AI Paper Atlas

Research map / Benchmarks / TextVQA

TextVQA leaderboard

158 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR report results on TextVQA. Best accuracy: 98.5 (BP-Checkpointing, ICML 2026).

Numbers are as each paper reports them, read by AI from its results tables and not checked by hand. Papers can differ in setting (training data, model size, extra data); the topic column shows what each paper is about. Open the paper before you compare or quote a number.

TextVQA · accuracy

111 papers, each with its best reported number for its own method; higher is better.

#accuracy ↑MethodPaperTopic
198.5BP-Checkpointing
Qwen2-VL 7B (QLoRA r=1)
Memory Savings at What Cost? A Study of Alternatives to Backpropagation
ICML 2026 · Table 3
RNNs and normalization
286.9+ SPOT-E
Qwen2.5-VL-7B
SPOT-E: Test-Time Entropy Shaping with Visual Spotlights for Frozen VLMs
ECCV 2026 · Table 1
Multimodal LLMs
385.9AttWarp-Chains
Qwen2.5-VL
Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping
ICLR 2026 · Table 1
Multimodal LLMs
485.3OURO-7B
Qwen2-VL-7B-Instruct
OURO: A Self-Bootstrapped Framework for Enhancing Multimodal Scene Understanding
ICCV 2025 · Table 3
Multimodal LLMs
585T-DRS
Qwen2.5-VL-7B
Remember Me: Bridging the Long-Range Gap in LVLMs with Three-Step Inference-Only Decay Resilience Strategies
AAAI 2026 · Table 1
Multimodal LLMs
684IP-Merging
Qwen-2-VL-7B-Instruct
Can MLLMs Absorb Math Reasoning Abilities from LLMs as Free Lunch?
NeurIPS 2025 · Table 2
Instruction and fine-tuning
783.9InternVL3-8B+DeepAlign
InternVL3-8B
DeepAlign: Mitigating Modality Conflict through Modality-Specific Alignment
CVPR 2026 · Table 1
Multimodal LLMs
883.5Lyra-Pro (74B)Lyra: An Efficient and Speech-Centric Framework for Omni-Cognition
ICCV 2025 · Table 3
Multimodal LLMs
983.2R2-T2 (MoVA-7B, NGD)
MoVA-7B
R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts
ICML 2025 · Table: comparison of three R2-
Human and embodied modeling
1082.9DeFacto
Qwen2.5-VL-7B
DeFacto: Counterfactual Thinking with Images for Enforcing Evidence-Grounded and Faithful Reasoning
ICML 2026 · Table 2
Multimodal LLMs
101 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

TextVQA · score

36 papers, each with its best reported number for its own method; higher is better.

#score ↑MethodPaperTopic
178.3Oryx-1.5-32B
Qwen-2.5-32B
Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution
ICLR 2025 · Image Understanding table
Video-language understanding
277.2IXC2-4KHD
InternLM2-7B
InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD
NeurIPS 2024 · Table: Comparison with closed-
Multimodal LLMs
377.1SAIL
Mistral-7B-v0.1
The Scalability of Simplicity: Empirical Analysis of Vision-Language Learning with a Single Transformer
ICCV 2025 · Table 2
Cross-modal learning
476.8Insight-V
Our Base Model
Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models
CVPR 2025 · Table 2
Multimodal LLMs
573.6Argus-X3-8B
Llama3-8B
Argus: Vision-Centric Reasoning with Grounded Chain-of-Thought
CVPR 2025 · Table 1
Multimodal LLMs
672.6CoMemo
InternLM-1.8B
CoMemo: LVLMs Need Image Context with Image Memory
ICML 2025 · Table 2
Multimodal LLMs
771.1EVEv2.0 (Ours)
Qwen2.5-7B
EVEv2: Improved Baselines for Encoder-Free Vision-Language Models
ICCV 2025 · Table 2 (main comparison)
Multimodal LLMs
870.9HoVLE (HD)
InternLM2-1.8B
HoVLE: Unleashing the Power of Monolithic Vision-Language Models with Holistic Vision-Language Embedding
CVPR 2025 · Table 2
Human and embodied modeling
970.7LLaVA-NeXT-Full
LLaVA-NeXT
COCONut-PanCap: Joint Panoptic Segmentation and Grounded Captions for Fine-Grained Understanding and Generation
NeurIPS 2025 · VQA Benchmark Results table
Human and embodied modeling
1070.6Hiwin-LLaVA-UHD
Qwen2-7B
LLaVA-UHD v2: Exploiting Hierarchical Vision Granularity in MLLMs via Inverse Semantic Pyramid
AAAI 2026 · Table 1
Multimodal LLMs
26 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

TextVQA · accuracy (val)

15 papers, each with its best reported number for its own method; higher is better.

#accuracy ↑MethodPaperTopic
186AIF
Qwen2.5-VL-7B
Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow
CVPR 2026 · Table 2
Multimodal LLMs
285.3AdaThinking-E (ATK)
Qwen2.5-VL-7B
AdaThinking-E: One-Token Entropy Regulation for Adaptive Thinking
ECCV 2026 · Table 1
Reasoning and chain of thought
384.4Manzano-30BMANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer
ICLR 2026 · Table 2
Multimodal LLMs
483.1Molmo-72B
Qwen2 72B
Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Vision-Language Models
CVPR 2025 · Table 1
Multimodal LLMs
580.3Qwen2.5-VL + Ours
Qwen2.5-3B
When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding
NeurIPS 2025 · Table: Experimental results
Multimodal LLMs
678DocMark-8B
InternLM-8B
Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding
CVPR 2025 · Table: Performance comparisons
Multimodal LLMs
764.7VisionLLM v2
Vicuna-7B
VisionLLM v2: An End-to-End Generalist Multimodal Large Language Model for Hundreds of Vision-Language Tasks
NeurIPS 2024 · Table 1
Multimodal LLMs
856.7PreSTU stages ablation
ViT-B/16 + mT5-base
PreSTU: Pre-Training for Scene-Text Understanding
ICCV 2023 · Table 5
Human and embodied modeling
954.71TAP††
M4C
TAP: Text-Aware Pre-Training for Text-VQA and Text-Caption
CVPR 2021 · TextVQA results table
Human and embodied modeling
1053.2MoME
Vicuna-v1.5 (7B)
MoME: Mixture of Multimodal Experts for Generalist Multimodal Large Language Models
NeurIPS 2024 · Table: Comparison with state-o
Multimodal LLMs
5 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

Not affiliated with any conference, journal or the benchmark's authors. Corrections: contact@latenticlabs.com.