4 papers accepted at NeurIPS 2026
Very proud of the work of my team and collaborators, with whom I have 4 papers at NeurIPS 2026: 3 in the main conference and one in the Datasets & Benchmarks track.
Behaviour4All: A Dependency-Aware Toolkit for In-the-Wild Facial Behaviour Analysis
Dimitrios Kollias, Chunchang Shao, Odysseus Kaloidas, Ioannis Patras
Behaviour4All is a comprehensive, open-source toolkit for in-the-wild facial behaviour analysis, integrating Face Localization, Valence-Arousal Estimation, Basic Expression Recognition and Action Unit Detection, all within a single framework. Available in both CPU-only and GPU-accelerated versions, Behaviour4All leverages 12 large-scale, in-the-wild datasets consisting of over 5 million images from diverse demographic groups.
When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment
Zhaohan Zhang, Junjie Liu, Chengzhengxu Li, Chen Shen, Xiaoming Liu, Chao Shen, Jieping Ye, Ziquan Liu, Ioannis Patras
CONFLENS is a framework for detecting shortcut reasoning in LLMs by tracking how model confidence evolves across chain-of-thought steps. We introduce the Distributional Answer Commitment Score (DACS), which estimates confidence from the entropy of the model’s probability distribution over answer commitment during reasoning. This does not require access to the ground truth or task-specific verifiers. We also incorporate the measure in Reinforcement Learning to mitigate reward models’ preference for responses that are the result of shortcut reasoning.
ViewSAM: Learning View-aware Semantics for Weakly Supervised Cross-view Referring Multi-Object Tracking
Jiawei Ge, Xintian Zhang, Jiuxin Cao, Bo Liu, Fabian Deuser, Chang Liu, Gong Wenkang, Siyou Li, Juexi Shao, Wenqing Wu, Chen Feng, Ioannis Patras
ViewSAM is a method for tracking multiple objects specified by natural language across multiple camera views, with globally consistent identities. ViewSAM bridges the gap between view-variant visual observations and view-invariant textual expressions, enabling robust cross-view referring tracking with only approximately 10% additional parameters.
CADMA: Capacity-Aware Recall Decomposition for Generative Model Assessment
Debin Meng, Zheng Gao, Ziquan Liu, Yanran Li, Ioannis Patras, Georgios Tzimiropoulos
A recall-oriented measure for generative model assessment that addresses the shortcomings of several existing measures, which do not capture well the degree to which synthetic samples cover the space of the real ones.