Controllable and Reliable AI Systems are AI frameworks designed to produce stable, trustworthy, and interpretable outputs across diverse real-world environments. Our research focuses on improving robustness, controllability, and reliability in complex reasoning tasks and high-stakes applications.

Logical Consistency & Calibration are methods for improving reasoning coherence, confidence estimation, and factual reliability in large language models. We study techniques that reduce hallucinations, maintain logical consistency across long reasoning chains, and align model confidence with prediction accuracy.

Lightweight Structural Modifications for LLMs are efficient architectural adaptations that enhance reasoning, verification, and inference capabilities without requiring large-scale retraining. Our work explores scalable and modular approaches that improve model performance while preserving computational efficiency.

Efficient Expert-Domain Adaptation is an approach for adapting foundation models to specialized scientific and expert domains using minimal supervision and efficient fine-tuning strategies. We focus on transferring domain-specific knowledge while maintaining strong general reasoning and language understanding abilities.

Reinforcement Learning is a research area that studies adaptive decision-making through interaction and feedback, with a strong focus on applying RL to NLP and LLMs.

Constraint-Guided Learning is a research area that incorporates logical, structural, and domain-specific constraints into learning and inference processes to improve consistency, robustness, and controllability. Our work focuses on integrating prior knowledge and formal rules into machine learning systems for more reliable and safer predictions.

Brain Foundation Models is a research area that explores biologically inspired architectures and neural representations for building scalable, adaptive, and generalizable AI systems. Our research investigates computational principles motivated by brain mechanisms to improve learning efficiency and adaptability.

Neuro-symbolic Approach is a research area that combines neural learning with symbolic reasoning and formal knowledge to enable interpretable, logically grounded, and reliable AI systems. We explore frameworks that integrate data-driven learning with explicit reasoning and structured knowledge representation.

Structured Prediction typically focuses on predicting structured outputs from machine learning models, which involves complex outputs such as sequences, trees, and graphs. We aim to expand the structured prediction problem to multiple (input, output) pairs in real-world applications, ensuring set consistency.

EBM (Energy-Based Model) is a form of generative model (GM) imported directly from statistical physics to learning. EBM provides a compatibility score of (x,y) pairs capturing their hidden relationships, which provides complex dependencies hidden in real data.

SPEN (Structured Prediction Energy Networks) combines the flexibility of neural networks with the benefits of structured prediction, enabling end-to-end training and inference for structured outputs. We are interested in using SPEN as a teacher model (SEAL, 2022) that can teach general feedforward neural net models.

AI for Science is a research area focused on applying machine learning systems to scientific discovery, physical modeling, and data-driven understanding across diverse scientific domains.

3D Generative Modeling is a research area that explores generative frameworks for structured 3D representations, human motion, and spatial dynamics through diffusion and sequential modeling approaches. We study models that capture complex geometric and temporal dependencies for realistic and controllable 3D generation.