Stable Domain Adaptation for Stronger Base Models

Determine how to obtain stable performance improvements when adapting the Qwen3.5-9B base model to the dedicated role-playing task, despite its already high initial character-consistency performance and the limitations of the current training data and pipeline.

Background

The paper reports that the training pipeline produced significant gains on the Qwen3 series but did not yield comparable gains when applied to Qwen3.5-9B. Because Qwen3.5-9B already has a high initial level of character consistency, the current distilled training data and adaptation pipeline were unable to produce stable further improvements. The authors attribute this difficulty partly to the training data being outdated relative to the capability improvement from Qwen3 to Qwen3.5, leaving stable adaptation to stronger base models unresolved.

The authors identify stronger data distillation, On-Policy Self-Distillation (OPSD), and other lossless domain-knowledge learning paradigms as directions for addressing this adaptation challenge. The unresolved problem is therefore to develop an adaptation method that can reliably improve role-playing performance on increasingly capable base models without sacrificing their existing capabilities.