Responsibilities include:

• Contribute to developing an open-source code framework aiming to apply LLM (e.g., GPT-4) agent to perform automatic research and development (R&D) in real-world scenarios.
• Identify research questions through the implementation and then formalize questions and expose the bottleneck of current LLMs.
• Demonstrate the promising future of LLM-driven automatic R&D and propose the effective technical methods (knowledge-augmented evolving strategies).

Responsibilities include:

• Investigate the decision rules of transformer-based AI judgers by feature attribution and find that they make legal judgment predictions according to irrelevant information in the given case descriptions.
• Causally analyze the AI judgers’ nature and argue that the lack of human knowledge, the imbalance of training data, and the incomprehensiveness of testing methods impede them from learning causal relationships.
• Propose two methods to infuse knowledge into data and model’s architecture, respectively.
• Further propose 8 kinds of legal-specific attacks to complete the testing methods.
• Experimentally demonstrate that knowledge improves models’ robustness and performance.