Selective Tie Learning
A debiasing method in preference optimization that uses theoretical analysis to identify the most effective tie data (pairs of responses that are equally preferred), achieving better debiasing with fewer ties.
Timeline (last 90 days)
A recent study proposes "selective tie learning", which uses theoretical analysis to screen out the most effective ties, thereby achieving debiasing effects comparable to or even better than uniform selection with fewer ties.
The experiments cover log-linear policies, neural network models, and large language models, and the results show that compared with uniform tie selection, the proposed selection rule achieves comparable effectiveness in reducing spurious-correlation dependence but requires fewer ties.
All Facts (2)
A recent study proposes "selective tie learning", which uses theoretical analysis to screen out the most effective ties, thereby achieving debiasing effects comparable to or even better than uniform selection with fewer ties.
50%UnverifiedThe experiments cover log-linear policies, neural network models, and large language models, and the results show that compared with uniform tie selection, the proposed selection rule achieves comparable effectiveness in reducing spurious-correlation dependence but requires fewer ties.
50%