LLMs share task-specific attention heads across prompting styles, with activation strength explaining performance differences and failures arising from competing representations.
cc/paper_files/paper/2020/file/ 1457c0d6bfcb4967418bfb8ac142f64a-Paper
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Shared Lexical Task Representations Explain Behavioral Variability In LLMs
LLMs share task-specific attention heads across prompting styles, with activation strength explaining performance differences and failures arising from competing representations.