SimPER aligns LLMs by directly optimizing inverse perplexity of chosen versus rejected responses, reporting strong benchmark results without a reference model or loss hyperparameters.
Discard any snails that are not alive
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
-
SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters
SimPER aligns LLMs by directly optimizing inverse perplexity of chosen versus rejected responses, reporting strong benchmark results without a reference model or loss hyperparameters.