AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

ImpText: A Benchmark and Tool-Augmented Framework for Implicit Text Reasoning

Litao Guo, Jinsong Zhou, Shuaibo Li, Man CHEN, Xinli Xu, Zixin Zhang, Harold Haodong Chen, Ying-Cong Chen

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs

Read the paperCheck your idea against 137,217 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 ImpText, a benchmark for implicit text (physically deformed, visually camouflaged, or cognitively suggested text) and a tool-augmented framework for reasoning over it in MLLMs.

Problem

MLLMs excel at standard text extraction but struggle with implicit text, where malicious content is concealed via physical deformation, visual camouflage, or cognitive suggestion to evade content moderation.

Approach

A benchmark targeting implicit text concealment techniques, paired with a tool-augmented framework that lets MLLMs invoke tools to recover and reason over hidden text. Based only on the abstract; the specific mechanism is not described.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add tool augmentation alone (ablation)Qwen3-VL 8B · ImpText-BenchF1: 26.09 -> 28.62No clear effect
add knowledge warm-up SFT (ablation)Qwen3-VL 8B + tools · ImpText-BenchTMS: 10.63 -> 35.79Improved
add oracle tool exploration (ablation)ImpText-Reader · ImpText-BenchTMS: 35.79 -> 46.83Improved
add boundary-aware RL (GRPO) (ablation)ImpText-Reader · ImpText-BenchF1: 78.49 -> 87.97Improved
filter training data to 0<P<1 instead of full dataset (ablation)Semantic Reasoner SFT · ImpText-BenchTMS: 39.06 -> 46.83Improved
decouple enhancer and reasoner training instead of joint (ablation)ImpText-Reader · ImpText-BenchTMS: 8.38 -> 35.79Improved
plain GRPO on Qwen3-VL 8B vs baseQwen3-VL 8B · ImpText-BenchTMS: 10.22 -> 27.40Improved
relax NED threshold from 0.5 to 0.9 (ablation)TMS metric · ImpText-BenchTMS: 48.26 -> 72.39 (ImpText-Reader)No clear effect
threshold tau=0.5 versus 0.1 or 0.9 for agreement with semantic judge (ablation)TMS metric · ImpText-Bench 1,000 subsetdecision agreement: 0.909 at tau=0.5 vs 0.840 / 0.827Improved

Experimental setup

Benchmarks: ImpText

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy