Full Program – August 28, 2025 (France time / CEST / GMT+2)
Track Talks (in-person and virtual): each talk is 20 minutes + 5 min Q&A (remote presenters will join via Zoom.)
Invited Talks (in-person and virtual): each talk is 30 minutes + 10 min Q&A including Q&A (remote presenters will join via Zoom.)
Please email Behnam Hedayatnia (ben.hedayat@gmail.com) for any questions.
09:30 - 09:45 Opening Remarks
Session Chair: Behnam Hedayatnia
09:45 - 10:25 Invited Talk by Verena Rieser
Title
Intentional, Plural, Deep: The Foundations of Beneficial AI Alignment
Abstract
We talk constantly about making AI "aligned," but we rarely ask the most important questions: aligned to what, and to whom? This keynote deconstructs our current assumptions and proposes that truly beneficial AI requires moving through three distinct depths of alignment. The first is Intentionality: we must stop assuming alignment will emerge on its own and start making deliberate choices about the goals we set. The second is Plurality: we must abandon the search for a single "gold standard" of human values and engineer systems that thrive on diversity. The third and most profound is Depth: we must push beyond optimizing for clicks and convenience and instead align AI with the complex, often contradictory, nature of long-term human well-being. This is a framework for the next frontier of AI — one that builds technology to support our human nature, not exploit it.
Session Chair: Emre Can
10:25 - 10:50 Track 1 Overview
Dialog System Evaluation: Dimensionality, Language, Culture and Safety at DSTC 12 [paper link]
John Mendonca, Lining Zhang, Rahul Mallidi, Alon Lavie, Isabel Trancoso, Luis Fernando D’Haro, João Sedoc
10:50 - 11:15 Oral Paper Presentation
Neural Models and Language Model Prompting for the Multidimensional Evaluation of Open-Ended Conversation [paper link]
Michelle Elizabeth, Alicja Kasicka, Natalia Krawczyk, Magalie Ochs, Gwénolé Lecorvé, Justyna Gromada, Lina M. Rojas-Barahona
11:15 - 11:30 Coffee break
Session Chair: Alexandru Coca
11:30 - 11:55 Track 2 Overview
Controllable Conversational Theme Detection Track at DSTC 12 [paper link]
Igor Shalyminov, Hang Su, Jake Vincent, Siffi Singh, Jason Cai, James Gung, Raphael Shu, Saab Mansour
11:55 - 12:20 Oral Paper Presentation
KSTC: Keyphrase-driven Sentence Embedding and Task Independent Prompting for Filling Slot in the Generation of Theme Label [paper link]
Sua Kim, Taeyoung Jeong, Seokyoung Hong, Seongjun Kim, Jeongpil Lee, Du-Seong Chang, Myoung-Wan Koo
12:20 - 13:50 Lunch break (1 hour 30 minutes)
13:50 - 14:15 Oral Paper Presentation
The Limits of Post-hoc Preference Adaptation: A Case Study on DSTC12 Clustering [paper link]
Jihyun Lee, Gary Geunbae Lee
14:15 - 14:40 Oral Paper Presentation
CATCH: A Controllable Theme Detection Framework with Contextualized Clustering and Hierarchical Generation [paper link]
Rui Ke, Jiahui Xu, Kuang Wang, Shenghao Yang, Feng Jiang, Haizhou Li
14:40 - 15:00 Coffee break
15:00 - 15:40 Invited Talk by Milica Gašić
Title
Dimensions of intelligence
Abstract
Large language models (LLMs) have transformed the area of artificial intelligence, achieving or surpassing human performance in a number of natural language processing tasks. Despite this tremendous success, they lack the ability to model multi-turn goal-directed conversation, they are largely uncalibrated and there is little insight into the way they operate. In this talk, I will present (1) multi-turn optimisation which integrates emotion, (2) confidence-based learning based on insights from human psychology and (3), in the direction of explainability, observations from topological data analysis applied to LLMs. To conclude, I will hypothesise how combining ideas from LLMs and task-oriented systems can lead to conversational agents encompassing a large spectrum of desired properties.
15:40 - 16:40 Panel Discussion
Moderator: Luis Fernando D'Haro
Panelists: Ryuichiro Higashinaka, Laurent Prévot, Tetsuro Takahashi, Milica Gašić
Session Chair: Behnam Hedayatnia
16:40 - 17:10 DSTC Discussion