Research Intern - Coding Agents & Post-Training Optimization

Microsoft • Redmond, WA

5d • Onsite

About The Position

Microsoft is hiring a Research Intern to improve coding agents through post-training methods, including reinforcement learning, fine-tuning, distillation, compression, and inference optimization. Coding agents plan and edit repositories, run tests, and fix failures across long, multi-step trajectories; the team works to make these agents reliable and affordable. The team trains, evaluates, and ships agentic coding models that power developer tools at scale.

The intern will own a research question from hypothesis through experimentation, evaluate results on realistic agent benchmarks, and work toward a research paper or product integration. The role involves collaborating with researchers and engineers across model training, agent infrastructure, and hardware acceleration. During the 12-week internship, interns are paired with mentors, collaborate with other interns and researchers, present findings, and contribute to the research community. Internships are offered year-round and typically begin in the summer.

Requirements

• Currently enrolled full time in a doctoral degree program in Computer Science, Computer Engineering, Electrical Engineering, or a related STEM field.
• Be physically located at the manager’s Microsoft worksite for the duration of the internship.
• Submit at least two reference letters, a cover letter, and relevant work or research samples.

Nice To Haves

• Have at least one additional quarter or semester of school remaining after the internship.
• Research experience in at least one of LLM post-training (reinforcement learning, supervised fine-tuning, or preference optimization), model compression, or efficient inference.
• Proficiency in Python and PyTorch, with hands-on experience training or fine-tuning transformer models.
• First-author publications at top venues (NeurIPS, ICML, ICLR, ACL, EMNLP, MLSys, ICSE, or FSE) or equivalent research impact.
• Experience with large-scale reinforcement learning for LLMs using frameworks such as TRL, verl, OpenRLHF, or DeepSpeed-Chat.
• Experience building or training code-generation models or coding agents, and using benchmarks such as SWE-bench, LiveCodeBench, Terminal-Bench, or HumanEval.
• Hands-on experience with quantization, distillation, pruning, or speculative decoding; familiarity with serving stacks such as vLLM, SGLang, TensorRT-LLM, or ONNX Runtime.
• Familiarity with distributed training (FSDP, DeepSpeed, or Megatron) and GPU profiling or kernel work (CUDA or Triton).
• Clear communication, self-direction, curiosity, and a bias toward hands-on experimentation.
• Experience collaborating across teams, disciplines, and cultures.

Responsibilities

• Own a research question end to end: form a hypothesis, build experiments, and measure outcomes on realistic agent benchmarks.
• Advance research toward a paper or product integration.
• Collaborate with researchers and engineers across model training, agent infrastructure, and hardware acceleration.
• Work with mentors and other Research Interns and researchers.
• Present findings and contribute to the research community.

What This Job Offers

• Job Type: Full-time
• Career Level: Intern
• Education Level: Ph.D. or professional degree