← Back to People
Researcher who co-authored a paper proving that perplexity may be unsuitable for model selection, as it does not always select for more accurate models.
1
Mentions
1
Articles
1
Outlets
Associated Topics
Coverage Patterns
How media typically covers Petar Veličković
Article Types
- research1
Media Angles
- technical1
Narrative Framing
- cautionary1
Associated AI Models
Transformer1
Articles
Most recent firstMentioned In
Referenced in coverage
Perplexity Cannot Always Tell Right From Wrong
arXivresearchcautiousFeb 4, 2026
Perplexity is an unsuitable metric for model selection because models with low perplexity can make incorrect predictions, and perplexity does not always favor more accurate models over less accurate ones.
“Researcher who co-authored a paper proving that perplexity may be unsuitable for model selection, as it does not always select for more accurate models.”
AI ResearchFoundation ModelsAI SafetyNLP