Generalization of CoVer to Additional Model Families

Establish whether the CoVer framework transfers effectively to additional language-model families, including DeepSeek-V2 and Gemma, beyond the evaluated Qwen2.5 and Llama-3.1 families.

Background

The experiments evaluate CoVer primarily with Qwen2.5-Instruct models at 7B and 14B parameters, with an additional cross-family experiment using Llama-3.1-8B-Instruct. Although the latter experiment shows improvement beyond Qwen, the paper does not establish whether the method generalizes to other model families such as DeepSeek-V2 or Gemma. Determining the extent of this transfer is therefore an unresolved empirical question concerning the model-family dependence of CoVer's gains.

References

Model family coverage. Our experiments use the Qwen2.5-Instruct family at 7B and 14B, as well as Llama-3.1-Instruct at 8B (Appendix~\ref{ssec:appendix-cross-family}); transfer to additional families (e.g., DeepSeek-V2, Gemma) is not yet established.

— Information-Gain Rewards over Diversity-Pruned Tests: GT-Anchored Verifier Co-Training for Reliable Code Generation  (2609.21208 - Nunez et al., 18 Sep 2026) in Limitations, Section 6