DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation
We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities.
Key points
- On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals.
- Yet we find that training on fewer tokens can outperform full-token OPD, challenging the intuition that more supervision improves learning.
- Existing disagreement-based criteria ignore probability scale: tokens assigned negligible probability by both models, termed low-low tokens, can receive large log-ratio rewards and hinder learning.
- Token-level evidence reveals that DIAL-OPD filters high-reward tokens with limited reasoning value while preserving supervision critical to reasoning correctness.
Sources (1)
- [1]DIAL-OPD: Learning More from Fewer Tokens in On-Policy DistillationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 10:36 AM
We propose DIAL-OPD, a token-selection method that bridges log-probability and probability spaces by weighting reward magnitude with the logarithmic mean of teacher and student probabilities.
On-policy distillation (OPD) supervises student-generated trajectories with token-level teacher signals.
Extractive summary: sentences quoted from the sources.