Model Compression and Reasoning Consistency: Are Distilled Models Truly "Reasoning Correctly"?

This study focuses on the reasoning consistency issue of compressed models after knowledge distillation, exploring whether compressed models truly understand the essence of problems when giving correct answers, rather than just mimicking surface patterns. Through multi-dimensional evaluation methods such as GradCAM (attention visualization), CKA (representational similarity analysis), and calibration analysis, it finds that test set accuracy cannot guarantee reasoning consistency, different distillation strategies have a significant impact on consistency, and puts forward practical suggestions for improving the distillation process and model selection, emphasizing the need to ensure model reasoning quality while improving efficiency.

Knowledge distillation technology for large language models can reduce inference costs while maintaining performance, but a core issue is often overlooked: do correct answers from compressed models stem from true understanding? In high-risk scenarios such as medical diagnosis and legal analysis, models need to make decisions based on correct logic and evidence. If they only "guess correctly" relying on surface features, it may lead to serious consequences. Therefore, reasoning consistency (correct answers with reasonable reasons) is a key consideration for the deployment of compressed models.

The study constructs a systematic evaluation framework to analyze reasoning behavior from three dimensions:

1. GradCAM: Compare the attention heatmaps of teacher and student models; if reasoning is consistent, the attention distribution will be similar;
2. CKA: Measure the similarity of intermediate layer representations to evaluate whether the student model has internalized the teacher's abstract representations;
3. Calibration Analysis: Check the matching degree between confidence and accuracy through reliability diagrams and ECE indicators; overconfidence is a signal of reasoning defects.

Key findings include:

1. Correct Answer ≠ Correct Reasoning: Some student models have accuracy comparable to the teacher model, but their attention distributions differ greatly, relying on spurious correlations or surface features;
2. Impact of Distillation Strategies: Output matching easily leads to surface learning, while feature/relation distillation is more effective in maintaining reasoning consistency;
3. Task Complexity Adjustment: Compressed models are more prone to "chain breaks" in complex reasoning tasks, requiring strict verification mechanisms.

Improvement suggestions:

1. Multi-objective Optimization: Combine output matching, feature alignment, and calibration constraints in the loss function;
2. Layered Distillation: Design differentiated distillation strategies for each layer;
3. Adversarial Verification: Test robustness using adversarial samples;
4. Human-in-the-Loop: Manually review heatmaps of key samples.

Model selection framework: In addition to accuracy, check GradCAM heatmaps, calibration curves, and CKA representation alignment.

Limitations: Focuses on visual and text classification tasks; reasoning consistency evaluation for generative tasks (such as translation and text generation) remains to be explored; interpretation tools (GradCAM/CKA) have limitations.

Future directions: Develop fine-grained reasoning path tracking tools, apply causal inference methods, and establish standardized reasoning consistency benchmark datasets.

Model compression makes large models easier to deploy, but efficiency improvement should not come at the cost of reasoning quality. It is necessary to examine the model's "thinking process" and ensure that compressed models make correct judgments based on correct reasons before they can be safely deployed in real complex scenarios.