Researcher & Ph.D. Student
Kangning Zhang
Ph.D. student, APEX Lab, Shanghai Jiao Tong University
I am a second-year Ph.D. student at the APEX Lab, Shanghai Jiao Tong University, advised by Prof. Weiwen Liu, Prof. Weinan Zhang, and Prof. Yong Yu.
My research focuses on LLM Post-Training, Agent Harness & Self-Evolution, and Long-Horizon Benchmarks & Evaluation—learning better policies, improving how agents execute, and verifying their capabilities on real tasks. I also work on robotics learning and recommender systems.
I received my bachelor’s degree in Computer Science from the IEEE Pilot Class at Shanghai Jiao Tong University in 2024.
I am always open to potential collaborations. If you are interested in working together, please feel free to reach out by email.
Education
Shanghai Jiao Tong University
Ph.D. in Computer Science
School of Computer Science
Shanghai Jiao Tong University
B.Eng. in Computer Science
IEEE Pilot Class
Internships
DeepExperience, Xiaohongshu
Research Intern
2025-Present
Research on agent tool calling and skills in the DeepExperience group.
Meituan
Research Intern
2024-2025
Research on generative recommendation with language models.
Shanghai Qi Zhi Institute
Research Intern
2023-2024
Research on robotics learning.
Huawei
Research Intern
2022-2023
Research on large language models for recommendation.
News
-
-
-
-
- HARNESS-R1 is available on arXiv for learning to edit executable runtime harnesses from agent failure trajectories.
-
- SKILLRISE is available on arXiv for cross-task skill evolution with agentic reinforcement learning.
-
-
- LOOPTOOL and the PROCESS REWARD MODEL SURVEY are accepted by ACL 2026 Main Conference.
- SWE-CYCLE is available on arXiv for benchmarking code agents across the complete issue resolution cycle.
- MUSEAGENT is available on arXiv for multimodal reasoning with stateful experiences.
- FINTS is available on arXiv for efficient inference-time personalization of LLMs.
- AUTOGRAPH is accepted by KDD 2025.
- ALIGNREC is available on arXiv and appears at CIKM 2024.
- 3D DIFFUSION POLICY is available on arXiv and appears at RSS 2024.
- CLICKPROMPT is available on arXiv and appears at WWW 2024.
- CODEAPEX is available on arXiv as a bilingual programming benchmark for large language models.
Publications
* Equal contribution. First/co-first-author papers are highlighted.
Scholar citations updated Sep 28, 2026; GitHub stars refresh automatically.
LLM Post-Training
6 papersTurning execution experience, teacher feedback, and sparse rewards into better policies through capability distillation, supervision attribution, and evolving training data.
arXiv preprint, 2026
Harness-S rescues failed rollouts with complementary sibling harnesses and transfers their successful behaviors into a single policy through reinforcement learning. The gains persist when the policy returns to its original harness across AppWorld, SWE-Gym, and DBBench.
arXiv preprint, 2026
VAD attributes teacher corrections to controlled visual evidence and reconstructs student-anchored supervision targets for multimodal on-policy distillation, outperforming direct privileged-view distillation across six fine-grained visual benchmarks.
Annual Meeting of the Association for Computational Linguistics (ACL), 2026
LoopTool closes the data-training loop for LLM tool use by evolving training data around a model's current weaknesses, improving robustness for multi-step tool-calling tasks.
ACM International Conference on Information and Knowledge Management (CIKM), 2026
HCGRec diagnoses when finite rollout groups cannot reach the target item, supplies the shortest useful target-prefix hint, and separates supervised credit for hinted context from GRPO credit for sampled suffix actions, reducing zero-advantage training samples from over 70% to below 20%.
Manuscript under review
Video-Thinker combines supervised fine-tuning and reinforcement learning to let multimodal models reason with videos through intrinsic grounding and captioning, generating visual evidence within the reasoning process.
Annual Meeting of the Association for Computational Linguistics (ACL), 2026
This survey reviews process reward models across process data construction, reward modeling, test-time scaling, and reinforcement learning for large language models.
Agent Harness & Self-Evolution
5 papersBuilding reusable multimodal skills and executable harnesses that help agents learn from failures, select useful experience, and improve across tasks.
Conference on Neural Information Processing Systems (NeurIPS), 2026
MMSkills represents reusable multimodal procedural knowledge for visual agents with textual procedures, runtime state cards, and multi-view keyframes, then uses branch loading to consult relevant visual evidence during decision making.
arXiv preprint, 2026
Harness-R1 trains a dedicated harness engineer to turn agent failure trajectories into validated executable runtime patches, improving target-agent success both before and after direct fine-tuning.
arXiv preprint, 2026
SkillGate trains skill selection in long-horizon agents by assigning separate credit to skill-naming and execution tokens, addressing selector credit starvation and improving a 9B policy's trial success from 40.8% to 53.2% across five agentic benchmarks.
arXiv preprint, 2026
SkillRise unifies cross-task solving and skill-document curation within reinforcement learning, enabling efficient skill evolution and transfer across progressively related tasks.
arXiv preprint, 2026
MuSEAgent distills interaction histories into stateful decision experiences and retrieves them through complementary search strategies for multimodal reasoning.
Long-Horizon Benchmarks & Evaluation
4 papersEvaluating agents through physical reconstruction, formalized research problems, and the complete software engineering cycle, alongside earlier work on programming evaluation.
arXiv preprint, 2026
PhysCodeGym evaluates coding agents on 156 video-to-Blender reconstruction tasks. Visual, motion, and physical checks—including counterfactual interventions—test whether agents build reliable physical models beyond matching appearances.
arXiv preprint, 2026
ReuF2F asks coding agents to prove or refute 244 formalized open research problems in Lean 4. A unified Leanimum-agent harness and independent verification with the Lean kernel and nanoda support trustworthy evaluation of long-horizon mathematical exploration.
arXiv preprint, 2026
SWE-Cycle evaluates code agents across environment reconstruction, implementation, test generation, and a full-cycle issue-resolution task.
arXiv preprint, 2023
CodeApex is a bilingual benchmark for evaluating large language models on programming comprehension, code generation, and code correction.
Robotics Learning
1 paperLearning robot manipulation policies with compact 3D representations and diffusion models for sample-efficient, generalizable behavior.
Robotics: Science and Systems (RSS), 2024
3D Diffusion Policy introduces compact point-cloud representations into diffusion-policy imitation learning for stronger robot manipulation generalization.
Recommender Systems
6 papersAligning multimodal content with collaborative signals, learning ID-free and cold-start item representations, and adapting language models to recommendation and personalization.
ACM International Conference on Information and Knowledge Management (CIKM), 2024
AlignRec studies representation misalignment in multimodal recommendation and introduces alignment-aware training for stronger multimodal user-item features.
European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD), 2026
MOTOR replaces item ID embeddings with learnable multimodal tokens and a token-crossing network, reducing reliance on sparse ID features.
European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD), 2026
DiffCold tackles cold-start recommendation with diffusion-based representation simulation, retrieval-enhanced aggregation, and representation alignment for cold items.
arXiv preprint, 2025
Fints performs inference-time personalization by selecting fine-grained, instance-tailored steering signals for dynamic user preferences and sparse personalization data.
ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025
This framework uses large language models to automate graph construction for recommendation, improving the graph learning substrate used by GNN-based recommenders.
The ACM Web Conference (WWW), 2024
ClickPrompt adapts language models to CTR prediction by using CTR models as prompt generators, combining semantic and collaborative signals.
