Research Intern, RL Team
Contributing to Kimi K3 and post-training for LLM agents, with work on Agents Last Exam and continual learning.
Jingsong Sun
孙靖淞
I work on post-training and evaluation for reliable, capable language model agents.
I am an undergraduate in Artificial Intelligence at Xi'an Jiaotong University. In Fall 2025, I was a visiting student at UC Berkeley, where I worked on tool-use agents and audio-agent evaluation with Prof. Joseph Gonzalez.
I am currently a research intern on the RL Team at Kimi (Moonshot AI). My current work focuses on post-training for LLM agents, continual learning, and tool-use evaluation.
Current
Experience
Selected work
Research
Real-Time Audio Tool Bench: Evaluating Real-Time Tool Calling in Multi-Turn Voice Interaction
A benchmark for evaluating real-time tool calling in multi-turn voice interaction, submitted to the NeurIPS 2026 Evaluations and Datasets Track.
Details in resumeTraining Dynamics of Loop Transformers on Random Graphs
We studied graph connectivity and showed that a single-layer loop Transformer can learn graph reachability through gradient descent, with provable convergence guarantees.
Details in resumeExploring the Robustness of In-Context Learning with Noisy Labels
We investigate how Transformers perform when demonstrations contain noisy labels, and find that noisy-label augmentation during training can further improve in-context learning robustness.
Published versionContact
Get in touch
Feel free to reach out to discuss research, explore potential collaborations, or simply connect.
sjs2294603251@gmail.com
WeChat: sjs2294603251