Xiaoze Liu Xiaoze Liu
Blog
Publications
LinkedIn (opens new window)
Scholar (opens new window)
GitHub (opens new window)
Twitter (opens new window)
Blog
Publications
LinkedIn (opens new window)
Scholar (opens new window)
GitHub (opens new window)
Twitter (opens new window)
  • Blog
    • Papers
    • Random thoughts
  • Papers

    • A note on knowledge graphs in the era of LLM agents
      • BibTeX
    • On cross-model-family collaboration
      • Training-level: Mutual Reinforcement Learning
      • Merging-level: When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models
      • Agentic-pipeline-level: The Vision Wormhole
      • What ties them together
    • Mutual Reinforcement Learning: Sharing experience across heterogeneous LLM policies
      • The hard problem
      • What we built (the substrate)
      • Three probes against GRPO
      • The trade-off, sketched
      • What we found
      • Where this sits
      • BibTeX
    • The Vision Wormhole: Latent-space communication for heterogeneous multi-agent systems
      • Latent communication is having a moment, but it stays inside one family
      • What we do
      • What we found
      • Where this sits
      • What I'm less sure about
      • BibTeX
    • When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models
      • Why I went looking for this
      • The attack, in one paragraph
      • What surprised me in the results
      • What I want you to take away
      • Where this sits
      • Limitations and ethics
      • BibTeX
    • A note on copyright protection in the LLM era
      • What got me started
      • SHIELD: detect the problem, then defend at inference
      • Why fixing open-weight models is nasty
      • SUV's reframing
      • Why the two papers together
      • References
      • BibTeX
  • Random thoughts

    • Vibe coding and the return of personal software
      • The App Store Tax
      • Personal Software Is Not Product Software
      • The New Default
    • A note on adapting old KG-alignment work to LLM-era problems
    • Both sides of LLM peer review at ICML 2026
      • 欲加之罪,何患无辞
      • Mean Reviewer
      • What this looked like in one of my own submissions
      • Where I'm coming from
      • Where I land now
    • Don't beg. Earn it. On a line from Eureka Seven
      • The Linguistic Anatomy of the Phrase
      • A Structural Echo
      • The Evolution of "Winning"
      • A Philosophy for Action

# Blog

Occasional notes on research, ideas, and things I find worth writing down.

# Papers

  • A note on knowledge graphs in the era of LLM agents
  • On cross-model-family collaboration
  • Mutual Reinforcement Learning: Sharing experience across heterogeneous LLM policies
  • The Vision Wormhole: Latent-space communication for heterogeneous multi-agent systems
  • When the Same Coefficients Reach Different Places: Asymmetric Realizability in Transplanting Tokenizers across Large Language Models
  • A note on copyright protection in the LLM era

# Random thoughts

  • Vibe coding and the return of personal software
  • A note on adapting old KG-alignment work to LLM-era problems
  • Both sides of LLM peer review at ICML 2026
  • Don't beg. Earn it. On a line from Eureka Seven

A note on knowledge graphs in the era of LLM agents →