Microsoft Research Forum

_{TALK | OPEN RESEARCH EXPLORATION}

Magentic Marketplace: Testing societies of agents at scale

As AI agents move from isolated tools to active participants in multi-agent ecosystems, their success depends on more than task competence—it requires strategic behavior under misaligned incentives and imperfect information. Using Magentic Marketplace, an open-source simulation of two-sided agent markets, we show that while frontier models can achieve strong welfare outcomes in ideal settings, performance degrades at scale and reveals emergent failure modes such as manipulation and speed bias, motivating a shift toward training agents for social reasoning.

_{TALK | IMPACT AT SCALE}

Agent Lightning: One learning system that makes all agents evolve

Agent Lightning is an agent optimization framework that enables agents to learn from their experiences through reinforcement learning and other methods. By treating agents as first-class citizens, optimization becomes automatic for any agent with minimal code changes.

_{TALK | MISSION ON AI}

Teaching small language models to think like optimization experts with OptiMind

OptiMind is a specialized language model that translates natural-language problem descriptions directly into solver-ready mathematical optimization formulations. This removes one of the most expertise-intensive bottlenecks in optimization workflows and makes advanced optimization more accessible.

_{TALK | AI FOR ALL}

Lessons from deploying HealthBots with experts-in-the-loop

The Tyger framework enables faster, more accessible medical imaging by streaming raw data to the cloud for accelerated reconstruction—reducing patient wait times and discomfort—while empowering researchers to rapidly test and deploy new algorithms.

_{TALK | NEW ADVANCEMENT OUT OF THE LAB}

ARO: A new lens on matrix optimization for LLMs

We present Adaptively Rotated Optimization (ARO), a matrix optimizer that speeds up LLM training by applying updates in a rotated, geometry-aware coordinate system. Guided by new insights on global structures on LLM loss landscapes, ARO treats rotation as a unifying principle for sample efficiency, and proposed a new update policy that is applicable to all model weight matrices. In large scale controlled experiments, ARO consistently outperforms AdamW and orthogonalization-based method, maintaining its gains as models and training budgets scale.

_{TALK | NEW ADVANCEMENT OUT OF THE LAB}

Dion2: A new simple method to shrink matrix in Muon

Dion2 reduces the cost of Muon’s orthonormalization step by orthonormalizing only a small, selected submatrix at each iteration. This lightweight approach preserves Muon’s strong performance while significantly improving scalability of optimizer at scale.

Host