Post-Training Increments Can Transfer Across Models

tuvllms · x · 2026-07-09

A recent study suggests that post-training shifts can be compressed into transferable logit deltas, capable of moving across different model scales and families. The author notes that migrating a delta extracted from a 1.5B Qwen model to Llama-3-8B resulted in a 15% boost on GSM8K.

The post connects this finding to concepts like task arithmetic, the Master Key Hypothesis, and fine-tuning/alignment transfer. It stresses that such transfers don't require models from the same family—just a simple vocabulary bridge to achieve logit grafting.

Original post →

More from Research

Research channel →