I am an AI safety researcher focused on building reliable and aligned AI systems, especially for improved reasoning, continual learning, and agentic deployment. I specialize in post-training (RL/SFT), identifying agentic misalignment and safety vulnerabilities, and improving the factuality and groundedness of frontier models.

Aligned and Safe Reasoning in LLMs: identifying hidden bias in process reward models and mitigating reward hacking and downstream policy misalignment in large reasoning models (LRMs), safer process supervision for policy learning (GRPO, RLHF, PPO) and policy search. I am interested to extend effective process-supervision for better reasoning and planning as my primary research interest.

Superalignment and Multi-LLM System: improving scalable oversight and weak-to-strong generalization with multiple LLMs, multi-LLM reasoning and LLM evaluation using uncertainty-aware answer selection for diverse LLMs.

Interpretability and Multimodal Robustness: reducing hallucinations in vision-language models through refined textual embeddings, studying diffusion language models for better reasoning, and improving robustness in multimodal systems.

I also have background in multi-agent reinforcement learning, robotics, and speech applications.