RESCUE-BENCH: Towards Relation-Aware Multi-Party Emotional Support Conversation Systems
Haichuan Hu, Yang Xiao, Mingni Tang, Jiawen Duan, Quanjun Zhang, Congqing He, Hao Zhang, Jiashuo Wang and 2 others
Hong Kong Polytechnic University · Vrije Universiteit Amsterdam · Nanjing University of Science and Technology · XU Exponential University of Applied Sciences
Published September 9, 2026 · on the feed September 10, 2026 · ▲ 3 · arXiv 2609.09657
The abstract
Existing emotional support conversation systems mainly focus on one-on-one seeker-supporter interactions and individual emotional states, leaving interpersonal relations in multi-party scenarios underexplored. In this work, we introduce relation-aware emotional support conversation, a new task that evaluates whether LLMs can capture and utilize the evolving dynamics of relationships to offer more effective emotional support. We construct RESCUE (Relation-aware Emotional Support Conversation Understanding and Evaluation Benchmark) from real couple and family interview conversations, containing 191 samples, 7,079 annotated turns, and 1,064.8 minutes of video. Based on rich annotations of socio-emotional and support-related dynamics, RESCUE defines six tasks that evaluate two core capabilities required for relation-aware emotional support: Relational Understanding and Relation-Sensitive Support. Experiments with ten LLMs show that current models perform relatively well on tasks relying on local emotional or intervention cues, but struggle with relation-intensive tasks such as relation pattern prediction, viewpoint prediction, and support strategy prediction. These findings reveal the limitations of current LLMs in modeling interpersonal relations and making relation-sensitive support decisions.
The Brief
2:18A short podcast summary of the paper. Written and voiced by AI.
Transcript
RESCUE-BENCH. Towards Relation-Aware Multi-Party Emotional Support Conversation Systems. By Haichuan Hu and nine others, from the Hong Kong Polytechnic University, the Vrije Universiteit Amsterdam, and two other institutions. Published on the ninth of September. Its code repository is public and real: twelve source files.
Emotional support conversation systems have mostly been built around a one-on-one exchange: a person seeking support, a supporter, and the emotional state of one individual. Conversations within a couple or a family involve more than one person, and the relations between those people shape what support would actually help. Multi-party scenarios, where those interpersonal relations matter, have been left underexplored.
The authors define a new task, relation-aware emotional support conversation. It asks whether language models can capture the evolving dynamics of relationships and use them to offer more effective support. Six tasks test two core capabilities: Relational Understanding, and Relation-Sensitive Support. The first is about reading the relationships in a conversation as they shift. The second is about acting on that reading, choosing support that fits the relationship rather than only the feelings of one speaker. A model could in principle do well at recognising emotions while failing at both, and the results suggest that is close to what happens.
The benchmark is called RESCUE, for Relation-aware Emotional Support Conversation Understanding and Evaluation Benchmark. It is built from real interview conversations with couples and families. It contains one hundred and ninety-one samples, seven thousand and seventy-nine annotated turns, and about one thousand and sixty-five minutes of video. The annotations are rich, covering socio-emotional dynamics and dynamics related to support, and the six tasks are defined on top of them.
Ten language models were tested. They do relatively well on tasks that rely on local cues, emotional ones or cues about an intervention. They struggle with the tasks that depend most on relations: predicting relation patterns, predicting viewpoints, and predicting which support strategy to use.
The pattern in the results is the finding. Current models read the emotion in a turn well enough, but falter when the right response depends on how the people in the room relate to one another. The authors present this as a limitation in modelling interpersonal relations and in making support decisions sensitive to them.
Paper Quest · 7 chapters · about 34 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Comforting one person is hard. Comforting a family is another thing entirely.
You want to know whether chat models can help a couple or a family through a hard talk, not just one upset person. Tensions between people shape what help would actually work.
The road ahead
- 1 · The problem ~6 min Chat models are tested on helping one upset person, but couples and families need help that minds the tensions between people.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min With a couple or family, support must help the whole group, track how relations shift, and decide when, whom and how to help.
- 4 · Choose your own research ~3 min
- 5 · How it works ~3 min
- 6 · The experiment ~6 min You’ll guess: Ten LLMs were tested on relation-aware emotional support tasks. Where did they struggle most?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.