Senior Researcher - Self-Evolving Agentic Systems

Microsoft Research India is a leading research lab advancing innovation across areas including Artificial Intelligence and Machine Learning, Theory, Technology and Empowerment, and Systems. Established two decades ago, the lab has been at the forefront of computer science research in India and is home to researchers and impactful projects.

AI agents are moving beyond fixed prompts and static policies. They can now plan, use tools, write code, and work across long, multi-step tasks. However, most agents still cannot continuously improve through their own experience, learn from failures, or adapt while operating.

Our team is building self-evolving agentic systems: agents that improve their own policies, skills, memory, and scaffolding through interaction, feedback, and reinforcement learning.

As a Senior Researcher – Self-Evolving Agentic Systems, you will advance the state of the art in how LLM-based agents learn and adapt. You will work across the full stack, from designing reinforcement learning algorithms and reward signals to building environments and evaluations and translating research ideas into systems that operate at scale.

You will collaborate with researchers and engineers across Microsoft Research India, Microsoft Research worldwide, and Microsoft product groups, enabling your work to contribute to both the research community and products used by millions of people.

We are looking for candidates who:

• Are passionate about advancing the state of the art in reinforcement learning, LLM post-training, and autonomous agents.
• Have a strong research track record demonstrated through publications at leading venues such as NeurIPS, ICML, ICLR, ACL, EMNLP, COLM, or similar venues, or through impactful open-source contributions.
• Combine scientific rigor with strong engineering skills and enjoy turning research ideas into working systems.
• Thrive in a highly collaborative research environment.
• Take end-to-end ownership of complex, open-ended research problems.

Responsibilities

• Lead research on self-evolving agents, including self-improvement loops, experience-driven learning, self-generated curricula, and automatic refinement of agent policies, tools, and memory.
• Design and apply reinforcement learning methods for LLM-based agents, including reward modeling, learning from verifiable and AI feedback, and credit assignment over long, multi-step tasks.
• Build scalable training and evaluation environments for agentic tasks such as coding, tool use, web and computer use, and knowledge work.
• Develop rigorous benchmarks and evaluation methods that measure agent capability, robustness, safety, and improvement over time.
• Form hypotheses, design experiments, and iterate quickly on model and agent performance.
• Publish at top-tier venues, release open-source artifacts, and help shape the research agenda in agentic AI.
• Collaborate with research and product teams across Microsoft to move research advances into real-world systems.
• Mentor interns and early-career researchers and contribute to Microsoft Research India’s academic engagement.
• Embody our culture and values.

Required Qualifications

• PhD in Computer Science, Machine Learning, Mathematics, or a related technical field, or equivalent experience.
• 1+ years of software engineering or research experience.
• Experience with reinforcement learning and LLMs.
• Strong research track record in machine learning, computer systems, or a closely related area.
• Demonstrated ability to conduct independent research and communicate results through publications, prototypes, or other impactful outcomes.
• Ability to work effectively across disciplinary boundaries and collaborate with diverse research and engineering teams.

Preferred Qualifications

• Experience building and scaling large AI systems.
• Experience applying reinforcement learning to AI models and agents.
• Experience with reward modeling, reinforcement learning, or other post-training techniques.
• Hands-on experience building datasets for LLM, VLM, or multimodal model pre-training or post-training.
• Experience with synthetic data, model graders, reward signals, hard-example mining, active learning, or data-mixture optimization.
• Experience building feedback loops that connect evaluation, failure mining, data generation, and training.
• Experience with multimodal or agentic datasets, including text, web, GUI, tool-use, or interaction trajectories.

How to Apply

Please apply using the application link above. In addition, please send an email to the Hiring Committee at [email protected].