Yi (Joshua) Ren

Research

Learning dynamics

Most of my work is built on learning dynamics: looking at how a model's prediction on one example changes when it learns from another. This turns loose notions like simplicity bias into quantities we can actually measure — learning speed, compression rate — and makes a lot of otherwise puzzling training behaviour tractable.

We have used it to find better supervisory signals (Ren et al., ICLR 2022), to design fine-tuning heads (Ren et al., ICLR 2023), and to explain what really happens to an LLM during finetuning (Ren et al., ICLR 2025). The same lens turns out to be surprisingly effective for RL post-training: it explains what negative gradients do to the output distribution (Deng et al., NeurIPS 2025), how token-level hidden rewards trade exploration against exploitation (Deng et al., ICLR 2026), and why GRPO collapses so readily in agentic search (Deng et al., ICML 2026).

I suspect these observations point at something more general: that compression and Occam's Razor are not merely descriptions of good models, but constraints on how learning has to work — a view put nicely in this talk on Compression for AGI.

Now: forgetting and continual learning

Lifelong (continual) learning lets AI systems keep adapting as new knowledge becomes available. Because learning inevitably interacts with what a model already knows, getting there requires a principled account of the different forms of forgetting — catastrophic forgetting, intentional unlearning, and time-dependent forgetting. We are working to uncover the mechanisms behind these phenomena across different learning algorithms and to bring them under one analytical framework, ultimately making such systems safer, more efficient and more capable.

Earlier: iterated learning

Before that I worked on iterated learning, an idea from cognitive science: a weak bias can be amplified generation after generation when each learner is trained on the output of the previous one. I started with emergent communication, where two agents have to invent a compositional language from scratch (Ren et al., ICLR 2020), then extended it to representation learning over vision, language and even molecular graphs during my visit to Aaron Courville's group at Mila (Ren et al., NeurIPS 2023). More recently we showed that the same framework partly explains how LLMs drift under repeated self-improvement, including diversity collapse and hidden bias amplification (Ren et al., NeurIPS 2024).

Talks

• 05/2026 Happy to share some understandings of the great paper "A Theory of Generalization in Deep Learning" by Elon Litman. Here is the notion page, any comments are welcome!
• 11/2025 Happy to share some interesting works about learning dynamics at different places (Slides)
• 07/2024 Happy to share some of my understandings about the Platonic Representation Hypothesis. (Slides)
• 03/2024 Happy to give a talk at Chalmers University of Technology about the application and understanding of neural iterated learning. (Slides)