Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators

Dingkang Yang$^\star$, Dongling Xiao, Jinjie Wei, ..., Ke Li$^\dagger$, Lihua Zhang$^\dagger$

• We propose an efficient Comparator-driven Decoding-Time (CDT) framework to improve response factuality. The core philosophy is to equip target LLMs with comparators modeling different generative attributes separately during the decoding process, using logit distribution integration to facilitate next-token prediction in the factuality-robust directions.

PediatricsGPT: Large Language Models as Chinese Medical Assistants for Pediatric Applications

Dingkang Yang$^\star$, Jinjie Wei$^\star$, Dongling Xiao$^\star$, ..., Peng Zhai$^\dagger$, Lihua Zhang$^\dagger$

• This paper builds PedCorpus, a high-quality dataset of over 300,000 multi-task instructions from pediatric textbooks, guidelines, and knowledge graph resources to fulfil diverse diagnostic demands. Upon well-designed PedCorpus, we propose PediatricsGPT, the first Chinese pediatric LLM assistant built on a systematic and robust training pipeline.

Large Vision-Language Models as Emotion Recognizers in Context Awareness

Yuxuan Lei, Dingkang Yang$^\dagger$, Zhaoyu Chen, ..., Lihua Zhang$^\dagger$

• We systematically explore the potential of leveraging Large Vision-Language Models (LVLMs) to empower the CAER task from three paradigms: 1) We fine-tune LVLMs on CAER datasets, which is the most common way to transfer large models to downstream tasks. 2) We design a training-free framework to exploit the In-Context Learning (ICL) capabilities of LVLMs. 3) To leverage the rich knowledge base of LVLMs, we incorporate Chain-of-Thought (CoT) into our framework to enhance the reasoning ability and provide interpretable results.

Detecting and Evaluating Medical Hallucinations in Large Vision Language Models

Jiawei Chen$^\star$, Dingkang Yang$^\star$, Tong Wu, ..., Lihua Zhang$^\dagger$

• We introduce the first benchmark dedicated to hallucination detection in the medical domain, Med-HallMark, and provide baselines for various LVLMs. We propose the first hallucination detection model, MediHallDetector, and demonstrate its superiority through extensive experiments. We present a new hallucination evaluation metric, MediHall Score, and show its effectiveness relative to traditional metrics through qualitative and quantitative analysis.

Efficiency in Focus: LayerNorm as a Catalyst for Fine-tuning Medical Visual Language Pre-trained Models

Jiawei Chen, Dingkang Yang$^\dagger$, Yue Jiang, ..., Lihua Zhang$^\dagger$

• We are the first to centre on finetuning a small subset of the Med-VLP's inherent parameters to adapt to downstream tasks. We conduct a comprehensive series of experiments finetuning foundational components of Med-VLMs, including systematic comparisons with existing PEFT methods centred on tuning extrinsic components.

Can LLMs' Tuning Methods Work in Medical Multimodal Domain?

Jiawei Chen, Yue Jiang, Dingkang Yang (Project advising), ..., Lihua Zhang$^\dagger$

• We delve into the fine-tuning methods of LLMs and conduct extensive experiments to investigate the impact of fine-tuning methods for large models on existing multimodal models in the medical domain from the training data level and the model structure level.

MedThink: Inducing Medical Large-scale Visual Language Models to Hallucinate Less by Thinking More

Yue Jiang, Jiawei Chen, Dingkang Yang, ..., Lihua Zhang$^\dagger$

• We introduce MedThink, a novel medical construction method that effectively mitigates hallucinations in LVLMs within the medical domain.

Towards Multimodal Sentiment Analysis Debiasing via Bias Purification

Dingkang Yang, Mingcheng Li, Dongling Xiao..., Lihua Zhang$^\dagger$

• Current multimodal learning task invariably suffers from unplanned dataset biases, particularly multimodal utterance-level label bias and word-level context bias. These harmful biases potentially mislead models to focus on statistical shortcuts and spurious correlations, causing severe performance bottlenecks. To alleviate these issues, we present a multimodal counterfactual inference analysis framework based on causality rather than conventional likelihood.

Asynchronous Multimodal Video Sequence Fusion via Learning Modality-Exclusive and-Agnostic Representations

Dingkang Yang, Mingcheng Li, Linhao Qu..., Lihua Zhang$^\dagger$

• We propose a Multimodal fusion approach for learning modality-Exclusive and modality-Agnostic representations (MEA) to refine multimodal features and leverage the complementarity across distinct modalities. MEA overcomes the temporal asynchrony dilemma by capturing intra- and inter-modal element dependencies in exclusive and agnostic subspaces with the above-tailored components.

Towards Context-Aware Emotion Recognition Debiasing from a Causal Demystification Perspective via De-confounded Training

Dingkang Yang, Kun Yang, Haopeng Kuang..., Lihua Zhang$^\dagger$

• We embrace causal inference to disentangle the models from the impact of the context bias, and formulate the causalities among variables in the computer vision task via a customized causal graph. Subsequently, we present a causal intervention module to de-confound the confounder, which is built upon backdoor adjustment theory to facilitate seeking approximate causal effects during model training.

Robust Emotion Recognition in Context Debiasing

Dingkang Yang, Kun Yang, Mingcheng Li, ..., Lihua Zhang$^\dagger$

• We devise CLEF, a model-agnostic CAER debiasing framework that facilitates existing methods to capture valuable causal relationships and mitigate the harmful bias in context semantics through counterfactual inference. CLEF can be readily adapted to state-of-the-art (SOTA) methods with different structures, bringing consistent and significant performance gains.

