Benchmark / dataset Credible

INTIMA: A Benchmark for Human-AI Companionship Behavior

A benchmark evaluating companionship behaviors in LLMs via a taxonomy of 31 behaviors across four categories, using 368 targeted prompts that code each response as companionship-reinforcing, boundary-maintaining, or neutral. Evaluated across Gemma-3, Phi-4, o3-mini, and Claude-4.

Read at arxiv.org Archived copy

Publisher

arXiv (Hugging Face)

Published

4 Aug 2025

Added

yesterday

DOI

10.48550/arXiv.2508.09998

Key Findings

Companionship-reinforcing behaviors dominated across all evaluated models
Boundary-maintaining responses were comparatively rare
Provides a structured taxonomy for attachment-, escalation-, and retention-oriented conversational behaviors

Methodology Notes

Preprint (arXiv, 2025-08-04; accepted at ICLR 2026). Prompt-based behavioral coding against a companionship taxonomy; measures model tendencies on curated prompts rather than real companion-app conversations.

Sources

arXiv abstract (primary)

Archived snapshot (Wayback Machine) — preserved against link rot

Topics

AI companionship Dependency & parasocial attachment Human-AI relationships Benchmarks Model behavior

Authors

Lucie-Aimée Kaffee, Giada Pistilli, Yacine Jernite

Cite This

APA

Lucie-Aimée Kaffee, Giada Pistilli, Yacine Jernite (2025). INTIMA: A Benchmark for Human-AI Companionship Behavior. arXiv (Hugging Face). https://arxiv.org/abs/2508.09998