ForensicsTok: Forensics-Guided Tokenized Modeling for Image Tampering Localization

A new method called ForensicsTok reformulates image tampering localization as an autoregressive sequence generation task, aiming to overcome limitations in existing multi-modal large language model approaches, according to research submitted to arXiv in 2026 [1][2]. The work, posted on the preprint server arXiv, proposes a unified architecture that directly generates spatially grounded token sequences for precise mask prediction without intermediary supervision [2]. The researchers introduce a Token Splatting Decoder (TSD) to map tokens to binary masks via codebook-aware code smoothing, a technique designed to mitigate sharp gradients from deterministic detokenizers [2]. A Hierarchical Expert Fusion (HEF) module is also proposed to inject multi-scale features from a forensic expert model, compensating for the lack of forensic priors in standard MLLMs [2]. Experiments across six benchmarks showed that ForensicsTok substantially improves over existing MLLM-based baselines and slightly improves over strong forensic expert baselines, while exhibiting stronger robustness to perturbations [2]. The paper was submitted to arXiv on June 23, 2026 [1]. arXiv, which began on August 14, 1991, is an open-access repository of electronic preprints that are moderated but not peer-reviewed, and as of November 2024 received about 24,000 submissions per month [6]. The research appears on the site with a suite of community-developed tools under the arXivLabs framework, which allows collaborators to build features such as bibliographic explorers and code finders directly on article pages [4][5].