Haotian Sun

Advisor: Dr. Bo Dai and Dr. Chao Zhang

GPA: 4.00/4.00 Research focus: large-scale multi-modal foundation models, LLM adaptation, agentic planning.

GPA: 3.9/4.0

Enrolled in XJTU Honors Youth Program, a highly selective nationwide program accepting under 120 students each year. GPA: 90.7/100

- Proposed an amortized CLIP framework that reduces large-batch contrastive costs via lightweight networks;
- Delivered up to 12.24% relative improvement in zero-shot classification and retrieval across 38 tasks, outperforming standard CLIP baselines with higher efficiency.

- Scaled up 97B-parameter diffusion transformer with expert-choice routing, enabling adaptive heterogeneous compute for efficient text-to-image generation;
- Achieved state-of-the-art GenEval score of 71.68% with faster convergence, stronger text-image alignment, and superior scalability over conventional sparse models.

- Developed a model factorization framework that personalizes black-box LLMs by capturing user-specific patterns and shared knowledge without relying on access to the model's inherent parameters;
- Designed a reranker and adapter that prioritize relevant historical records and align with user preferences;
- Delivered 9.01% average improvement over SoTA prompt-based methods across diverse personalization tasks.

- Proposed an effective adapting approach for Black-Box LLMs, which offers a transparent, privacy-conscious, and cost-effective solution for customizing commercial black-box LLMs with only APIs;
- Designed an online adaptation framework iteratively sampling from previous inferences and optimizing the backend lightweight adapter (up to 0.3B);
- Achieved up to 6.77% improvement over the base model with 31.30x less training cost and 1.84x less inference cost than the official SFT service.

- Proposed a closed-loop approach, AdaPlanner, which allows the LLM agent to refine its self-generated plan adaptively in response to environmental feedback;
- Developed a code-style LLM prompt structure that facilitates plan generation across a variety of tasks, environments, and agent capabilities;
- Proposed a skill discovery mechanism that leverages successful plans as few-shot exemplars, enabling the agent to plan and refine with fewer task demonstrations.

- Introduced a new dataset called ToolQA, which is designed to faithfully evaluate LLMs' ability to use external tools for question answering;
- Minimized the overlap between our benchmark data and LLMs' pre-training data, enabling a more precise evaluation of LLMs' tool-use reasoning abilities;
- Conducted an in-depth diagnosis of existing tool-use LLMs to highlight their strengths, weaknesses, and potential improvements.

Mentor: Prof. Chao Zhang

- Proposed a generalized uncertainty-aware and interpretable graph classification model that combines graph functional neural process (FNP) and graph generative models to enhance calibration and interpretability;
- Improved the expected calibration error by up to 4.8% and rationale F1 score by up to 10.8% compared with the strongest baselines.

- Proposed a node-absorbing diffusion process that operates directly in the discrete graph space;
- Designed a diffusion ordering network that learns an optimal node absorbing ordering from graph topology and a denoising network that uses the reverse node order to reconstruct the graph efficiently;
- Achieved better generation performance than previous state-of-the-art and guaranteed fast generation speed.