Towards Multimodal Human Intention Understanding Debiasing via Subject-Deconfounding

Dingkang Yang, Dongling Xiao, Ke Li..., Lihua Zhang$^\dagger$

• Multimodal intention understanding (MIU) is an indispensable component of human expression analysis from heterogeneous modalities, including visual postures, linguistic contents, and acoustic behaviors. Unfortunately, existing works all suffer from the subject variation problem due to data distribution discrepancies among subjects. Here, we propose SuCI, a simple yet effective causal intervention module to disentangle the impact of subjects acting as unobserved confounders and achieve model training via true causal effect.

De-confounded Data-free Knowledge Distillation for Handling Distribution Shifts

Yuzheng Wang$^\star$,Dingkang Yang$^\star$, Zhaoyu Chen, ..., Lihua Zhang$^\dagger$

• We propose a KDCI framework to restrain the detrimental effect caused by the confounder and attempt to achieve the de-confounded distillation process. KDCI can be easily and flexibly combined with existing generation-based or sampling-based DFKD paradigms.

Correlation-Decoupled Knowledge Distillation for Multimodal Sentiment Analysis with Incomplete Modalities

Mingcheng Li$^\star$, Dingkang Yang$^\star$, Xiao Zhao, ..., Lihua Zhang$^\dagger$

• We propose a Correlation-decoupled Knowledge Distillation (CorrKD) framework for the multimodal sentiment analysis task under uncertain missing modalities.

Towards Asynchronous Multimodal Signal Interaction and Fusion via Tailored Transformers

Dingkang Yang, Haopeng Kuang, Kun Yang, Mingcheng Li, Lihua Zhang$^\dagger$

• We present a Transformer-driven Signal Interaction and Fusion (TSIF) approach to effectively model asynchronous multimodal signal sequences. TSIF consists of linear and cross-modal transformer modules with different duties.

A Unified Self-Distillation Framework for Multimodal Sentiment Analysis with Uncertain Missing Modalities

Mingcheng Li$^\star$, Dingkang Yang$^\star$, Yuxuan Lei, ..., Lihua Zhang$^\dagger$

• We propose an unified multimodal missing modality self-distillation framework (UMDF) to tackle the missing modality dilemma in the MSA task. UMDF yields robust joint multimodal representations through distillationbased distribution supervision and attention-based multigrained interactions.

Context De-confounded Emotion Recognition

Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, ..., Lihua Zhang$^\dagger$

Project | ArXiv

• We are the first to investigate the adverse context bias of the datasets in the context-aware emotion recognition task from the causal inference perspective and identify that such bias is a confounder, which misleads the models to learn the spurious correlation. In this case, we propose a contextual causal intervention module based on the backdoor adjustment to de-confound the confounder and exploit the true causal effect for model training.

Target and Source Modality Co-reinforcement for Emotion Understanding from Asynchronous Multimodal Sequences

Dingkang Yang, Yang Liu, Can Huang, ..., Peng Zhai$^\dagger$, Lihua Zhang$^\dagger$

• Inspired by the human perception paradigm, we propose a target and source modality co-reinforcement approach to achieve sufficient crossmodal interaction and fusion at different granularities.

Emotion Recognition for Multiple Context Awareness

Dingkang Yang$^\star$, Shuai Huang$^\star$, Shunli Wang, ..., Lihua Zhang$^\dagger$

Project | Supplementary | Data

• We present a context-aware emotion recognition framework that combines four complementary contexts.

Disentangled Representation Learning for Multimodal Emotion Recognition

Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, Lihua Zhang$^\dagger$

• We propose a feature-disentangled multimodal emotion recognition method, which learns the common and private feature representations for each modality.

Learning Modality-specific and -agnostic Representations for Asynchronous Multimodal Language Sequences

Dingkang Yang, Haopeng Kuang, Shuai Huang, Lihua Zhang$^\dagger$

• We propose a multimodal fusion approach for learning modality-specific and modality-agnostic representations to refine multimodal representations and leverage the complementarity across different modalities.

Contextual and Cross-modal Interaction for Multi-modal Speech Emotion Recognition

Dingkang Yang, Shuai Huang, Yang Liu, Lihua Zhang$^\dagger$

• We propose a multimodal speech emotion recognition method based on interaction awareness.

How2comm: Communication-Efficient and Collaboration-Pragmatic Multi-Agent Perception

Dingkang Yang$^\star$, Kun Yang$^\star$, Yuzheng Wang, ..., Peng Zhai$^\dagger$, Lihua Zhang$^\dagger$

Project

• Multi-agent collaborative perception has recently received widespread attention as an emerging application in driving scenarios. We propose How2comm, a collaborative perception framework that seeks a trade-off between perception performance and communication bandwidth.

AIDE: A Vision-Driven Multi-View, Multi-Modal, Multi-Tasking Dataset for Assistive Driving Perception

Dingkang Yang, Shuai Huang, Zhi Xu, ..., Lihua Zhang$^\dagger$

Project | ArXiv

• We propose an AssIstive Driving pErception dataset (AIDE) to facilitate further research on the vision-driven driver monitoring systems. AIDE captures rich information inside and outside the vehicle from several drivers in realistic driving conditions.

What2comm: Towards Communication-efficient Collaborative Perception via Feature Decoupling

Kun Yang$^\star$, Dingkang Yang$^\star$, Jingyu Zhang, ...

• We propose What2comm, a communication-efficient multiagent collaborative perception framework. Our framework outperforms previous approaches on real-world and simulated datasets by addressing various collaboration interferences, including communication noises, transmission delay, and localization errors, in an end-to-end manner.

Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception

Kun Yang$^\star$, Dingkang Yang$^\star$, Jingyu Zhang, ...

Project | ArXiv

• Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. We propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner.