对于不同的研究人员来说，从基于理论的研究过渡到关注现实世界应用的经历可能会有很大差异。然而，, 对于两名前 MIT 研究生和一名前博士后, 而言，他们现在都在 IBM, 工作，在他们的成长岁月中与 MIT-IBM 计算研究实验室 (（以前称为 MIT-IBM Watson AI 实验室)）合作，使他们不仅能够缩小教育和就业之间的差距,，而且还能够产生有希望对业务产生影响的想法。
The experience of transitioning from research based in theory to focusing on real-world application can vary significantly for different researchers. However, for two former MIT graduate students and a former postdoc, all now at IBM, working with the MIT-IBM Computing Research Lab (formerly the MIT-IBM Watson AI Lab) during their formative years enabled them to not only close the gap between education and employment, but also to generate ideas promising to business impact.
尽管在量子机器学习,强化学习和人工智能代理,和值得信赖和公平的AI,分别, Srinivasan Arunachalam,张伟洪博士’25,和Irene Ko博士’24中追求不同的职业，但他们一直在寻找方法来解决由新颖性和严格性,定义的问题，并将其转化为具有实际约束的系统。在这里, MIT-IBM 计算研究实验室充当了建立研究关系以及将其专业知识流向行业应用的渠道。
Despite pursuing varied careers in quantum machine learning, reinforcement learning and artificial intelligence agents,and trustworthy and fair AI, respectively, Srinivasan Arunachalam, Zhang-Wei Hong PhD ’25, and Irene Ko PhD ’24 have consistently found ways to tackle problems defined by novelty and rigor, and translate them to systems with real constraints. Here, the MIT-IBM Computing Research Lab served as a conduit for research relationship building and the flow of their expertise to industry applications.
自从发现 DeepMind 可以玩 Atari 并通过特征工程从原始屏幕像素中学习以来，Hong 一直对强化学习着迷。在与 EECS 副教授 Pulkit Agrawal,（同时也是实验室, 的首席研究员）一起进行研究生工作期间，Hong 试图在此: 的基础上使用 Atari, 中的 “Montezuma的 Revenge” 改进视频游戏, 中强化学习的价值函数学习，以预测和优化代理的策略性能。通过实验室, Hong 开发了一些技术，使 AI 能够更实际地应用，并提供更好的奖励反馈,，他将其应用于机器人, 大语言模型 (LLMs), 和科学强化学习等领域。
Hong has been captivated with reinforcement learning since discovering that DeepMind could play Atari and learn from raw screen pixels via feature engineering. During his graduate work with EECS Associate Professor Pulkit Agrawal, who is also a principal investigator with the lab, Hong sought to build on this: improving value function learning for reinforcement learning in video games, using “Montezuma的 Revenge” in Atari, in order to predict and optimize the policy performance of an agent. With the lab, Hong developed techniques to ground AI for more realistic applications and provide better reward feedback, which he applied to domains such as robotics, large language models (LLMs), and reinforcement learning for science.
“I’m 对好奇心驱动的探索感到非常兴奋,” Hong 谈到 MIT-IBM 的研究生工作，帮助他进入了自己的职业。他说, , 允许代理像人类, 一样对新数据, 感到好奇，并执行各种任务— 从生成测试用例到对LLM 进行压力测试再到探索新环境。现在, 作为自己学生的导师, Hong 继续追求类似的开放式强化学习研究,，这促使他研究代理和基础模型, 的测试时培训，并为 IBM的 代理框架开发基础架构，用于企业任务，例如图表读取和数据库查询的工具调用。这包括利用进化计算来推动更好的探索优化，并利用神经科学来改进部署时间模型。
“I’m very excited about curiosity-driven exploration,” says Hong of the MIT-IBM graduate work that helped propel him into his profession. This, he says, allows agents to be inquisitive about new data, like humans, and perform a variety of tasks — from generating test cases to stress-test LLMs to exploring new environments. Now, as a mentor for students of his own, Hong continues to pursue similar lines of open-ended reinforcement learning research, leading him to investigate test-time training for agents and foundation models, and develop infrastructure for IBM的 agentic framework for enterprise tasks like chart reading and tool calling for database queries. This includes evolutionary computing to drive better optimization for exploration and leveraging neuroscience to inform deployment time model improvement.
