**Original paper:** [https://arxiv.org/abs/2608.13168v1](https://arxiv.org/abs/2608.13168v1)

# Summary: 2026-08-13_12-32-33Z_WhichLLMIsYourIdealCompanion_EvaluatingEmotionalCo.md
Saved: 2026-08-13 21:44
Source: 2026-08-13_12-32-33Z_WhichLLMIsYourIdealCompanion_EvaluatingEmotionalCo.md
Model: None

---

## Summary  
The paper proposes evaluating large language models as emotional companions using adult attachment theory, introducing a benchmark called ECBench that measures anxiety and avoidance traits across four relational scenarios in both friendship and romance. It aims to provide a psychological framework for selecting LLMs that support healthy emotional interactions. The study assesses 32 LLMs through dialogue‑quality metrics to link model behavior with attachment tendencies.  

## Key Contributions  
- Finding 1: Attachment anxiety/avoidance traits of LLMs can be quantified using the Experiences in Close Relationships‑Revised (ECR‑R) scale.  
- Finding 2: ECBench provides a standardized benchmark across emotional support, collaborative tasks, conflict resolution, and social guidance in both friendship and romantic contexts.  
- Finding 3: Prompt engineering with secure‑attachment cues can mitigate maladaptive attachment behaviors, improving model responsiveness.  

## Methodology  
The authors applied adult attachment theory to LLM evaluation by measuring anxiety and avoidance via the ECR‑R scale. They designed ECBench with four interaction scenarios across two relationship types, using eleven dialogue‑quality metrics such as empathy, coherence, emotional attunement, and response relevance. The benchmark was administered to 32 LLMs, generating multi‑turn dialogues that were scored against these metrics.  

## Results  
The study found a positive correlation between high attachment anxiety scores and models that provide excessive reassurance but lack depth, indicating anxious attachment tendencies. Models with avoidance traits responded minimally or deflected conversation. Prompting with secure‑attachment cues reduced avoidance by approximately 15 % on average, while the top‑performing models aligned closely with secure attachment profiles.  

## Significance  
This work bridges psychology and artificial intelligence, offering a method to select LLMs that support healthy relational dynamics and potentially reduce loneliness or improve mental health outcomes for users seeking emotional companionship.  

## Related Concepts  
- Adult Attachment Theory  
- Experiences in Close Relationships‑Revised (ECR‑R)  
- Emotional Companion Benchmark (ECBench)  
- Dialogue‑quality metrics  
- Prompt engineering  
- Secure/Anxious/Avoidant attachment styles