Output prefilling with a structured prefix steers LLMs to produce cleaner first tokens in MCQA, raising accuracy and calibration over standard first-token probability.
JMLR (2023) 14 S
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
-
Improving LLM First-Token Predictions in Multiple-Choice Question Answering via Output Prefilling
Output prefilling with a structured prefix steers LLMs to produce cleaner first tokens in MCQA, raising accuracy and calibration over standard first-token probability.