Relative causes of code-retrieval performance decline

Determine the relative contribution of backbone capability retention and code-specific supervision coverage to the decline in CoFree-4B's code-retrieval performance, particularly on the LeetCode and Pony subsets of BRIGHT.

Background

The paper reports that CoFree-4B performs worse than its Qwen3-Embedding-4B backbone on the LeetCode and Pony subsets of BRIGHT. Contrastive-only fine-tuning on the same query–document pairs also reduces performance on these datasets, suggesting that the decline may result from continued training altering code-retrieval capabilities learned by the backbone. The authors further note that limited coverage of task-relevant code reasoning in the retained supervision could contribute, but they do not establish how much each factor matters.