Jyu'Log

Jun Yu

I am a researcher/builder currently interested in increasing human<->AI bandwidth. As models get more capable, I think the bottleneck shifts to how well we can convey intent, share context, and judge results. I am building Nodex: a shared workspace for humans and agents (or, if that helps, an open-source alternative to Notion+Codex).

Previously, I was a research assistant at the NUS CuriOSity Research Group, advised by Prof. Zhenkai Liang (2024–2025). My research focused on safety in multi-step LLM and agent workflows. I also worked on Stable Signer, a model-watermarking framework for image provenance, where I contributed to the framework design and experiments, balancing the different needs of model providers, regulators, and users.

I received my B.Eng. in Computer Science and Technology from Zhejiang University in 2024. At ZJU, I led the product and engineering team at QSC Mobile, working on GPA, grade, and timetable features for students. I’m interested in how technical choices shape the way people actually use a product.

Selected Writing

  1. Weak-to-Strong Exploitation: Weak LLMs Orchestrating Aligned Models Toward Harmful Goals

    Weak LLMs as malicious orchestrators: a threat model, three mechanisms, causal evaluation, and future research on RL.

  2. Stable Signer: Balancing Three Stakeholders’ Interests in Image Provenance

    A message can change a generative model's decoder. The harder question is who chooses the change, who can read it, and who gets to make a claim from it.