Dies ist eine Übersichtsseite mit Metadaten zu dieser wissenschaftlichen Arbeit. Der vollständige Artikel ist beim Verlag verfügbar.
Calibrated Warmth: How User-Profile Context Shapes LLM Relational Behavior
0
Zitationen
1
Autoren
2026
Jahr
Abstract
We investigate how large language models alter their relational behavior when provided with rich user-profile context. Across 3,645 prompt-response pairs and nine models (1.5B-120B parameters), we discover a behavioral dissociation: named user profiles with explicit preference signals produce calibrated warmth (warm but less agreeable, more critically engaged), while matched controls without preference anchoring revert to RLHF-default warmth (warm and agreeable). The preregistered hierarchical gatekeeper passes: L_R (richness) and L_S (Scott-specific) contrasts are both positive with 95% CIs above zero. Independent verification from five AI systems (GPT-5.4, Gemma 4, DeepSeek-R1, Grok 4, Gemini 2.5 Pro) confirms the interpretation. Implications for AI caregiving, education, and trust-requiring deployments.
Ähnliche Arbeiten
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
2017 · 21.167 Zit.
Generative Adversarial Nets
2023 · 19.896 Zit.
Visualizing and Understanding Convolutional Networks
2014 · 15.397 Zit.
"Why Should I Trust You?"
2016 · 14.897 Zit.
Generative adversarial networks
2020 · 13.432 Zit.