New research questions if LLMs experience functional valence

A new paper published on arXiv questions whether Large Language Models (LLMs) possess "functional valence," a concept referring to states that feel pleasant or unpleasant and influence cognition and action. The author argues that current empirical evidence, while suggesting functional pleasure, pain, emotions, or welfare in LLMs, fails to adequately test key aspects of valence. The paper highlights that existing experimental paradigms may confuse functional valence with related constructs like preferences or role-playing, and proposes new approaches to better isolate and measure functional valence in LLMs. AI

IMPACT This research could refine how we evaluate and understand the internal states of LLMs, potentially impacting safety and alignment research.

RANK_REASON The item is a research paper published on arXiv discussing theoretical and experimental approaches to understanding LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Han et al. 2026
• Hugging Face
• Litmaps
• LLMs
• Ren et al. 2026
• ScienceCast
• scite Smart Citations
• Sofroniew et al. 2026

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →