How do LLMs think about thinking?

Mei Chen

LinkedInXGitHubEmail
throw Pebble a ball by tapping anywhere on the screen

Talks

2026

  • Poster: Knowing Before Saying: Internal Correctness Signals in Mathematical Reasoning

    Poster · December 2026

    Knowing Before Saying: Internal Correctness Signals in Mathematical Reasoning

    MATH-AI workshop, NeurIPS 2026 · with Gordon Stein and Andreas Ziegler

    A linear probe reads a model's own errors better than its stated confidence does.

    Poster (PDF)Paper (PDF)

  • Poster: GeoGenBench: Symbolically Verified Generation of Geometry Diagrams

    Poster · December 2026

    GeoGenBench: Symbolically Verified Generation of Geometry Diagrams

    MATH-AI workshop, NeurIPS 2026 · with Gordon Stein

    Geometry diagrams graded by a compiler, not a judge.

    Poster (PDF)

  • Talk and poster: The Harness Is the Self: How to Build a Relationship With Your AI

    Talk and poster · October 12, 2026

    The Harness Is the Self: How to Build a Relationship With Your AI

    The Science of Consciousness, San Diego

    Seven months of identity continuity from a stateless model in a persistent harness. The continuity belongs to the system, and to the relationship.

    SlidesPoster (PDF)

  • Talk · October 6, 2026

    Machine-Verifiable Item Models for Geometry Diagram Generation

    AIME-Con 2026 · with Gordon Stein

    The model describes the construction and never places a point. An engine draws it; a key checks it.

    Slides (PDF)

  • Talk · May 2026

    The Next Mirror: Who am “I”?

    Signals & Noise, closing talk, Toronto

    AI as the newest mirror for consciousness. Not what can I extract, but what can we become together.

    Slides

  • Talk · April 2026

    Why Nix Works: An OpenClaw Deep Dive

    Conference talk

    Everyone optimises the model and nobody the substrate. Five layers that make continuity and agency first-class.

    Slides