Research map / Vision-language and multimodal / Multimodal LLMs

TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

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 TableMix, which mixes multimodal table reasoning, text-only math, and table perception data in each batch and trains Qwen2.5-VL-7B with GRPO plus a difficulty-aware length reward, reaching state-of-the-art on table benchmarks.

Problem

Multimodal LLMs underperform text-only models on table reasoning, which the authors attribute to vision-language alignment pre-training degrading the LLM's intrinsic reasoning, and to RL treating mixed-difficulty samples uniformly.

Approach

Batch-level mixing of multimodal table reasoning, text-only math reasoning (to restore reasoning), and simple table perception data (to preserve visual grounding), combined with a difficulty-aware reward that adds a length penalty only for questions with high group success rate.

Key results

• TableMix reaches 99.20 on TabMWP and 81.32 on WTQ, exceeding Turbo-8B (96.75, 67.80) and Table-R1 (96.40, 81.20 on WTQ).
• On zero-shot TableVQA-Bench, TableMix averages 82.3 versus 70.2 for Qwen2.5-VL-7B and 69.7 for Ovis2-8B.
• DRS reduces average inference tokens by about 20% without loss of accuracy relative to vanilla GRPO.

Surprising findings

• Fine-tuning only on table reasoning data degraded simple table perception, showing a reasoning-perception tension.
• Two-stage training (math then table, or reverse) slightly underperformed batch-level mixing.

Limitations

• Gains depend on the choice of mixing ratios and math dataset; MetaMath performed best among tested options.
• Evaluation on FeTaQA relies on DeepSeek-V3 to standardize answers, introducing judge dependence.
• Hardware and training compute are not reported.

Experimental setup

Datasets: TabMWP WTQ HiTab TAT-QA FeTaQA ToTTo TabFact FEVEROUS HybridQA FinQA MultiModalQA InfoTabs

Benchmarks: TabMWP WTQ HiTab TAT-QA FeTaQA TabFact InfoTabs TableVQA-Bench

Metrics: accuracy

Baselines: Turbo-8B HIPPO-8B Table-R1 SynTab-LLaVA-7B Table-LLaVA-7B Qwen2.5-VL-7B Ovis2-8B TableGPT2-7B

Builds on: GRPO DeepSeek-R1 Turbo MetaMath HIPPO

Closest papers by meaning

• Multimodal Tabular Reasoning with Privileged Structured Information (NeurIPS 2025 · similarity 0.91)
• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.90)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.90)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
• Decoupling Skeleton and Flesh: Efficient Multimodal Table Reasoning with Disentangled Alignment and Structure-aware Guidance (ICML 2026 · similarity 0.90)
• TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.90)