Token-Level Off-Policy Learning for Faithful Generation Under Distribution Shift
TOPL trains models to judge token correctness instead of copying off-policy target tokens.
The paper says that token-level signal improves faithful generation under distribution shift. In summarization experiments, TOPL generalizes across 11 out-of-distribution datasets better than sequence- and token-level baselines. The authors also report transfer to machine translation. Their ablations say the token-level objective is the key part; sequence-level versions do not show the same benefit. HF Daily Papers' note
The paper says that token-level signal improves faithful generation under distribution shift. In summarization experiments, TOPL generalizes across 11 out-of-distribution datasets better than sequence- and token-level baselines. The authors also report transfer to machine translation. Their ablations say the token-level objective is the key part; sequence-level versions do not show the same benefit. HF Daily Papers' note
score 5