# Interpretability

R1-Omni is an innovative multimodal emotion recognition model that enhances model reasoning and generalization capabilities through reinforcement learning. Developed based on HumanOmni-0.5B, it focuses on emotion recognition tasks and can perform emotion analysis using visual and audio modal information. Its main advantages include strong reasoning capabilities, significantly improved emotion recognition performance, and excellent performance on out-of-distribution data. This model is suitable for scenarios requiring multimodal understanding, such as sentiment analysis and intelligent customer service, and has significant research and application value.

FakeShield is a multimodal framework designed to address two primary challenges in the field of Image Forensics Detection and Localization (IFDL): the black-box nature of detection mechanisms and the limited generalization across different tampering methods. By leveraging GPT-4o to enhance existing IFDL datasets, FakeShield has created a Multimodal Tampering Description Dataset (MMTD-Set) to train its tampering analysis capabilities. The framework includes domain label-guided interpretable detection modules (DTE-FDM) and localization modules (MFLM) that can interpret various types of tampering detection and guide localization through detailed textual descriptions. FakeShield outperforms other methods in detection accuracy and F1 scores, providing a superior and interpretable solution.

This paper investigates the problem of conceptual interpretability for video Transformer representations. Specifically, we aim to explain the decision-making process of video Transformers based on high-level spatio-temporal concepts that are automatically discovered. Previous research on concept-based interpretability has primarily focused on image-level tasks. In contrast, video models handle the additional time dimension, increasing complexity and posing challenges in identifying dynamic concepts that evolve over time. In this work, we systematically address these challenges by introducing the first video Transformer Concept Discovery (VTCD) algorithm. To this end, we propose an effective unsupervised method for identifying video Transformer representation units (concepts) and rank their importance in the model output. The obtained concepts exhibit high interpretability, revealing the spatio-temporal reasoning mechanisms and object-centric representations within black-box video models. Through joint analysis on diverse supervised and self-supervised representations, we discover that some of these mechanisms are prevalent across video Transformers. Finally, we demonstrate that VTCD can be used to improve the performance of models on fine-grained tasks.

Patchscope is a unified framework for probing the hidden representations of large language models (LLMs). It enables the interpretation of model behavior and the validation of its alignment with human values. By leveraging the model's own capacity to generate human-understandable text, we propose utilizing the model itself to explain its internal natural language representations. We demonstrate how the Patchscope framework can be used to answer a wide range of research questions about LLM computation. We show that prior interpretability methods based on projecting representations into the vocabulary space and intervening with LLM computation can be viewed as special instances of this framework. Furthermore, Patchscope opens new possibilities, such as using more powerful models to interpret the representations of smaller models and unlocking novel applications like self-correction and multi-hop reasoning.