Across synthetic spurious-correlation tasks, DPO and KTO are more robust in math reasoning while SFT is stronger in document QA, and higher spuriousness does not uniformly degrade performance.
Disentangling length from quality in direct preference optimization
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
-
Assessing Robustness to Spurious Correlations in Post-Training Language Models
Across synthetic spurious-correlation tasks, DPO and KTO are more robust in math reasoning while SFT is stronger in document QA, and higher spuriousness does not uniformly degrade performance.