Microsoft Research Forum

_{TALK | IMPACT AT SCALE}

Pushing boundaries of complex reasoning in small language models

This talk explores Phi-4-Reasoning and Phi-4-Reasoning-Plus—two 14B models designed to push the boundaries of complex reasoning in small language models. By introducing a dedicated “thinking block” into the model response and applying supervised fine-tuning and reinforcement learning on carefully curated, reasoning-intensive STEM datasets, we achieved significant advances in the models’ problem-solving capabilities.

_{TALK | MISSION ON AI}

Dion: The distributed orthonormal update revolution is here

We introduce Dion, an orthonormal-update optimizer in the style of Muon that adds a tunable rank axis—orthonormalizing only the top-r subspace via amortized power iteration—to retain Muon’s fast convergence while significantly reducing compute and communication, and scaling efficiently with FSDP/TP for very large models.

_{TALK | NEW ADVANCEMENT OUT OF THE LAB}

Disrupting AI Datacenters with MicroLEDs

The performance of modern AI systems is increasingly constrained by limitations in memory and network resources, driven by fundamental scaling challenges in current communication technologies. Copper links are power-efficient and reliable but have very limited reach (< 2 m). Optical links offer longer reach but at the expense of high-power consumption and lower reliability. A collaborative effort across MSR, Azure Networking, Azure Hardware, and M365 has been investigating the use of microLEDs to develop a novel technology that can break this trade-off by providing high-distance connectivity with low power consumption, low cost, and high reliability, opening up exciting opportunities for radical new AI cluster designs.

_{TALK | AI FOR ALL}

Scalable emulation of protein equilibrium ensembles with BioEmu

The next frontier in molecular biology is predicting how biomolecules change shape to drive biological function, a key step for drug discovery. MSR AI for Science’s Biomolecular Emulator (BioEmu) recently reached a major milestone, emulating protein structure changes and stabilities with near-experimental accuracy while running up to 100,000x faster than traditional simulations.

_{TALK | RESEARCH EXPLORATION}

Dehumanizing Machines: Making Sense of AI Systems that Seem Human

GenAI systems are the focus of growing scrutiny in part because they are increasingly anthropomorphic—i.e., they are designed or perceived to be human-like. Guiding practitioners in appropriately and deliberately designing anthropomorphic AI systems requires equipping them with clarity regarding what about these systems’ design and behavior makes them anthropomorphic, and which design decisions and system behaviors are desirable and effective.

Host remarks