Research

I study reasoning, in models and in people, and try to build AI whose work can be checked.

The logo is the “therefore” sign ∴. Click it: its three dots are also the three terms of a syllogism.

Three questions I keep asking

Do models reason?

Classical logic lets me separate reasoning from world knowledge. On 11,000 syllogisms, five frontier LLMs fail on the same forms, say “yes” too readily, and depend on where the middle term sits in a sentence. That looks like surface patterns, not logical structure.

Paper Code ∴ Try it

What does AI do to how we write?

When people worry that their writing sounds like AI, do they start writing differently? I study how this AI-likeness pressure reshapes the way people express themselves.

AAAI 2027 · under review

Can we audit what AI makes?

LLM outputs are mostly judged by how good they look. I argue they should also be auditable, and I build systems that make this concrete, like DeepTrace, a research agent whose judge checks every citation against its source.

DeepTrace ICDM 2026 BlueSky

Before this: AI for health. I built a wearable physiological-signal database and a model that predicts continuous blood-pressure waveforms, and co-authored papers on AI in remote cardiac rehabilitation and on the philosophy of psycho-cardiology.

Questions I want to answer next

  1. Does chain-of-thought prompting remove the middle-term bias, or only hide it?
  2. Do humans and LLMs share the mechanisms behind their reasoning errors, or only the errors?
  3. Do these syntactic biases carry over to quantifiers like “most” and “few”, and to richer natural logic?

Publications

Topic
Type

Under review

  • Am I the AI? How AI-Likeness Pressure Reshapes Human Writing Expression
    Limeng Ge, Mingjia Qian
    AAAI 2027 AI for Social Impact Track Under review

2026

2024