
Xi Xiao
I am a Ph.D. Candidate in Computer Science at the University of Alabama at Birmingham, co-advised by Prof. Min Xu and Prof. Tianyang Wang.
I study how post-training shapes the intelligence and behavior of foundation models. My current research focuses on multimodal reasoning and agentic AI, using reinforcement learning (RL) and other post-training techniques to improve model and agent capabilities. I am particularly interested in agent memory, harness engineering, recursive self-improvement (RSI), and reasoning.
I'm currently a Research Scientist Intern at Oak Ridge National Laboratory, where I design and train an AI4Weather foundation model for the Genesis Mission. Previously, I was an Applied Scientist Intern at Amazon AGI Foundations, where I post-trained a 200B-parameter frontier MLLM across more than 100 compute nodes.
Seeking Research Internship opportunities in Summer 2027.
Email (xxiao [at] uab (dot) edu) Scholar GitHub LinkedIn Twitter CV WeChat
News
- One paper accepted to EACL 2027 Findings!
- Recognized as a CVPR 2026 Outstanding Reviewer.
- Two papers accepted by WACV 2027! See you in Orlando!
- Four papers accepted by NeurIPS 2026! See you in Atlanta!
- One paper accepted by TMLR!
- We are organizing the 1st Latent Visual Reasoning (LVR) Workshop at WACV 2027. See you in Orlando!
More newsLess news
- Two papers accepted by EMNLP 2026 Findings!
- Join Oak Ridge National Laboratory as a Research Scientist Intern again!
- We are organizing the Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data Workshop at SC 2026!
- One paper accepted by TMLR!
- One paper accepted by ACM MM 2026!
- Five papers accepted by ECCV 2026!
- New preprint is out: Agent Harness Engineering: A Survey.
- Recognized as an ICML 2026 Silver Reviewer.
- One paper accepted by MICCAI 2026 Spotlight π!
- Join Amazon AGI as an Applied Scientist Intern!
- One paper accepted by ICML 2026!
- Two papers accepted by ACL 2026 (one Main conference, one Findings)! See you in San Diego!
- One paper accepted by ICME 2026!
- Three papers accepted by CVPR 2026! See you in Denver!
- One paper accepted by TMLR!
- Two papers accepted by ICASSP 2026 Oral π!
- π Received a Tinker Research Grant ($5,000) from Thinking Machines Lab to support my post-training related research!
- One paper accepted by SC 2025 β Best Paper Award π Deeply honored to be the only student author.
- π Big News: Our Work Was Featured by AMD! Read the story
- Two papers accepted by WACV 2026!
- One paper accepted by AAAI 2026!
- One paper accepted by NeurIPS 2025! See you in San Diego!
- One paper accepted by WACV 2026! Round 1 Acceptance (85/1329 β 6.4%)
- One paper accepted by EMNLP 2025 (Findings)!
- Our work selected as a finalist for the ACM Gordon Bell Prize π
- One paper accepted by COLM 2025!
- One paper accepted by ACM MM 2025!
- One paper accepted by ICCV 2025!
- One paper accepted by ECML-PKDD 2025!
- Join Oak Ridge National Laboratory as a Research Scientist Intern!
- One paper accepted by ICASSP 2025!
- Started Ph.D. in Computer Science at the University of Alabama at Birmingham!
Selected Work
Organised by theme. * Equal contribution; β corresponding author.
Multimodal Reasoning & Agents
Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence
MLLM Reasoning | Scaling Law | Post-training
Work done during an internship at Amazon AGI.
Ranked #1 on Hugging Face Daily Papers.
FORGE: Form-Optimal Routing of Grounded Evidence for Frozen LLM Agents
Agentic Reasoning | Adaptive Reasoning | Agentic Routing
Trust Is Not Enough: Influence Calibration for On-Policy Self-Distillation in Agentic RL
Post-training | Agentic RL | On-Policy Self-Distillation
Staying VIGILant: Mitigating Visual Laziness via Counterfactual Visual Alignment in MLLMs
Post-Training | MLLM Reasoning | Visual Alignment
Efficient Foundation Model Adaptation
Beyond Endpoints: BΓ©zier Paths for Vision Foundation Model Adaptation
Post-training | Foundation Model Adaptation
Adapting Vision Foundation Models with Cascaded Semantics
Post-training | Foundation Model Adaptation
Layer-Specific Prompt Fusion Discovery via Differentiable Search in Vision Foundation Models
Post-training | Foundation Model Adaptation
Not All Directions Matter: Towards Structured and Task-Aware Low-Rank Model Adaptation
Post-training | Foundation Model Adaptation
Rethinking Layer-Wise Information Allocation for Vision Foundation Model Adaptation
Post-training | Foundation Model Adaptation
Dispersion Loss Counteracts Embedding Condensation and Improves Generalization in Small Language Models
Language Model Pretraining | Representation Learning
Prompt-based Adaptation in Large-scale Vision Models: A Survey
Post-training | Foundation Model Adaptation | Survey
Visual Instance-aware Prompt Tuning
Post-training | Foundation Model Adaptation
Efficient Computer Vision Systems
MagicID: Hybrid Preference Optimization for ID-Consistent and Dynamic-Preserved Video Customization
Post-training | Personalized Video Generation
ORBIT-2: Scaling Exascale Vision Foundation Models for Weather and Climate Downscaling
AI for Science | Exascale Pretraining | Foundation Model
Best Paper Award, Gordon Bell Prize Finalist
Work conducted during an internship at ORNL
In collaboration with ORNL, RIKEN, AMD, and NVIDIA
Publications
* Equal Contribution
Full list on Google Scholar; see also other publications.
