Investigate APSD at Larger Model Scales

Investigate the effectiveness of Asymmetric Policy Optimization via Self-Distillation (APSD) on language models substantially larger than 8 billion parameters.

Background

The experiments assess APSD on models ranging from 1.7 billion to 8 billion parameters. Consequently, the paper does not determine whether the observed effectiveness extends to substantially larger LLMs, identifying model-scale generalization as unresolved.

References

Furthermore, due to computational resource constraints, we evaluate APSD on models with up to 8B parameters, and its effectiveness on substantially larger models remains to be investigated.

— Correctness-Dependent Asymmetric Target Distributions for Self-Distillation Policy Optimization  (2610.25973 - Yoon et al., 19 Oct 2026) in Section: Limitations