Research map / Vision-language and multimodal / Multimodal LLMs

TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 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 TableDART, a frozen-expert framework that routes each table-query pair to Text-only, Image-only, or LLM-agent Fusion paths via a 2.59M-parameter MLP gate, reaching new open-source SOTA on seven table benchmarks.

Problem

Table-as-Multimodality MLLM approaches statically process both text and image views of every table-query pair, causing redundancy and conflicts while requiring costly MLLM fine-tuning.

Approach

Reuse frozen single-modality table experts (TableGPT2-7B for text, Ovis2-8B for image) and train only a lightweight gating MLP that routes each instance to text-only, image-only, or fusion; fusion is done by a training-free Gemini 2.0 Flash agent that arbitrates or rescues answers.

Key results

• TableDART (TableGPT2-7B + Ovis2-8B) averages 74.86% accuracy across accuracy benchmarks, surpassing HIPPO-8B (70.84%) by 4.02 points.
• Dynamic routing reduces mean latency by 24.5% (2.92s to 2.20s) versus the Non-Adaptive Fusion baseline.
• Only the 2.59M-parameter gate is trained, versus 25.87M trainable LoRA parameters in HIPPO, on a single H100 for about 13.5 hours.

Surprising findings

• Random routing hurts badly (e.g., WTQ 65.40 vs 70.58), while the non-adaptive fusion baseline is competitive on several datasets.
• Lower resource-loss weight (lambda=0.00) gives best results on several datasets, yet the chosen lambda=0.15 is kept for its efficiency trade-off.

Limitations

• Fusion path depends on a closed-source Gemini 2.0 Flash API, adding cost and reproducibility concerns.
• Gate trained on only 10k samples from five benchmarks; two benchmarks used as zero-shot evaluation.
• Routing results on TFV/TQA show mixed per-dataset gains; not uniformly best across all benchmarks.

Experimental setup

Datasets: TABMWP WTQ TAT-QA TabFact InfoTabs

Benchmarks: WTQ TABMWP TAT-QA HiTab FeTaQA TabFact InfoTabs

Metrics: accuracy BLEU

Baselines: HIPPO-8B Gemini 2.0 Flash TableGPT2-7B Ovis2-8B Qwen2.5-VL-7B SynTab-LLaVA-7B MiniCPM-V-2.6-8B TableLlama-7B Table-LLaVA-7B Random Routing

Builds on: HIPPO TableGPT2 Ovis2

Closest papers by meaning

• Multimodal Tabular Reasoning with Privileged Structured Information (NeurIPS 2025 · similarity 0.91)
• TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.91)
• Decoupling Skeleton and Flesh: Efficient Multimodal Table Reasoning with Disentangled Alignment and Structure-aware Guidance (ICML 2026 · similarity 0.90)
• TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective (CVPR 2026 · similarity 0.90)
• NeedleInATable: Exploring Long-Context Capability of Large Language Models towards Long-Structured Tables (NeurIPS 2025 · similarity 0.89)
• SynTab-LLaVA: Enhancing Multimodal Table Understanding with Decoupled Synthesis (CVPR 2025 · similarity 0.89)