Extend PARSER to training and fine-tuning

Extend the theoretical framework of PARSER beyond inference-time compression to the training and fine-tuning of mixture-of-experts large language models, where the objectives and optimization dynamics differ from those of inference compression.

Background

PARSER is designed for compressing mixture-of-experts LLMs for inference. The paper notes that training and fine-tuning compression involve different objectives and optimization dynamics, so the inference-oriented theoretical analysis does not directly resolve those settings.

The authors explicitly identify extending PARSER’s theoretical framework to training and fine-tuning as an unresolved direction.

References

Extending the theoretical framework of PARSER to these different settings remains an open direction.

— Residual Sparsification via Output Importance for Compressing Mixture-of-Experts LLMs  (2609.00575 - Jung et al., 1 Sep 2026) in Section Limitations, subsection “Compression for training and fine-tuning”