Value representations from token logits, sequence perplexity, and text generation are all sensitive to prompt and option changes, and their correlation with model behavior in value scenarios is weak.
ElNokrashy, Mai AlKhamissi, and Mona Diab
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Revisiting LLM Value Probing Strategies: Are They Robust and Expressive?
Value representations from token logits, sequence perplexity, and text generation are all sensitive to prompt and option changes, and their correlation with model behavior in value scenarios is weak.