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

Brought self-improving agents into TikTok LIVE recommendation research, supporting data queries, experiment analysis, and ranking diagnosis.

Built an Auto Research Agent Harness with persistent Leads coordinating research and scoped Workers executing tasks. Integrated internal-platform MCP interfaces with semantic, episodic, and procedural memory, using snapshots and indexes for on-demand context loading.

Built an evaluation system for harness and skill updates using replayable virtual environments derived from historical task snapshots. Injected permission failures and API drift to test skill routing, API invocation, and parameter binding, achieving over 90% exact invocation accuracy.

Applied the harness to a next-day user-creator LTV project that improved adjusted live watch duration by 0.58% in A/B testing. A comparison across standard research modules showed 49% less human support time.

Collaborated with Nanjing Weiyou Education to build AI Tutor, a curriculum-grounded math tutoring system for Chinese junior-high entrance exam preparation.

Built an end-to-end pipeline from tutoring PDFs to a searchable problem bank, combining VLM/OCR parsing, structured content processing, and embedding retrieval.

Developed a tutoring agent with curriculum-grounded retrieval, structured solutions, learner memory, and a sandboxed drawing agent for SVG explanations.

Evaluated the system on 100+ questions from five years of Nanjing entrance exams; improved curriculum alignment by 32% and correctness by 9% relative to the strongest base model tested.

Built an AI-powered whiteboard that turns ambiguous ideas into structured, editable diagrams, combining generation, refinement, and recording in one platform.

Launched a multi-agent diagram generation product with 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 architecture, capturing artifacts and failure modes for systematic prompt and 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 RL stacks across Isaac Gym/MuJoCo and JAX/PyTorch into a unified JAX + MuJoCo Playground platform, standardizing configuration and reducing cross-task maintenance.

Validated Unitree G1 humanoid locomotion end to end, from policy training and simulation evaluation to sim-to-real deployment; reduced training time by 37%.

Wrote interaction documentation, usage notes, and debugging guidance to lower onboarding costs and support cross-team integration and daily development.

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, Ding Zhao

Zili Tang, Yuming Feng, Meng Guo

Yuming Feng, Yuan Tian, Erica Zhao

Yuming Feng, Christy Yang

Yuming Feng, Xinrui Jiang

Zhengxu Yan, Han Li, Yuming Feng