Xinke Kong 孔信轲

I am currently a first-year Master's student at Tianjin University, supervised by Prof. Changqing Zhang. I received my Bachelor's degree in Computer Science and Technology from Tianjin University in 2025.

My research focuses on LLM reasoning reinforcement (RLVR) and test-time training (TTT), including: 1) self-rewarding mechanisms for large reasoning model optimization (under review); 2) test-time training (ICLR'25).

Email  /  Google Scholar  /  Github

[2025-01] One paper accepted by ICLR, thanks to all co-authors!

An EM-based test-time training framework that alternates Critic calibration (E-step) and policy optimization (M-step), enabling reasoning LLMs to keep improving post-deployment with sustained gains over 350+ steps.

A conservative entropy minimization approach for test-time adaptation that explicitly models prediction uncertainty to prevent over-confidence and avoid common collapse modes.

National Scholarship (1%) 2022

Conference Reviewer: ICML 2026, NeurIPS 2026