Position Title: Research Intern for Machine Learning Group – Discriminative Foundation Models and Efficient Decision-Making
Job Type: Full-time Intern
Number of Openings: 2-4
Location: Beijing
Group Introduction:
We study intelligent agents that can perceive, reason, plan, and act in complex interactive environments. Our research brings together decision models, large language models, multimodal models, and reinforcement learning, enabling agents to learn continuously from interaction and improve their decision-making capabilities. Over the past few years, we have published numerous papers in leading international conferences and journals. We also have extensive expertise in reinforcement learning: our Mahjong AI system, Suphx, became the first AI system to achieve the rank of ten-dan on the internationally renowned competitive Mahjong platform Tenhou, surpassing the average performance of top human players in the platform’s public rooms.
This position focuses on Decision Foundation Models for discriminative tasks—models that turn natural-language task descriptions and input context into fast, structured decisions that software can use directly. Rather than generating free-form text, these models select among specified options, make binary judgments, or assign scores, with an emphasis on well-calibrated probabilities and meaningful uncertainty estimates. We explore large-scale multitask learning, reinforcement learning for calibrated decisions, data synthesis and selection, lightweight architectures, and efficient inference. Our goal is to build reusable decision-making capabilities that generalize across tasks while balancing accuracy, reliability, latency, and cost for agents and automated workflows.
Job Responsibilities:
Contribute to the research and development of Decision Foundation Models, building on the team’s existing training framework. Investigate data, learning objectives, model architectures, and inference methods to improve performance across diverse discriminative tasks. Specific responsibilities include:
1. Training and Post-Training Methods: Investigate multitask learning and reinforcement learning objectives for accurate, well-calibrated decisions, including research directions motivated by Reinforcement Learning for Calibrated Decisions (RLCD). Study how to jointly improve predictive accuracy, probability calibration, and uncertainty estimation.
2. Data Development and Optimization: Collect, curate, and synthesize training data for classification, option selection, binary judgment, and rubric-based scoring. Explore data quality assessment, selection, and task-mixture strategies to improve data efficiency and generalization to new task descriptions and label sets.
3. Lightweight Architectures and Efficient Inference: Explore architectures and prediction mechanisms tailored to structured decision outputs, alongside knowledge distillation and inference optimization. Reduce latency, GPU memory usage, and serving costs while preserving predictive quality.
4. Evaluation and Research Iteration: Develop reproducible evaluations across domains, unseen tasks, and distribution shifts. Assess accuracy, calibration, robustness, and inference efficiency, including how uncertainty estimates support abstention or escalation to human review. Identify research questions through systematic experiments and failure analysis.
Qualifications:
- PhD in Computer Science, Software Engineering, Electrical Engineering, or other related fields. Undergraduate and Master’s students aspiring to scientific research are also welcome to apply.
- Strong Python programming skills and proficiency in using AI coding tools to accelerate research and development, with the ability to understand, validate, and improve generated code. Good communication and teamwork skills.
- A solid foundation in machine learning and deep learning, familiarity with frameworks such as PyTorch, and the ability to read current research papers and conduct experiments. A strong interest in discriminative foundation models, structured prediction, calibrated decision-making, and efficient inference.
- Bonus points: Familiarity or hands-on experience with models designed for structured decision-making, such as Jev; research experience in classification and scoring models, reinforcement learning, probability calibration and uncertainty estimation, synthetic data generation and data selection, model distillation, or lightweight architectures; achievements in mathematics, physics, or informatics competitions; or publications in top international conferences and journals..
Internship Duration Requirements:
Must obtain permission from your academic advisor and commit to at least three months of internship.
Please be sure to download and complete the application form (Application form link: https://aka.ms/InternApplication (opens in new tab) ) and send it along with a complete English and Chinese resume (in PDF/Word format) to: MSRAih@microsoft.com & lizo@microsoft.com. Please include ” Research Intern for Machine Learning Group – Discriminative Foundation Models and Efficient Decision-Making ” in the email subject line.
To learn more about the “Stars of Tomorrow Internship Program,” please visit: https://www.msra.cn/zh-cn/connections/academic-programs/tomorrowstars (opens in new tab)
岗位名称：机器学习组研究实习生 – 判别式基础模型与高效决策方向
工作性质：全职实习生
招聘人数：2-4人
工作地点：北京
组别简介：
我们致力于研究能够在复杂交互环境中感知、推理、规划和行动的智能体，探索决策模型与大语言模型、多模态模型、强化学习的融合，使智能体能够通过环境交互持续学习并提升决策能力。过去几年，我们在国际顶级会议和期刊上发表了大量论文。我们在强化学习领域也有深厚积累：我们的麻将 AI 系统 Suphx 成为首个在国际知名专业麻将平台“天凤”上达到十段的 AI 系统，在该平台公开房间中超越了顶尖人类玩家的平均水平。
本岗位聚焦于面向判别式任务的决策基础模型（Decision Foundation Models），使模型能够根据任务描述直接输出分类、判断、选择或评分结果，并提供可靠的概率与置信度。我们探索大规模多任务训练、面向校准决策的强化学习、数据合成与筛选，以及轻量化架构与高效推理，致力于实现跨任务泛化、低延迟、低成本的可靠决策，为智能体与自动化系统提供可复用的基础决策能力。
工作职责：
参与决策基础模型的研发，在团队已有训练框架上持续迭代，围绕数据、训练方法、模型结构与推理效率开展研究，提升模型在广泛判别式任务上的泛化能力、预测可靠性与计算效率。具体包括：
- 训练与后训练算法：探索多任务学习，以及面向校准决策的强化学习方法（如 RLCD 相关方向），联合提升决策准确率、概率校准与不确定性估计能力。
- 数据构建与优化：收集、清洗并合成覆盖分类、判断、选择和评分等任务的训练数据，探索数据筛选、质量评估与任务配比策略，提升数据效率与任务覆盖度。
- 轻量化与高效推理：研究适合判别式任务的模型结构、知识蒸馏与推理优化方法，降低延迟、显存占用和部署成本，优化准确率与效率之间的权衡。
- 评测与研究迭代：建立覆盖跨领域、未见任务及分布变化的评测体系，从预测质量、校准程度和推理成本等维度分析模型表现，提炼研究问题并开展可复现的实验验证。
任职要求：
1. 计算机科学、软件工程、电子工程或其它相关专业博士，有志于科研的本科生或硕士也欢迎申请；
2. 具备扎实的 Python 编程能力，能够熟练使用 AI coding 工具高效开展研发，并对生成代码进行理解、验证与改进；具备良好的沟通能力和团队协作精神。
3. 具备机器学习与深度学习基础，熟悉 PyTorch 等框架，能够阅读相关领域的前沿论文并开展实验；对判别式基础模型、结构化决策、概率校准与高效推理有浓厚兴趣。
4. **加分项：**了解或使用过 Jev 等面向结构化决策的模型；具有分类与评分模型、强化学习、概率校准与不确定性估计、合成数据与数据筛选、模型蒸馏或轻量化架构等相关研究经验；有数学、物理或信息学竞赛成绩，或在国际顶级会议和期刊上发表过论文。
工作时间要求：
能获得导师许可并保证至少三个月的实习。
请务必下载并填写申请表（申请表链接：https://aka.ms/InternApplication (opens in new tab) ）并将其与完整的中英文简历（PDF/Word形式）一同发送至： MSRAih@microsoft.com & lizo@microsoft.com，邮件标题中注明：机器学习组—判别式基础模型与高效决策方向实习生。
了解更多“明日之星实习生计划”，请访问：微软亚洲研究院“明日之星”实习生项目 – Microsoft Research