Atif Quamar

Visiting Researcher at MBZUAI

About

I am currently a Visiting Researcher at MBZUAI, where I work under the supervision of Prof. Junpei Komiyama. Previously, at UC San Diego, I worked on latent reasoning in vision-language models, and at Purdue University, my work was focused on inference-time alignment of language models. I also founded Insituate, where we shipped agentic systems to banks and judiciary. I earned my bachelors degree in Computer Science and Biosciences from IIIT Delhi in 2024.

I like to understand how foundation models reason, and how we can make that reasoning more efficient and trustworthy. My work focuses on improving reasoning in settings that involve language and vision, while ensuring these models remain safe, reliable, and aligned with human values.

I am currently looking for research opportunities in both academia (PhD) and industry for a 2026 start.

Interests

Multimodal Reasoning
Trustworthy AI
AI Alignment and Safety
Reinforcement Learning

Education

B.Tech in Computer Science and Biosciences, 2020-2024
IIIT - Delhi

Publications

Quickly discover relevant content by filtering publications.

M. Atif Quamar, M. Areeb, N. Sharma, A. Shreekumar, J. Rosenthal, M. Kuznetsov, M. Ozgur Ozmen, Z. Berkay Celik (2025). Adaptive Blockwise Search: Inference-Time Alignment for Large Language Models.

Project arXiv

Y. Shao, K. Zhou, Z. Xu, M. Atif Quamar, S. Hao, Z. Wang, Z. Hu, B. Huang (2025). Learning Modal-Mixed Chain-of-Thought Reasoning with Latent Embeddings. Under review at The 43rd International Conference of Machine Learning (ICML 2026).

Project arXiv

M. Atif Quamar, M. Areeb, M. Kuznetsov, M. Ozgur Ozmen, Z. Berkay Celik (2025). STARS: Segment-level Token Alignment via Rejection Sampling in Large Language Models. Frontiers in Probabilistic Inference: Sampling Meets Learning Workshop at NeurIPS 2025.

Project arXiv

M. Atif Quamar, M. Areeb (2025). Logit–Entropy Adaptive Stopping Heuristic for Efficient Chain-of-Thought Reasoning. Efficient Reasoning Workshop at NeurIPS 2025.

arXiv OpenReview

N. Sharma, M. Atif Quamar, P. Xie (2025). Decoding Histone Modification Signatures of Non-Coding RNAs via Foundation Models. Multi-modal Foundation Models and Large Language Models for Life Sciences Workshop at NeurIPS 2025.

OpenReview

Experience

Visiting Researcher

MBZUAI

Feb 2026 – Present Abu Dhabi, UAE

Working with Dr. Junpei Komiyama on advancing the reasoning capabilities of language models.

Research Intern

University of Virginia

Oct 2025 – Feb 2026 Charlottesville, Virginia, United States

Worked on developing a drift-resilient memory framework for code-execution agents using KL-constrained adapter updates, mitigating embedding distribution shifts during online learning to reduce unsafe code generation without compromising task success rates.

Research Intern

University of California - San Diego

Jul 2025 – Oct 2025 San Diego, California, United States

Worked on multimodal reasoning – interleaving text and visuals within the chain-of-thought, enabling models to “think” with sketches, diagrams, and images. This work bridges language and vision to solve complex problems with richer, more interpretable reasoning.

Research Intern

Purdue University

Feb 2025 – Jul 2025 West Lafayette, Indiana, United States

Worked on inference-time alignment method that outperforms Best-of-N decoding by over 30%, while reducing reward model calls by 20%. Aligned LLMs in reducing harmlessness, improved reasoning and positive sentiment generation.

Co-Founder

Insituate

Sep 2023 – Feb 2025 New Delhi, India

Built agentic software for the Supreme Court of India, Mizuho Bank, PNC Bank and Indian High Courts.

Insituate is a no-code platform that enables companies to make custom AI agents for industry-specific needs, and allowing them to productionize these copilots 10x faster securely on their data.

Projects

Adaptive Blockwise Search : Inference-Time Alignment for Large Language Models

M. Atif Quamar ¹ , M. Areeb , N. Sharma , A. Shreekumar , J. Rosenthal , M. Kuznetsov , M. Ozgur Ozmen , Z. Berkay Celik

Adaptively focuses computation on the most critical early tokens during LLM decoding, boosting alignment performance across multiple tasks compared to Best-of-N and fine-tuning.

Adaptive Blockwise Search : Inference-Time Alignment for Large Language Models

Learning Modal-Mixed Chain-of-Thought Reasoning with Latent Embeddings

Y. Shao , K. Zhou , Z. Xu , M. Atif Quamar , S. Hao , Z. Wang , Z. Hu , B. Huang

Modal-mixed chain-of-thought lets a VLM interleave text with compact latent visual “sketches”, using a diffusion-based latent decoder with SFT+RL training to boost vision-intensive reasoning while adding only modest inference overhead.

Learning Modal-Mixed Chain-of-Thought Reasoning with Latent Embeddings

STARS - Segment-level Token Alignment via Rejection Sampling in Large Language Models

M. Atif Quamar ¹ , M. Areeb , M. Kuznetsov , M. Ozgur Ozmen , Z. Berkay Celik

Decoding method that aligns large language models with human preferences at inference time by accepting only high-reward text segments, boosting quality without retraining.

STARS - Segment-level Token Alignment via Rejection Sampling in Large Language Models

Contact

atif7102@gmail.com