Training on all tokens (next token prediction) beats training only on answer tokens (critical token prediction) on small-scale reasoning benchmarks, an effect the authors attribute to noise-induced regularization.
Title resolution pending
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
-
Reasoning Bias of Next Token Prediction Training
Training on all tokens (next token prediction) beats training only on answer tokens (critical token prediction) on small-scale reasoning benchmarks, an effect the authors attribute to noise-induced regularization.