Gang Xie

Hangzhou Dianzi University

Hangzhou, China

Expected graduation: June 2027

I am Gang Xie (谢刚), a master’s student in Computer Science at Hangzhou Dianzi University, expected to graduate in June 2027. My research focuses on brain-inspired multimodal affective computing, with an emphasis on robust multimodal emotion and sentiment understanding.

I also work on reliable AI systems for content safety, including multimodal safety evaluation, intent-aware harmful-content detection, and LLM-based prompt optimization. My recent work connects cognitive signals, semantic reasoning, and deployable machine-learning systems.

Research interests

Brain-inspired multimodal affective computing
Multimodal emotion recognition and sentiment analysis
Robustness, signal-noise separation, and cross-dataset generalization
LLM safety, prompt optimization, and multimodal content moderation

For research inquiries, please contact me at siegang@163.com. You can also find my work on Google Scholar, OpenReview, GitHub, and ORCID.

selected publications

IntentSafe-CN: A Multi-Layer Diagnostic Benchmark for Intent-Aware Chinese Harmful Content Detection

Gang Xie and others

In Proceedings of the AAAI Conference on Artificial Intelligence, 2027

Under review; co-first author

Abs

A multi-layer diagnostic benchmark that makes surface-form variants, targets, emotions, and communicative intent observable for Chinese content-safety evaluation.
From Enumeration to Principle: Robust Prompt Optimization via Semantic Reflection and Multi-Objective Evolution

Gang Xie and others

In Findings of the Association for Computational Linguistics: EMNLP 2026, 2026

Accepted; first author

Abs HTML

PROSE uses semantic reflection, multi-objective evolution, and ensemble-based selection to reduce shortcut learning in prompt optimization for adversarial tasks.
HPC-Net: A Brain-Inspired Isolation-Integration Framework for Robust Multi-Dataset Multimodal Emotion Recognition

Gang Xie and others

In Proceedings of the AAAI Conference on Artificial Intelligence, 2027

Under review; first author

Abs

HPC-Net separates auxiliary supervision from target optimization and uses target-conditioned integration to transfer useful representations across heterogeneous affective-computing datasets.
NeuroSIFT: A Biologically-Inspired Framework with Explicit Signal-Noise Separation for Robust Multimodal Emotion Recognition

Gang Xie and others

In 2026 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2026

Accepted; first author

Abs HTML

NeuroSIFT explicitly separates emotional signals from structured noise with biologically inspired selective inhibition and dual-stream adversarial training.
Cognition-driven Adaptive Semantic Decoding Framework for Multimodal Sentiment Analysis

Gang Xie and others

ACM Transactions on Multimedia Computing, Communications, and Applications, 2026

Accepted; student first author

Abs HTML

A cognition-driven framework for adapting to arbitrary modality combinations and aligning heterogeneous sentiment evaluation spaces.
Cognition-Guided Coupled Learning for Camouflaged Object Detection

Gang Xie and others

IEEE Transactions on Human-Machine Systems, 2026

Accepted; student second author

Abs HTML

A cognitive-visual coupling framework that uses EEG during training to improve visual-only camouflaged object detection at inference time.