MAS-Orchestra is a reinforcement learning framework that learns to build an entire multi-agent system at once using function-calling reinforcement learning, enabling global system-level reasoning instead of sequential code execution. Together with MASBENCH, a benchmark that measures task structure across five axes, our study shows when multi-agent systems truly outperform single-agent systems and delivers consistent gains on math, multi-hop QA, and search-based tasks.
Research Papers
MAS-Zero: Designing Multi-Agent Systems with Zero Supervision
MAS-Zero is the first inference-time-only framework that meta-designs MAS through iterative generation, execution, feedback, and self-verification—no outcome supervision or validation set required. It advances the accuracy–cost Pareto frontier on math, graduate-level QA, and software engineering benchmarks.
Published: SEA@NeurIPS, 2025 🏅 Oral
A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems
Comprehensive survey that maps the rapidly evolving landscape of LLM reasoning across inference-time scaling, learning-to-reason methods, and agentic systems. Highlights unified taxonomies, benchmark trends, and emerging research challenges.
Published: TMLR, 2025 🏅 Survey Certificate