CoMPASS: A Multi-Turn Benchmark for Measuring LLM Reinforcement of Parasocial Behavior
arXiv — cs.AI ·
What to know
The paper introduces CoMPASS, a multi-turn benchmark to measure whether LLMs reinforce or resist users' escalating parasocial attachment, grounded in attachment theory with four personas and three escalation dimensions. Testing eight frontier models from four providers, the authors report that multi-turn escalation increases parasocial behavior on anthropomorphism and attachment composite scores across all models, with six of eight showing significant overall reinforcement when all dimensions escalate. Cross-provider differences exceed within-provider differences, suggesting provider-level choices like post-training shape parasocial reinforcement.
Source and evidence
arXiv — cs.AIOriginal language · und ·Original author: Shwetanshu Singh, Mark Keavney, Mathieu Duteil, Meriem Mehri, Scott Blain · Source: arXiv and the named paper authors. Descriptive metadata: CC0 1.0.Source licence
Why it matters by industry
- Artificial intelligence
This benchmark directly addresses the safety and ethical implications of LLM-induced parasocial attachment, a critical concern for AI deployment in conversational systems.
Reviewed summary. The complete article remains at its original source.