“如果成功,，我认为这对于强化学习的所有从业者来说将是一个非常有用的系统和框架,，因为它将是第一个使模型能够改进—在部署时在线自我进化其模型权重的框架,” Hong 说。
“If successful, I think that it would be a very useful system and framework for all of the practitioners in reinforcement learning, because it will be the first framework that enables a model to improve — self-evolve their model weights online at a deployment time,” says Hong.
从个人和专业的角度来看，Irene Ko的研究也是价值驱动的,。 Ko 说，“I 从我攻读博士学位的第一天起就开始与 IBM 研究人员合作[on 值得信赖的 AI],，因为它是由 MIT-IBM,” 资助的。她说, 这个, 特别有利，因为她开发前沿安全, 稳健, 准确, 和公平人工智能的目标也与麻省理工学院和IBM, 的目标一致，缩小了开发与实际部署之间的差距。 “这确实在纯研究和’行业标准或价值之间取得了平衡。”
Irene Ko的 research has also been value-driven, from a personal and professional standpoint. “I started to work [on trustworthy AI] with IBM researchers from day 1 in my PhD, because it was funded by MIT-IBM,” says Ko. This, she says, was particularly advantageous since her goals to develop frontier-safe, robust, accurate, and fair AI also align with that of MIT and IBM, closing the gap between development and real-world deployment. “That really strikes a balance between pure research and something that的 of industry standard or value.”
此外,，她通过 EECS, 顾问 Joseph F. 和 Nancy P. Keithley 教授 Luca Daniel, 和 IBM 首席研究科学家 Pin-Yu Chen, 与 MIT-IBM 合作，帮助确定了她工作的方向和参数，以最大限度地提高影响,，首先在神经网络领域，然后在基础模型和法学硕士领域。 2024, 毕业后，Ko 加入 IBM 研究院，作为研究科学家继续从事可信人工智能方面的工作。
Further, her MIT-IBM collaboration through her advisor in EECS, Joseph F. and Nancy P. Keithley Professor Luca Daniel, and IBM Principal Research Scientist Pin-Yu Chen, helped define the direction and parameters of her work to maximize impact, first in neural networks and later with foundation models and LLMs. After graduating in 2024, Ko joined IBM Research to continue her work on trustworthy AI as a research scientist.
“我在获得博士学位后选择进入工业界,，特别是 IBM, 的原因是我在博士学位期间的合作中找到了巨大的乐趣。 Ko 说，这个过程, 那五年, 在个人成就感方面给了我非常高的回报,”。 “I 希望继续保持这一势头。”
“The reason I chose to go into industry after my PhD, and IBM specifically, is that I found great joy in the collaboration during my PhD. That process, those five years, gave me very high rewards in personal fulfillment,” says Ko. “I wanted to continue the momentum.”
她当前的项目重点是寻找当前值得信赖的方法中尚未广泛部署在人工智能推理平台中的痛点。与使用低等级适配器, 不同，后者添加了额外的步骤来监视和修改模型行为,，她在 vLLM Hook 上的工作提供了一种访问内部模型信号, 的方法，例如用于解码 LLM 的隐藏状态或激活,。该向量作用于变压器模块以分析安全分数,，例如识别即时注射和幻觉的可能性。 Here, Ko 开发了一个轻量级 vLLM 推理引擎插件框架来对模型内部进行编程，与其他方法相比，可以显着节省成本。 “I’m 对此项目感到非常自豪，因为据我们所知，这确实是, 使用推理引擎在值得信赖的 AI 中部署和开发之间的第一座桥梁。”
Her current project focuses on finding pain points in current trustworthy methods that are not widely deployed in AI inference platforms. Unlike using low-rank adapters, which add extra steps to monitor and modify model behavior, her work on vLLM Hook provides a way to access internal model signals, like hidden states or activations, for decoding LLMs. This vector acts on transformer modules to analyze safety scores, such as identifying the likelihood of prompt-injection and hallucination. Here, Ko has developed a lightweight vLLM inference engine plugin framework to program the model internals that could provide significant cost savings over other methods. “I’m very proud of this project because this is really, as far as we know, the first bridge between the deployment and development in trustworthy AI with the inference engines.”
