Benchmark RadarSocialPersona
SocialPersona evaluates multimodal LLMs on recovering revealed preferences from longitudinal social-media timelines and using them in dialogue. It includes 171 user timelines, 2,597 human-verified preference tags across seven domains, and supports profile construction and response generation tasks.
- Released
- 2026-06-25
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Existing personalization benchmarks rely on explicitly stated preferences; SocialPersona tests inference from natural multimodal traces, a harder capability. It provides a reusable benchmark for measuring progress on long-horizon user modeling and personalized dialogue.
Motivation
Personalized language-model assistants are often evaluated through a memory lens: can a model recall preferences users have explicitly stated in dialogue?
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.