Songcheng Cai

Songcheng Cai

Master Student at University of Waterloo.

Hi, this is Songcheng. I am a second-year MMath student at David R. Cheriton School of Computer Science, University of Waterloo, supervised by Professor Krzysztof Czarnecki. In addition to my work with my advisor, I am fortunate to collaborate closely with Professor Wenhu Chen. Previously, I obtained my B.Eng. in Information Engineering from Zhejiang University. During my undergraduate studies, I was a research intern at Toyota Technological Institute at Chicago (TTIC), where I was fortunate to be advised by Karen Livescu and Hongyuan Mei.

My research focus is on building reliable language and multimodal foundation models for complex, real-world tasks. I am currently pursuing three parallel directions: (1) Rigorous evaluation of foundation models, including benchmarks that capture realistic task complexity through live environments, executable verification, and multimodal reasoning, as well as meta-evaluations that scrutinize the metrics themselves to make sure they reflect what we actually care about; (2) Agentic post-training with rich automatic feedback, where I design pipelines that turn execution outcomes, automatic verifiers, and self-evolving reward models into scalable supervision, reducing the reliance on expensive human preferences; (3) Vision-language-action (VLA) models for autonomous driving, where I focus on multimodal fusion of LiDAR and camera to ground language-conditioned decisions in accurate and robust 3D perception.

🔥 News

  • 2026.07: RewardHarness is accepted by COLM 2026.
  • 2026.05: Rewardharness is released. Check out our paper and code!
  • 2026.04: ClawBench is released. We find frontier LLM agents still struggle with everyday tasks on real websites (Claude Sonnet 4.6: 33.3%). Check out our paper and code!
  • 2026.04: SWE-QA-Pro is accepted by Findings of ACL 2026. Check out our paper and huggingface collections.
  • 2026.01: VisCoder2 is accepted by ICLR 2026.
  • 2025.10: VisCoder2 is released. Check out our paper and huggingface collections!
  • 2024.10: LFTQA-Eval is accepted by EMNLP 2024. Check out our paper and code.
  • 2024.09: Started my thesis-based MMath program at University of Waterloo.