CiCi Yutong Cheng
Ph.D. candidate @Virginia Tech, Department of Computer Science.
About Me
Hi, I’m CiCi! I am a Computer Science PhD student at Virginia Tech. My research treats code as the representation in which intelligence can be executed, verified, and improved. I believe coding intelligence opens the last mile toward AGI: recursive self-improvement and world-model control can both be achieved via coding agents and their maintained artifacts. Specifically, my research circles around three pillars:
-
Recursive Self-Improvement: agents that improve from their own trajectories and interactions. At the system level, the harness, tools, and memory are formulated as programs and evolved as coding tasks; at the model level, the policy is fine-tuned on execution-verified rollouts, distilling inference-time search into the weights. -
Inference-Time Scaling: trading test-time compute for capability. The agent searches over program edits and rollouts under verifiers grounded in execution, and the resulting feedback serves as a scalable reward signal for both search and RL. -
Code World Model: code as the state management for environments and agents. Environments, trajectories, and agent states are represented as executable programs, casting planning as search within a verifiable simulator, toward worlds that are persistent, open-ended, and continuously improved.
Internships & Experience
-
Research intern, NEC Laboratories America, advised by Dr. Wei Cheng, 05/2026–08/2026, working on multi-objective quality–diversity optimization of compact test suites that act as behavioral gates for self-evolving agents rewriting their own harness.
-
Research intern, NEC Laboratories America, advised by Dr. Wei Cheng, 01/2026–03/2026, working on inference-time tree search to optimize code documentation for agent-oriented code reimplementation and migration.
-
Research intern, MetaGPT (Atoms.dev), 10/2024–12/2024, working on testing LLM-generated software via static and dynamic analysis in sandboxed environments.
Awards
-
2026 Future Leaders of AI, ACM AI Leadership Summit.
-
2026 ICML Golden Reviewer Award.
-
2025 CCI SWVA Cyber Innovation Scholarship.
-
2024 CCI SWVA Cyber Innovation Scholarship.
-
2024 Bitshares Fellowship.
Services
-
Program Committee, AAAI 2027.
-
Reviewer, ICML 2026, NeurIPS 2026, COLM 2026, ICLR 2027.
-
Student Organizer, 2024 DMV Security Workshop.
Resources
- awesome-world-models — a curated reading list on world models for embodied AI, tracing the generative and action-centric schools across foundation architectures, video and latent models, 3D scene generation, VLA models, simulators, and benchmarks.
Selected Publications
- NeurIPS 2026SpecForge: Recursive Self-Improvement for Coding Agents via Behavioral SpecificationsIn Proceedings of the 40th Conference on Neural Information Processing Systems, NeurIPS, 2026
- ICML 2026Escaping Whack-a-Mole: Optimizing Documentation as Repo-Specific Playbooks for Coding AgentsIn Proceedings of the 43rd International Conference on Machine Learning, ICML, 2026Adopted by NEC for repo-level test and documentation generation across large-scale Go and Java legacy codebases.
- CTINexus: Automatic Cyber Threat Intelligence Knowledge Graph Construction Using Large Language ModelsIn Proceedings of the 10th IEEE European Symposium on Security and Privacy, Euro S&P, 2025Adopted by Palo Alto Networks, ThreatConnect, and multiple other security companies for automated threat intelligence analysisTutorial presented at the PRISM Workshop of NDSS 2026.