Multimodal Reasoning & Agents
Staying VIGILant: Mitigating Visual Laziness via Counterfactual Visual Alignment in MLLMs
SpaMEM: Benchmarking Dynamic Spatial Reasoning via Perception-Memory Integration in Embodied Environments
Leveraging Latent Visual Reasoning in Silence
ComfySearch: Autonomous Exploration and Reasoning for ComfyUI Workflows
Dynamic Hub-and-Spoke Memory for Streaming Video Understanding
Efficient Foundation Model Adaptation
Not All Directions Matter: Towards Structured and Task-Aware Low-Rank Model Adaptation
HyperAdaLoRA: Accelerating LoRA Rank Allocation During Training via Hypernetworks without Sacrificing Performance
Layer-Specific Prompt Fusion Discovery via Differentiable Search in Vision Foundation Models
Rethinking Layer-Wise Information Allocation for Vision Foundation Model Adaptation
Personalize Your Large Vision-Language Models With In-Context Prompt Tuning
Dispersion Loss Counteracts Embedding Condensation and Improves Generalization in Small Language Models
Visual Instance-aware Prompt Tuning
MΒ²IV: Towards Efficient and Fine-grained Multimodal In-Context Learning via Representation Engineering
Sensitivity-LoRA: Low-Load Sensitivity-Based Fine-Tuning for Large Language Models
Efficient Computer Vision Systems
Learning Straight Flows: Variational Flow Matching for Efficient Generation
HIERAMP: Coarse-to-Fine Autoregressive Amplification for Generative Dataset Distillation
fMRI-LM: Towards a Universal Foundation Model for Language-Aligned fMRI Understanding
CAD-VAE: Leveraging Correlation-Aware Latents for Comprehensive Fair Disentanglement
Stochastic Interpolants via Conditional Dependent Coupling
Drift Flow Matching
Less Tokens, Better Forecasts: Sparse Residual Routing for Efficient Weather Prediction
UtoMe: Observation-Uncertainty-Guided Token Merging for Weather Foundation Models
MagicID: Hybrid Preference Optimization for ID-Consistent and Dynamic-Preserved Video Customization
MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding
Experience
Education
Awards & Honors
- [2026] CVPR 2026 Outstanding Reviewer.
- [2026] ICML 2026 Silver Reviewer.
- [2025] Best Paper Award, SC 2025 for ORBIT-2: Scaling Exascale Vision Foundation Models for Weather and Climate Downscaling.
- [2025] Finalist, ACM Gordon Bell Prize (team award) for exascale climate foundation modeling on Frontier.
- [2025] Tinker Research Grant ($5,000) from Thinking Machines Lab to support research on efficient post-training of large-scale models.
Talks & Media
- AMD story on ORBIT-2 and Frontier exascale climate modeling
- UAB News: UAB students' research wins Best Paper Award at an international computing conference
- More invited talks and interviews coming soon.
Academic Service
- Conference Reviewer
- International Conference on Learning Representations (ICLR) 2027
- IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2027
- AAAI Conference on Artificial Intelligence (AAAI) 2026, 2027
- International Conference on Machine Learning (ICML) 2026 Silver Reviewer
- International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2026
- IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026
- European Conference on Computer Vision (ECCV) 2026
- Annual Conference on Neural Information Processing Systems (NeurIPS) 2025, 2026
- ACM Multimedia (MM) 2025, 2026
- International Conference on Artificial Neural Networks (ICANN) 2025, 2026
- International Joint Conference on Neural Networks (IJCNN) 2025, 2026
- International Conference on Neural Information Processing (ICONIP) 2024, 2025
- Journal Reviewer
- Transactions on Machine Learning Research (TMLR)
- Information Fusion
- Information Sciences
- Information Processing & Management (IPM)
- IEEE Transactions on Circuits and Systems for Video Technology (TCSVT)
- IEEE Transactions on Multimedia (TMM)
- IEEE Transactions on Systems, Man, and Cybernetics: Systems
- npj Digital Medicine
- Remote Sensing of Environment (RSE)
- ISPRS Journal of Photogrammetry and Remote Sensing (ISPRS JPRS)
- Mechanical Systems and Signal Processing (MSSP)
- Engineering Applications of Artificial Intelligence
- ISPRS Open Journal of Photogrammetry and Remote Sensing
- Frontiers in Imaging
- International Journal of Human-Computer Interaction
- Journal of Magnetic Resonance Imaging
- Workshop Reviewer
- Multi-modal Foundation Models and Large Language Models for Life Sciences (FM4LS), ICML 2025, 2026; NeurIPS 2025
- Multi-Turn Interactions in Large Language Models (MTI-LLM), NeurIPS 2025
- Multimodal Foundation Models for Spatial Intelligence (MFMSI), ACM Multimedia 2025
- Meta-Agents: Managing Agents that Manage Agents, NeurIPS 2026
- Women in Computer Vision (WiCV), ECCV 2024
- Professional Activities
- Organizer, the 1st Latent Visual Reasoning (LVR) Workshop, WACV 2027
- Program Committee, Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data Workshop, SC 2026
- IEEE Graduate Student Member
Scan to connect on WeChat