-stage Table-GLS generates more tokens than direct answering, increasing latency.
• Evaluation focuses on table-centric tasks; generalization to other structured visual reasoning is untested.

Experimental setup

Datasets: MMTab TABMWP WTQ FeTaQA HiTab TAT-QA TabFact InfoTabs ToTTo Rotowire WikiBIO

Benchmarks: MMTab table understanding tasks (TSD, TCL, RCE, MCD, TCE) WTQ HiTab TAT-QA AIT-QA TabMCQ TabFact InfoTabs PubHealthTab ScienceQA CRPE HallusionBench

Metrics: accuracy cell-level accuracy cell-level F1 inference time (s) average accuracy

Baselines: Table-LLaVA 7B Table-LLaVA 13B Table-R1 HIPPO GPT-4o-mini GPT-5.4-mini ReFocus Chain-of-Thought Row-of-Thought TableLlama

Builds on: MMTab Table-LLaVA Row-of-Thought ReFocus LoRA

Closest papers by meaning

• Multimodal Tabular Reasoning with Privileged Structured Information (NeurIPS 2025 · similarity 0.90)
• TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.90)
• TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective (CVPR 2026 · similarity 0.90)
• VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.87)
• TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.87)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.87)