← Back to Home

Research Projects

Study on Semantic Constraints in LLM Decoding

Ongoing Research Project

Dec 2024 - Present

University of Toronto

Contributions

  • Surveyed SOTA LLMs on syntactical constrained decoding performance
  • Developed LLM Output Debugger with llama cpp backend that allows better user interaction in exploring LLM output guidance
  • Evaluated SOTA LLMs on tool usage performance
llama-cpp quantization self-hosted LLM automated reasoning

Study on LLM In-Context Learning using Psychological Theory

Ongoing Research Project

Jan 2025 - Present

University of Toronto

Contributions

  • Evaluate similarity of SOTA LLM on the psychological level between AI agents and humans in terms of planning in academic problem solving
  • Suggest new guideline for ways of thinking in prompting LLM for In-Context Learning for high-difficulty tasks
  • Propose self iterative ICL framework based on the theory discovered
Agents In-Context Learning Human-Computer Interaction OpenAI Cognitive Psychology