From Explanation to Evaluation
A Mechanistic CT Scan for Your AI Models
Powered by interaction-based explanation theory, we penetrate black boxes to enable mechanistic-level evaluation and security reinforcement.
Output accuracy = Mechanistic reliability ?
The high accuracy of the model outputs does not imply reliable underlying mechanisms.
Experientially judged data quality = The real utility of data?
AND-OR Interaction Mechanistic Evaluation
Mechanistic Evaluation for Reliable AI
Decomposign complex decision-making logic underlying billions of parameters into 50 to 150 interaction mechanisms, revealing potential risky representations.
Representation Risk Evaluation for Autonomous Driving Models
In pedestrian detection, we found that despite correct outputs, over 60% of interactions risk "cancellation effects," over 60% interactions represented overfitted patterns.
Mechanistic Evaluation of Legal LLM Judgments
While judgments were correct, the logic relied on spurious correlations, many interaction mechanisms employed by the LLM indicated spurious correlations.
Innovation
Mechanistic Evaluation & Safety Enhancement
Risk-Averse Applications
Detect hidden failure modes in safety-critical systems，even when outputs look correct.
High-Trust Decisions
Verify internal reasoning for decision-critical domains like finance and law.
Model Compression
Quantify and prevent mechanism distortion during compression for reliable edge deployment.
Identify deep-seated risk representations
Enhance model optimization efficiency
Pinpoint data causing model overfitting
