Verifiable Social Reasoning for LLM Assistants
Researchers introduced Fuse, a multi-agent simulation framework for studying user-mediated social reasoning in LLM assistants. Fuse involves a target agent with a hidden motive interacting with other agents, including a user agent, who then consults the evaluated assistant to infer the target's motive, providing verifiable ground truth. The framework was validated through a human study with 24k annotations. Fuse was applied to 12 LLMs, revealing that user mediation compounds social reasoning difficulty, LLMs show systematic sensitivity to biased user framing, models may require more details than humans for correct predictions, and longer conversations do not always improve performance.
A new multi-agent simulation framework called Fuse enables verifiable evaluation of LLM assistants' social reasoning in user-mediated consultation settings. By constructing scenarios with hidden motives and verifiable ground truth, Fuse was validated with 24k human annotations and applied to 12 LLMs. Key findings include increased difficulty due to user mediation, sensitivity to biased framing, higher detail requirements than humans, and inconsistent benefits from longer conversations.
Fuse provides a controlled environment where social reasoning can be evaluated with ground truth, addressing the lack of verifiable social properties in real-world consultations. The framework isolates factors such as user mediation, framing bias, detail requirements, and conversation length, enabling systematic analysis of LLM social reasoning capabilities.
This research highlights a gap in current LLM evaluation for social advice applications, suggesting that real-world user interactions may degrade performance compared to direct reasoning tasks. The findings could inform development of more robust assistants for social consultation and user-facing AI products.
For companies deploying LLM assistants in social advice or customer-facing roles, Fuse offers a method to benchmark and improve social reasoning reliability. Insights into user mediation and framing bias can guide product design to mitigate errors and enhance user trust.
Future work may extend Fuse to more complex social scenarios, additional LLMs, and integration with other evaluation frameworks. The observed sensitivity to biased framing and detail requirements suggests potential improvements in prompt design, user interaction handling, and model training for social reasoning.