How Social is It? A Benchmark for LLMs' Capabilities in Multi-user Multi-turn Social Agent Tasks

Yusen Wu, Junwu Xiong, Xiaotie Deng

Keywords: LLM, Multi-user, Social agent, Evaluation

TL;DR: A benchmark for agent capabilities in multi-user multi-turn complex social interaction based on sociology and LLM study with novel dataset, pipeline and metrics.

Abstract:

Expanding the application of large language models (LLMs) to societal life, instead of primary function only as auxiliary assistants to communicate with only one person at a time, necessitates LLMs' capabilities to independently play roles in multi-user, multi-turn social agent tasks within complex social settings. However, currently the capability has not been systematically measured with available benchmarks. To address this gap, we first introduce an agent task leveling framework grounded in sociological principles. Concurrently, we propose a novel benchmark, How Social Is It (we call it HSII below), designed to assess LLM's social capabilities in comprehensive social agents tasks and benchmark representative models. HSII comprises four stages: format parsing, target selection, target switching conversation, and stable conversation, which collectively evaluate the communication and task completion capabilities of LLMs within realistic social interaction scenarios dataset, HSII-Dataset. The dataset is derived step by step from news dataset. We perform an ablation study by doing clustering to the dataset. Additionally, we investigate the impact of chain of thought (COT) method on enhancing LLMs' social performance. Since COT cost more computation, we further introduce a new statistical metric, COT-complexity, to quantify the efficiency of certain LLMs with COTs for specific social tasks and strike a better trade-off between measurement of correctness and efficiency. Various results of our experiments demonstrate that our benchmark is well-suited for evaluating social skills in LLMs.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4133