Kaveh Alim

🧠 Inference-time scaling lets LLMs spend more compute to solve harder problems, but not every question needs that! After all, we don’t use a whiteboard to solve 1 + 1. So why should an LLM? Introducing Instance-Adaptive Inference-Time Scaling, a smarter way to allocate compute during inference. 💡 By calibrating PRMs, we can estimate how likely each reasoning step is to lead to a correct answer. This lets us: - Use less compute on easy questions and on likely incorrect reasoning paths - Spend more compute on harder questions and on promising reasoning steps 🎯 Same accuracy, up to 50% less compute. Paper → arxiv.org/abs/2506.09338 Code → github.com/azizanlab/inst… Calibration dataset → huggingface.co/datasets/young… Calibrated PRMs → huggingface.co/collections/yo… Joint work with @KGreenewald, @KavehAlim, @HW_HaoWang, and @NavidAzizan.