Independent Researcher

Brian K Chen

I’m an independent AI researcher with a background in mathematics and statistics. I work on model efficiency and reasoning, which I see as two sides of the same coin. Much of my research is about understanding how language models use the context they’re given, and whether we can make that process more efficient without losing their ability to reason.

During my PhD, I mainly worked on connections between in-context learning and model weights, as well as methods for adapting models more efficiently. Recently, I’ve been looking at how memory and attention limit what transformers can do. I’m also studying diffusion language models as a possible alternative, particularly how to make their decoding more efficient.

I enjoy bringing what I’ve learned from working with researchers in industry into my own research. That experience has given me a clearer sense of the problems that are most pressing in practice and helps guide the questions I choose to work on. I believe theory should guide experiments and help us understand what’s happening at the frontier, rather than serve as an excuse to ignore it.

I did my PhD in Computer Science at the National University of Singapore with Kenji Kawaguchi, after studying Mathematics and Statistics at Oxford. My work has appeared at ICML and ICLR.

Brian K Chen by the waterfront
Context & memory Efficient reasoning Diffusion language models

Selected publications

Equal contribution

  1. ICLR2026
  2. npj Biomedical
    Innovations
    2026
  3. ICML2024
  4. NeurIPSWorkshop2024

Current research

Recent work on the limits of memory and attention, probing reasoning, and efficient language model inference.

Submitted manuscripts ICLR 2027 submissions

  1. The Value of Memory in Language Models: Joint Limits on Language Model Expressivity

    Brian K Chen, Cui Kai

  2. How Far Can Sliding Window Attention See?

    Maxime Meyer*, Brian K Chen*, Nikolaj Murašov, Caroline Chaux, Vincent Y. F. Tan

  3. Reasoning Under Intervention: Learned Soft Prefixes as Behavioral Probes

    Brian K Chen

  4. Speculative Correction: Draft-Then-Refine Decoding for Diffusion Language Models

    Brian K Chen, Chong Wu, Kenji Kawaguchi

Work in progress5 projects
  • Bottlenecks in Hybrid Sliding-Window Attention

    Brian K Chen, Maxime Meyer

  • The Expressivity of the Indexer Controls the Expressivity of the Language Model

    Brian K Chen, Maxime Meyer

  • BFLA: Block-Filtered Long-Context Attention Mechanism

    Wu Chong, Brian K Chen, et al.

  • Index-Predictor: Does DeepSeek Need as Many Static Heads?

    Brian K Chen, Wu Chong

  • Amortized Memory Systems for Long-Horizon Agentic Harnesses

    Brian K Chen

Get in touch

Let’s talk research.

For research discussions and collaboration, you can reach me by email.

e0694208@u.nus.edu