Coursework: Deep Learning, Natural Language Processing with Deep Learning, Human-Centered LLMs, Apps with LLMs Inside, Understanding Users

Visiting scholar, Mechanical Engineering. Multi-agent collaborative manipulation and long-horizon quadrupedal pushing.

Coursework: Introduction to Computation, Data Structure and Algorithm, Multi-Agent Systems, Thermodynamics and Statistical Physics, Quantum Mechanics

Designed a TikTok Live long-term value ranking target for US For You mixed ranking, modeling user-creator "app-open reason for live" labels on T+1.

Built an end-to-end pipeline converting traditional tutoring PDFs into AI-ready corpora, using VLM/OCR parsing, structured Markdown normalization, corpus splitting, problem-bank construction, and embedding retrieval for grounded tutoring.

Designed a curriculum-grounded agent system for Chinese math exams with structured solve outputs, problem-bank RAG, user-level memory, and a sandboxed drawing subagent for Python-generated SVG explanations.

Built an exam-paper evaluation harness over recent regional exams, showing our agentic pipeline improved curriculum fit by 32% and correctness by 9% over the base frontier model.

Launched a multi-agent diagram generation product that turns intent into editable diagrams through clarification, generation, canvas refinement, and one-click recording; reached 10K+ impressions and 300+ users.

Designed a telemetry-driven iteration loop around the Planner-Judge-Executor-Validator agent architecture, capturing artifacts and failure modes to support systematic prompt/model debugging and workflow refinement.

Reduced generation cost by about 87% and latency by about 60% through DSL compilation, model routing, and prompt caching while preserving output quality.

Consolidated fragmented training stacks across simulation backends and ML frameworks into a unified JAX + MuJoCo Playground platform, standardizing configuration and reducing cross-task maintenance.

Validated the end-to-end ML workflow from policy training and simulation evaluation to deployment handoff, reducing training time by 37% versus the previous fragmented workflow.

Documented configuration, debugging, and deployment handoff workflows to lower onboarding costs, standardize experiment transfer, and support cross-team integration.

Built a reproducible evaluation harness for small-model LLM alignment, measuring task performance and optimization behavior across SFT, DPO, full fine-tuning, and LoRA under matched data and metric settings.

Tested GPT-2 (124M) on paraphrase detection and sonnet generation, showing that DPO gains are limited in small-model, low-data regimes and depend on task format, data scale, and SFT baseline strength.

Quantified adaptation trade-offs: full fine-tuning outperformed LoRA by about 3% Acc/F1, while LoRA enabled up to 2x larger batch sizes under the same compute budget.

Trained persona-steerable review summarization with SFT (asymmetric distillation) plus RLAIF (PPO), gaining ~70% on average across key metrics vs. the base model.

Built a full evaluation system with rule-, LLM-, and user-based metrics, analyzing how these signals correlate with each other and with the model’s observed effectiveness in practice.

Formulated conceptual novelty estimation as a pairwise ranking problem rather than only binary classification, raising the evaluation ceiling from 32% F1 to 75% agreement.

Built a retrieval-augmented LLM training pipeline where FAISS neighbors and LLM-generated similarity reports form comparison context; used SFT to teach Qwen3-4B the pairwise novelty-judgment format, then applied DPO on preference pairs to optimize ranking decisions.

Benchmarked compact encoder models, fine-tuned Qwen3-4B, and zero-shot frontier models, showing that task-specific supervision and semantically aligned comparison pairs matter more than raw model scale.

Proposed a three-level hierarchical MARL framework for long-horizon quadrupedal pushing: RRT plus a centralized adaptive policy generates object subgoals, a decentralized goal-conditioned policy coordinates quadrupeds, and a pretrained locomotion policy executes local motion, improving success rate by 36.0% and reducing completion time by 24.5% over best baselines.

Designed an Occlusion-Based (OCB) reward to guide agents toward better contact points on the target object, improving average success rate by 49.5% in ablation.

Enabled Sim2Real transfer using friction-based domain randomization and deployed the policy on real Unitree Go1 Team, pushing T-shaped blocks and cubes for 10-20 m at 0.5 m/s.

Published at IEEE ICRA 2025.

Proposed a hybrid search algorithm that decomposes mode sequences and navigation paths to generate feasible cluttered-scene pushing trajectories, reducing execution time by 18% and improving trajectory smoothness by 67.8% over uniform path partitioning.

Proposed a fast, scalable multi-agent contact-mode switching algorithm; used the intermediate value theorem to prove 100% collision-free switching and establish a theoretical time bound for a single mode switch.

Validated the planning-control pipeline in PyBullet with 100% multi-scenario success and approximately 0.03 m simulation tracking error, then deployed it on ROS hardware with less than 0.05 m tracking error.

Published at Robotics: Science and Systems (RSS) 2024.

Yuming Feng, Chuye Hong, Yaru Niu, Shiqi Liu, Yuxiang Yang, Wenhao Yu, Tingnan Zhang, Jie Tan, Ding Zhao

Zili Tang, Yuming Feng, Meng Guo