Generalization of collaboration-oriented post-training across model families

Ascertain whether collaboration-oriented post-training methods (supervised finetuning on multi-turn data and reinforcement learning with PPO using conversational rewards) that increase genuine-followup rates for Qwen3.5-2B also generalize to other model families without degrading task accuracy.

Background

The authors apply collaboration-oriented post-training (SFT and PPO-based RL) to Qwen3.5-2B and observe increased genuine-followup rates, with differing effects on task accuracy. This provides preliminary evidence that interaction awareness can be trained.

They explicitly note that this study is preliminary and that it remains unknown whether similar gains would hold for other model families, indicating a need for broader validation.