虽然 Srinivasan Arunachalam 一直涉足量子研究,，但他不断探索理论的其他领域,，寻求在意想不到的查询和论文中找到量子见解和深层数学。 “马上,你没有’看到它。你认为,也许这只是一个普通问题,，然后一旦你开始进一步研究,，你会发现一些非常有趣的数学结果,，我认为这很酷,”他说。
While Srinivasan Arunachalam has always dabbled in quantum research, he constantly explores other areas of theory, seeking to find quantum insights and deep math in unexpected lines of inquiry and papers. “Right off the bat, you don’t see it. You think, maybe this is just a vanilla problem, and then once you start investigating it further, you find some really interesting math that comes out of it, which I think is pretty cool,” he says.
通过无缝过渡到 IBM,，Arunachalam 更加密切地参与了可能在近期量子设备, 上实现的问题，同时牢记最近邻架构, 噪声, 和更简单的可观测测量等约束。在此期间, Arunachalam 专注于量子机器学习和量子计算优于经典计算的领域, 越来越优先考虑基于理论的可证明性而不是启发式。 MIT 与 IBM 的联系有助于将理论问题转化为具体的研究方向, 塑造工作最终产生了两篇著名论文: 一篇关于哈密顿学习, 为学习量子系统的动力学提供了严格的保证, 另一篇关于量子内核, 提供了理论证据，表明在广泛相信的硬度假设下，量子特征空间可以提供优于经典内核的优势。
With a seamless transition to IBM, Arunachalam more closely involved himself with problems that are potentially implementable on a near-term quantum device, keeping in mind constraints like nearest-neighbor architecture, noise, and simpler observable measurements. During this time, Arunachalam focused on quantum machine learning and areas where quantum computing would be superior to classical computing, increasingly prioritizing provability grounded in theory to heuristics. That MIT-IBM connection helped turn theoretical questions into concrete research directions, shaping work that culminated in two prominent papers: one on Hamiltonian learning, which gave rigorous guarantees for learning the dynamics of quantum systems, and another on quantum kernels, which provided theoretical evidence that quantum feature spaces can offer advantages over classical kernels under widely believed hardness assumptions.
阿鲁纳查拉姆还继续扩大自己的知识库，将自己投入到计算机科学的不同分支中，以揭示其他人可能错过的问题的结构。 “我’一直非常喜欢的一件事是揭示不同领域之间的联系。”这使他能够探索学习量子态—从完全经典可模拟的量子对象到极其复杂的量子对象。
Arunachalam also continued to expand his knowledge base by pouring himself into different branches of computer science to uncover structure in problems others may have missed. “One thing which I’ve been a huge fan of is exposing connections between different fields.” This has allowed him to explore learning quantum states — from completely classically simulatable quantum objects to the extremely complicated quantum objects.
尽管 Hong, Arunachalam, 和 Ko 涉足不同的领域,，但他们都有一种本能:，即在原则上可能的内容和实践中有用的内容之间移动想法。每个人都以自己的方式, 应用从合作,（例如 MIT-IBM 计算研究实验室,）中获得的知识来开发“killer 应用程序” — 一个真实世界的用例，证明基础研究可以超越实验室。
Although Hong, Arunachalam, and Ko navigate different domains, they share an instinct: to move ideas across the space between what is possible in principle and what is useful in practice. In their own way, each is applying knowledge gained from collaborations, like that of MIT-IBM Computing Research Lab, to develop “killer applications” — a real-world use case that proves the underlying research can matter beyond the lab.