I am an AI Researcher
I am a Ph.D. Candidate in Computer Science at the University of Alabama at Birmingham, co-advised by Prof. Tianyang Wang and Prof. Min Xu. I'm also a student researcher at Oak Ridge National Laboratory working with Dr. Xiao Wang since May 2025.
CV (PDF) Scholar GitHub LinkedIn WeChat xxiao [at] uab (dot) edu
My earlier research focused on understanding and shaping the intelligence and behavior of foundation models through post-training techniques. I'm currently extending this focus to agents, exploring how reinforcement learning (RL) and other post-training techniques can improve their capabilities and behavior, with a particular focus on agent memory, harness engineering, recursive self-improvement (RSI), and reasoning.
Scroll for earlier updates
MLLM Reasoning | Scaling Law | Post-training
We expose an evidence-credit gap in latent visual reasoning and train continuous latent trajectories with explicit visual supervision, grounding reasoning without changing inference.
Agentic Reasoning | Adaptive Reasoning | Agentic Routing
A lightweight router jointly selects evidence form and reasoning depth for each query, improving quality-cost trade-offs around frozen language models without changing their weights.
Post-training | Foundation Model Adaptation
Color, texture, shape, and attention-derived priors guide visual prompts across transformer layers, enabling efficient adaptation of frozen vision models to diverse recognition tasks.
Post-Training | MLLM Reasoning | Visual Alignment
By contrasting seeing and counterfactually blind responses, VIGIL discourages reliance on language priors and improves visual grounding during multimodal preference-based post-training.
Post-training | Foundation Model Adaptation
Differentiable architecture search discovers how each transformer layer should fuse prompts with image tokens, improving adaptation over a single fixed prompt-injection rule.
Post-training | Foundation Model Adaptation
Information-bottleneck filtering removes task-irrelevant low-rank directions, while a training-only graph coordinates layer updates to improve adaptation without adding inference overhead.
Post-training | Foundation Model Adaptation
Prompted Information Bottlenecks regulate what representations retain and suppress across layers, promoting task-relevant evidence, reduced shortcut reliance, and more robust frozen-model adaptation.
Language Model Pretraining | Representation Learning
We identify token embeddings collapsing into narrow cones in smaller language models and introduce dispersion loss to preserve representational diversity and improve generalization.
Post-training | Foundation Model Adaptation | Survey
A unified framework distinguishes pixel-level visual prompting from token-level prompt tuning, organizing methods, applications, benchmarks, and open challenges in adapting large-scale vision models.
Post-training | Foundation Model Adaptation
Image-conditioned prompt generation combines instance-aware and dataset-level prompts, while PCA-based propagation preserves useful prompting information as it moves through frozen transformer layers.
Post-training | Personalized Video Generation
Hybrid preference optimization balances identity preservation and motion quality, learning from paired customized videos to produce identity-consistent outputs with richer natural dynamics.
AI for Science | Exascale Pretraining | Foundation Model
A scalable vision foundation model combines efficient residual prediction and tile-wise sequence processing to enable high-resolution climate downscaling on exascale computing systems.
Talks & Media
Academic Service
Scan to connect on WeChat
