🎉 Medical Image Spatial Grounding with Semantic Sampling Accepted to MICCAI 2026
Hi, I'm Mohsen!
and I love math ❤️. I work on these things.
I'm always excited to start new collaborations, especially when it's something new to me. Feel free to reach out by email; you'll find it in the footer.
News
View all🎉 Ranking Reasoning LLMs under Test-Time Scaling Accepted to ACL 2026 Main
🎉 Quantize What Counts: More for Keys, Less for Values Accepted to ACL 2026 Findings
🎲 Don’t Pass@𝑘: A Bayesian Framework for Large Language Model Evaluation Accepted to ICLR 2026
📦 Julia & Python pkgs for the Bayesian framework are out!
📦 vLLM × DFloat11: run your model with 30% less memory!
✨ 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float Accepted to NeurIPS 2025
Recent Research
View all
Quantize What Counts: More for Keys, Less for Values
• Mohsen Hariri, Alan Luo, Weicong Chen, Tianyi Zhang, Qifan Wang, Xiaotian Han, Vipin Chaudhary
A geometry-driven mixed-precision KV-cache quantization poster showing that keys carry more information than values, so key-favored bit allocation preserves accuracy while reducing memory.
Ranking Reasoning LLMs under Test-Time Scaling
• Mohsen Hariri, Michael Hinczewski, Jing Ma, Vipin Chaudhary
Ranking reasoning LLMs under repeated sampling, comparing 72 ranking methods across four Olympiad-style math benchmarks and packaging them in Scorio.
Don't Pass@k: A Bayesian Framework for Large Language Model Evaluation
• Mohsen Hariri, Amirhossein Samandar, Michael Hinczewski, Vipin Chaudhary
Proposed a Bayesian framework that estimates models' success probabilities with quantified uncertainty, yielding more reliable rankings and enabling categorical evaluation of LLMs.
70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float
• Tianyi Zhang, Mohsen Hariri, Shaochen Zhong, Vipin Chaudhary, Yang Sui, Xia Hu, Anshumali Shrivastava
NeurIPS 2025 poster on DFloat11: a lossless compression framework that shrinks LLMs and diffusion transformers to ~70% of their size with bit-for-bit identical outputs, plus a GPU kernel that decompresses on the fly.
Recent Posts
View allEntropy of bfloat16 During Training: How Optimizers Shape Weight Distributions
• Training, Information Theory, Optimizers
Entropy of bfloat16: 8 Bits Are Doing 2.6 Bits of Work
• LLMs, Information Theory, Efficiency
Simulating LLM Evaluation Datasets Using Psychometric Models
• Simulation, LLMs, Reasoning
Recent Slides
View allServing Reasoning LLMs Efficiently and Reliably [No Anime]
• LLMs, Efficiency, Inference
Serving reasoning LLMs efficiently and reliably: lossless DFloat11 compression, KV-cache quantization, and Bayes@N evaluation and ranking under test-time scaling.
Serving Reasoning LLMs Efficiently and Reliably
• LLMs, Efficiency, Inference
Serving reasoning LLMs efficiently and reliably: lossless DFloat11 compression, KV-cache quantization, and Bayes@N evaluation and ranking under test-time scaling.
Python Environments
• Tools, Tutorial
Python environments, how to create and reproduce them, and when to use pip, conda, micromamba, uv, pipx, lockfiles, and containers.