LogicLens: Visual-Logical Co-Reasoning for Text-Centric Forgery Analysis

TL;DR

LogicLens tackles the challenge of robust, explainable text-centric forgery analysis by unifying detection, grounding, and explanation into a single autoregressive report. It introduces the Cross-Cue Thought Chain for iterative cross-modal reasoning between visual artifacts and textual cues, and optimizes outputs with a GRPO-based multi-task reward. The RealText dataset and PR² pipeline provide scalable, cognitively aligned annotations to train and evaluate the model. Empirical results show state-of-the-art macro-F1 on RealText, strong zero-shot generalization to public benchmarks, and robustness on dense-text datasets, demonstrating the value of visual-textual co-reasoning for forgery defense.

Abstract

Sophisticated text-centric forgeries, fueled by rapid AIGC advancements, pose a significant threat to societal security and information authenticity. Current methods for text-centric forgery analysis are often limited to coarse-grained visual analysis and lack the capacity for sophisticated reasoning. Moreover, they typically treat detection, grounding, and explanation as discrete sub-tasks, overlooking their intrinsic relationships for holistic performance enhancement. To address these challenges, we introduce LogicLens, a unified framework for Visual-Textual Co-reasoning that reformulates these objectives into a joint task. The deep reasoning of LogicLens is powered by our novel Cross-Cues-aware Chain of Thought (CCT) mechanism, which iteratively cross-validates visual cues against textual logic. To ensure robust alignment across all tasks, we further propose a weighted multi-task reward function for GRPO-based optimization. Complementing this framework, we first designed the PR$^2$ (Perceiver, Reasoner, Reviewer) pipeline, a hierarchical and iterative multi-agent system that generates high-quality, cognitively-aligned annotations. Then, we constructed RealText, a diverse dataset comprising 5,397 images with fine-grained annotations, including textual explanations, pixel-level segmentation, and authenticity labels for model training. Extensive experiments demonstrate the superiority of LogicLens across multiple benchmarks. In a zero-shot evaluation on T-IC13, it surpasses the specialized framework by 41.4% and GPT-4o by 23.4% in macro-average F1 score. Moreover, on the challenging dense-text T-SROIE dataset, it establishes a significant lead over other MLLM-based methods in mF1, CSS, and the macro-average F1. Our dataset, model, and code will be made publicly available.

Figures (5)

• Figure 1: LogicLens(pink area) achieves state-of-the-art performance on our RealText benchmark across detection (D), grounding (G), and explanation (E), with M-F1 (macro-average F1 over the three tasks) as the unified evaluation metric.
• Figure 2: Overview of the LogicLens framework. Our unified generative model takes an image and a prompt to autoregressively produce a structured textual forgery analysis Report. Central to this process is the Cross-Cues-aware Chain of Thought (CCT), an iterative reasoning mechanism that analyzes and cross-validates visual and logical cues. The entire model is optimized via a GRPO-based weighted multi-task reward. The resulting machine-readable report is a versatile output that can power various downstream applications, such as the professional forensic report shown on the right.
• Figure 3: The architecture of our PR² (Perceiver, Reasoner, Reviewer) pipeline. After an initial data collection stage, our multi-agent system generates annotations through a collaborative, iterative process. The Perceiver drafts an analysis, the Reasoner structures it to target CCT & analysis report, and the Reviewer validates its quality, initiating a refinement loop if necessary. This cycle, indicated by the solid $\bm{\rightarrow}$ and dashed $\bm{\dashleftarrow}$ feedback loops, ensures the final output is a high-fidelity, structured JSON annotation.
• Figure 4: Training curves for the GRPO-based alignment stage, showing smoothed metrics. The simultaneous increase in task-specific rewards (Grounding and Format) alongside a decrease in entropy and mean output length demonstrates the effectiveness of our weighted multi-task reward. This strategy successfully guides the model toward a more stable, confident, and efficient policy, rather than simply a verbose one.
• Figure 5: Qualitative comparison of artifact grounding and explanations across different methods. LogicLens demonstrates superior performance, accurately identifying both visual artifacts and logical cues. Shaded regions indicate the localized areas.