Research map / Vision-language and multimodal / Multimodal LLMs

TabFlash: Efficient Table Understanding with Progressive Question Conditioning and Token Focusing

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 2 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 TabFlash, an efficient table-image MLLM that injects questions into ViT layers at progressively increasing frequency, prunes low-L2-norm background tokens, and uses a token focusing loss to concentrate answer information in retained tokens.

Problem

Table images need question-specific focus and contain much redundant background, but existing MLLMs produce question-agnostic, redundant visual tokens that are costly for the LM to process.

Approach

Progressive question conditioning (question injected into ViT layers at increasing frequency), L2-norm-based background token pruning compatible with FlashAttention, and a token focusing loss (promotion loss on retained tokens plus uniform-KL suppression on pruned tokens) that moves answer information into retained tokens.

Key results

• TabFlash (3B) averages 60.5 across seven table QA benchmarks, outperforming InternVL-2.5 (3B) at 57.5 and the best open-source baseline.
• TabFlash (3B) uses 10.38 TFLOPs and 17.3G memory versus 14.23 TFLOPs and 24.7G for InternVL-2.5 (3B), a 27% FLOPs and 30% memory reduction.
• Token focusing with pruning (p=0.3) keeps 96% of the unpruned 1B model's accuracy (48.4 vs 50.3) while cutting TFLOPs by 31%; removing the promotion loss drops accuracy to 37.2.

Surprising findings

• Negative cross-entropy as the suppression loss performs worse than omitting it (27.2 vs 47.3), so the uniform-KL suppression is needed.
• Injecting the question into all ViT layers degrades performance (43.4) relative to conditioning only a subset of layers, showing that overly frequent conditioning hurts.

Limitations

• Pruning at higher rates (p>=0.4) causes notable accuracy drops, so the efficiency-accuracy trade-off is tunable but imposes a cost.
• Gains over InternVL-2.5 on generation tasks (Rotowire, WikiBio) are not consistent; TabFlash trails SynTab-LLaVA on some metrics.
• Proprietary model results rely on a subset evaluation for GPT-4V and API results for GPT-4o and Gemini.

Experimental setup

Datasets: MMTab-pre MMTab-instruct

Benchmarks: TABMWP WTQ HiTab TAT-QA FeTaQA AIT-QA TabMCQ TabFact InfoTabs PubHealthTab HiTab_t2t Rotowire

Metrics: accuracy BLEU TFLOPs peak GPU memory throughput

Baselines: GPT-4o Gemini 2.5 Pro GPT-4V LLaVA-1.5 Vary-toy Monkey TabPedia mPLUG-DocOwl1.5 Table-LLaVA SynTab-LLaVA InternVL-2.5

Builds on: InternVL-2.5 Table-LLaVA FlashAttention LoRA

Closest papers by meaning

• TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.89)
• Multimodal Tabular Reasoning with Privileged Structured Information (NeurIPS 2025 · similarity 0.89)
• FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.87)
• Benchmarking Attention for Tabular Foundation Models (NeurIPS 2026 · similarity 0.87)
• TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.87)
• SynTab-LLaVA: Enhancing Multimodal Table Understanding with Decoupled Synthesis (CVPR 2025 · similarity 0.87)