Research map / Vision-language and multimodal / Multimodal LLMs

MMTIT-Bench: A Multilingual and Multi-Scenario Benchmark with Cognition-Perception-Reasoning Guided Text-Image Machine Translation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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.

Introduces MMTIT-Bench, a 1,400-image multilingual multi-scenario end-to-end text-image machine translation benchmark, and CPR-Trans, a cognition-perception-reasoning CoT data paradigm that fine-tunes Qwen2.5-VL-3B/7B for TIMT.

Problem

End-to-end text-image machine translation lacks evaluation resources covering diverse real-world scenes and low-resource languages, and effective reasoning-oriented (CoT) data designs for TIMT are underexplored.

Approach

A unified long-CoT training format that sequences scene cognition, text perception (OCR-aligned layout description), and translation reasoning, generated by a VLLM pipeline from human OCR and translation annotations, instead of cascaded OCR-then-translate or language-only distilled reasoning.

Key results

• On MMTIT-Bench, CPR-Trans fine-tuned Qwen2.5-VL-7B reaches 83.98 (Gemini judge, other2en) versus 53.98 for the original Qwen2.5-VL-7B and 68.40 for direct SFT, and exceeds Gemini 2.5 Flash's 82.94 on that metric.
• At 3B, CPR-Trans scores 82.00 (Gemini judge other2en) vs 72.31 for Simple CoT and 64.98 for Direct translation.
• Ablation: the full CPR-Trans improves over the OCR+translation baseline by an average of +11.2 under the Gemini judge and +8.2 under the Qwen3-VL judge.

Surprising findings

• A fine-tuned 7B model (Qwen2.5-VL-7B with CPR-Trans) reaches judge scores competitive with or above much larger models such as Gemini 2.5 Flash on the benchmark's Gemini judge metric.
• Distillation from a VLLM thinking teacher underperforms the simpler Simple CoT (OCR plus translation) setting, attributed to noisy, redundant reasoning traces.

Limitations

• Benchmark and training data release is promised only upon acceptance; not yet released.
• Training data is generated and judged largely by Gemini 2.5 Flash and Qwen3-VL models, so evaluation relies on LLM judges that may share biases with the annotators.
• Evaluation covers only Other-to-English and Other-to-Chinese directions and 14 languages.

Experimental setup

Datasets: MMTIT-Bench training split (manual annotations) SynthDog synthetic image-text pairs

Benchmarks: MMTIT-Bench

Metrics: LLM-judge score (faithfulness, fluency, readability, termino COMET

Baselines: Gemini 2.5 Flash Qwen3-VL-235B-A22B-Instruct Qwen3-VL-235B-A22B-Thinking dots.vlm1 MiMo-VL Seed1.6 Seed1.6-Thinking MinerU2.5+Qwen3 Dots.ocr+Qwen3

Builds on: Qwen2.5-VL COMET SynthDog

Closest papers by meaning

• Don't Show Pixels, Show Cues: Unlocking Visual Tool Reasoning in Language Models via Perception Programs (CVPR 2026 · similarity 0.89)
• JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (NeurIPS 2024 · similarity 0.89)
• Toward Cognitive Supersensing in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.89)
• Multilingual Diversity Improves Vision-Language Representations (NeurIPS 2024 · similarity 0.88)
• Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.88)
• Cross-Lingual Text-Rich Visual Comprehension: An Information Theory Perspective (AAAI 2025 · similarity 0.88)