Study on Reasoning Ability Degradation of Qwen3-4B: Why Does Model Generalization Decline After Fine-Tuning?

This article conducts an in-depth analysis of the reasoning ability degradation phenomenon of the Qwen3-4B model after fine-tuning on specific downstream tasks, explores the trade-off between model generalization and specialization, and provides important references for LLM fine-tuning practices. Recently, the Qwen3-4B-Reasoning-Degradation project on GitHub has attracted widespread attention from the community. This project systematically studies this degradation phenomenon and has important warning significance for developers and researchers who are currently or planning to perform model fine-tuning.

Qwen3 is an open-source large language model series launched by the Alibaba Cloud Tongyi Qianwen team. The 4B parameter version provides excellent basic capabilities while maintaining a small size, making it suitable for deployment in resource-constrained environments. It has demonstrated good reasoning, code generation, and mathematical calculation abilities during the pre-training phase. However, after fine-tuning in specific domains, the base model often exhibits "catastrophic forgetting" or capability drift phenomena, which this study explores.

Core findings of the study: The general reasoning ability of Qwen3-4B decreases significantly after fine-tuning on specific tasks. Specific manifestations include: 1. The performance on the fine-tuned task improves, but the integrity of the reasoning chain for unseen new tasks decreases; 2. The error rate for complex multi-step reasoning problems increases significantly; 3. The performance on cross-domain transfer tasks declines, confirming that generalization ability is impaired.

Analysis of degradation causes: 1. Weight update conflict: The new task objectives are inconsistent with the optimization direction of the original general capabilities, and parameter updates overwrite or distort the original knowledge representation; 2. Data distribution shift: The data distribution of downstream tasks differs greatly from the pre-training data, leading the model to over-adapt to the specific distribution; 3. Singularity of optimization objectives: Standard fine-tuning only focuses on task-specific loss functions and lacks explicit constraints to maintain general capabilities.

Practical recommendations for the degradation phenomenon: 1. Progressive fine-tuning: Use parameter-efficient fine-tuning methods such as small learning rates, regularization, or LoRA; 2. Mixed training data: Incorporate data related to general capabilities; 3. Continuous evaluation of general capabilities: Regularly test on independent evaluation sets; 4. Multi-task fine-tuning: Fine-tune on multiple related tasks simultaneously; 5. Explore alignment technologies: Such as RLHF, etc.

Industry significance: Reminds enterprises to pay attention to the maintenance of the model's overall capabilities during fine-tuning to build reliable AI systems. Future research directions: Develop more intelligent fine-tuning algorithms, establish comprehensive evaluation frameworks, explore modular architectures, etc., to achieve a balance between specialization and generalization.