A single-pass reward model trained with a max-consistency (TD) loss provides fast reward-guided decoding with competitive alignment quality.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Towards Cost-Effective Reward Guided Text Generation
A single-pass reward model trained with a max-consistency (TD) loss provides fast reward-guided decoding with competitive alignment quality.