Jonathan Light 🔮

Hey there 👋

I'm Jonathan

I build LLM agents that improve themselves through search and reinforcement learning.

My research asks how a model can propose its own problems, explore solutions with inference-time search, and learn from what it finds. Lately that has meant post-training coding agents.

I am a Visiting Researcher at Caltech, advised by Yisong Yue. I am pursuing a Ph.D. in Computer Science at Rensselaer Polytechnic Institute, advised by Santiago Paternain and previously by Ziniu Hu, and expect to graduate in 2027. I received my M.S. from the University of Chicago, where I worked with Haifeng Xu and Dacheng Xiu. Before that, I trained as an economist.

I have also spent time in industry at Microsoft Research, NEC Laboratories America, and Asari.ai, working on post-training coding agents and reinforcement learning for LLMs.

My research has been featured in the State of AI Report by Air Street Capital, appearing in the 2024 edition and 2026 edition.

Get in touch

Research

I think of self-improvement as a loop with three moving parts. I've been gradually working on improving different parts of the loop, and the whole loop runs on a harness and a set of environments.

Self-improvement loop: the model proposes tasks, explores trajectories, and learns from them. Harness & environments Propose what to solve next Explore how to solve it better Learn what to keep Model Tasks Trajectories
The loop closes: what the model learns changes which tasks are worth proposing next.
  1. 1. Propose
    what to solve next

    Curriculum and task generation: choosing the problems whose solutions would improve the model the most right now.

    Actor-Curator FrogNano
    more on Curriculum Learning →
  2. 2. Explore
    how to solve it better

    Inference-time search and scaling: spending compute to find better solutions than a single forward pass would.

    DISC Scattered Forest Search Strategist PIANIST
    more on Inference Scaling →
  3. 3. Learn
    what to keep

    Feedback internalization: turning the attempts into a signal and absorbing it, into weights or into memory.

    ROFT Strategist Dataset Distillation for RL
    more on RL Post-Training →
  4. Runs on a harness and environments

    The substrate the loop executes in: how the agent is scaffolded and which worlds it acts in.

    JAZ AvalonBench

News

Show older news (7)

Selected Publications

View all
Figure from Actor-Curator: Co-adaptive Curriculum Learning via Policy-Improvement Bandits for RL Post-Training

Actor-Curator: Co-adaptive Curriculum Learning via Policy-Improvement Bandits for RL Post-Training

COLM 2026
Zhengyao Gu*, Jonathan Light*, Raul Astudillo, Ziyu Ye, Langzhou He, Henry Peng Zou, Wei Cheng, Santiago Paternain, Philip S. Yu, Yisong Yue

A scalable, fully automated curriculum-learning framework for RL post-training of LLMs. Learns a neural curator that selects training problems to directly maximize expected policy improvement, formulated as a non-stationary bandit with regret guarantees. Delivers +28.6% on AIME2024, +30.5% on ARC-1D, and up to 80% speedup over the strongest baseline. Outstanding Paper Award and Oral (top 3.9% of accepted papers) at the ICLR 2026 Logical Reasoning Workshop.

RL Post-Training Curriculum Learning LLM Reasoning
Code Website PDF
Figure from Strategist: Self-improvement of LLM Decision Making via Bi-Level Tree Search

Strategist: Self-improvement of LLM Decision Making via Bi-Level Tree Search

ICLR 2025
Jonathan Light, Min Cai, Weiqin Chen, Guanzhi Wang, Xiusi Chen, Wei Cheng, Yisong Yue, Ziniu Hu

LLM + MCTS hybrid that learns to optimize its own decision process through self-play and strategy refinement. Outperforms RL and existing LLM agents, achieves human-level performance in complex games. Highlighted in the State of AI Report 2024 as a breakthrough in autonomous agent design.

Tree Search Multi-Agent & Games LLM Agents
Code Website PDF