Transcript

Microsoft Research Forum | Season 2, Episode 2

From Microsoft Research · Episode page · Watch on YouTube

· 34 min · 4,785 words · 42 timestamped sections

This is the full, timestamped transcript of "Microsoft Research Forum | Season 2, Episode 2", an episode of Microsoft Research published December 9, 2025 (34 min, about 4,785 words). It mentions 13 people, works and ideas, including Microsoft Research, Josh Benalo, Ida Momenad, Tyler Payne, Tiger.

What this episode is about

The content delves into the innovative research at Microsoft focused on the intersection of neuroscience and artificial intelligence, specifically exploring how brain-inspired architectures can enhance the multi-step reasoning capabilities of large language models. Researcher Ida Momenad discusses how collaborative mechanisms akin to neuron interactions can improve planning and reduce hallucinations in AI systems, showcasing the potential for transformative applications in real-world scenarios.

Key takeaways

  • Brain-inspired AI models could elevate multi-step reasoning, tackling complex tasks that stump traditional LLMs.
  • Surprisingly, smaller LLMs like Llama 7TB outperform larger models like GPT-4 in specific reasoning tasks.
  • MAP architecture reduces hallucinations to 0%, setting a new standard for reliability in AI systems.
  • Curiosity-driven research signals a shift: AI's future lies in collaboration, not just computational power.
  • From academic theory to practical application, brain-inspired designs could change how AI systems serve real-world needs.

Who speaks in this episode?

What is discussed, and when

  1. – 5:00: Microsoft Research, Ida Momenad, Karen Easterbrook
  2. – 10:00: Microsoft Research, Josh Benalo, Karen Easterbrook, Llama 7TB
  3. – 15:00: Microsoft Research, Tyler Payne, Election Guard, Magent 1
  4. – 20:00: Microsoft Research, Ilana Rosenberg, Tyler Payne, Magent 1
  5. – 25:00: Microsoft Research
  6. – 30:00: Microsoft Research, Chunlong Wong
  7. – 34:02: Microsoft Research

Notable quotes from "Microsoft Research Forum | Season 2, Episode 2"

  • Great to celebrate the incredible breadth of research happening across our labs with you. At Microsoft Research, we believe progress starts with curiosity with open exploration…
    about Microsoft Research · at
  • Hello, I'm Idam Manhad, principal researcher at Microsoft Research New York City. Today I will talk to you a little bit about how brain inspired agentic architectures built with LLMs.
    about Ida Momenad · at
  • Josh Benalo, senior principal cryptographer here at Microsoft Research Redmond, has spent decades pioneering cryptographic tools that bring transparency and trust to elections.
    about Josh Benalo · at
  • Responsibly by understanding not just what these systems can do, but how and why they sometimes fail. Tyler Payne, a senior research software engineer with Microsoft Research…
    about Tyler Payne · at
  • For example, Magent 1 is a popular generalist agent developed by my collaborators here at MSR. It is designed as a multi- aent system which is a useful programming abstraction that.
    about Magent 1 · at
  • To share how AI and cloud computing can make health care faster, more efficient, and more equitable, we'll hear from Ilana Rosenberg, principal product manager in Microsoft…
    about Ilana Rosenberg · at
  • Tiger, developed in Microsoft Research, alleviates these challenges. Tiger is an open-source and cloud-based remote signal processing framework that enables online MRI reconstructions.
    about Tiger · at
  • Organizations harness technology in their everyday work. A fellow Redmondbased colleague, Cheng Lang Wong, principal researcher, works with his team to reimagine how analysts…
    about Chunlong Wong · at

Who and what is mentioned in this episode?

People

  • Josh Benalo — Person · Main speaker · 2 mentions · first at
    Explains Microsoft Research work on publicly verifiable elections and cryptographic approaches that improve transparency and voter trust, including research removing cryptographic keys via commitments.
  • Ida Momenad — Person · Main speaker · 1 mention · first at
    Presents research on brain-inspired agentic architectures (MAP) using multiple LLM modules to improve multi-step reasoning, planning, and reduce hallucinations.
  • Tyler Payne — Person · Main speaker · 3 mentions · first at
    Discusses research on LLM agents using many tools via MCP and the phenomenon of tool space interference; introduces the MCP Interviewer tool and findings from analyzing public MCP servers.
  • Karen Easterbrook — Person · Host · 2 mentions · first at
    Opens and frames the Microsoft Research Forum episode, introduces speakers and connects their work to MSR’s mission and responsible innovation.
    Impact from scientific discovery to innovation that shapes how people live and work. I'm Karen Easterbrook, senior director of the cryptography team here at Microsoft Research…
  • Ilana Rosenberg — Person · Main speaker · 1 mention · first at
    Presents the Tiger open-source cloud framework for streaming MRI raw data for real-time reconstruction, enabling faster prototyping and potentially improving access and throughput in medical imaging.
  • Chunlong Wong — Person · Main speaker · 1 mention · first at
    Introduces Data Formulator, an interactive AI-powered visualization tool combining natural language and UI-driven iteration to help users explore data while staying in control.

Books & papers

  • Artificial Life — Book · Referenced work · 1 mention · first at
    A publication exploring multi-agent architectures for collective innovation, referenced in the research.
    Improve accuracy, reliability, and potentially safety in AI systems. In parallel work published in artificial life, we explored topologies of multi-agent architectures for…

Companies & orgs

  • Microsoft Research — Company · Business focus · 14 mentions · first at
    The organization hosting the forum and presenting various research projects.

Products & tech

  • Tiger — Product · Case study · 3 mentions
    An open-source framework for remote signal processing in MRI imaging, improving healthcare efficiency.
  • Magent 1 — Product · Case study · 2 mentions · first at
    A multi-agent system designed for generalist tasks, demonstrating tool space interference issues.
  • Election Guard — Product · Case study · 1 mention · first at
    A tool for verifiable elections, ensuring transparency and trust in the voting process.
    Microsoft has built free open-source tools which allow voters to do this kind of confirmation and these tools can be integrated into new and existing systems so that elections can be made.
  • Azure — Product · Referenced work · 2 mentions
    Microsoft's cloud computing platform used in the Tiger framework for MRI data processing.
    Here's how. As the MRI scan is being completed, images are streamed to the cloud real time. Data from the scanner is divided into blocks and uploaded as buffers to Azure blob…
  • Llama 7TB — Product · Referenced work · 1 mention · first at
    A smaller, cost-efficient LLM used in research to demonstrate superior performance in certain tasks.
    Hallucination. MAP also shows superior transfer learning between tasks and out of distribution. A notable outcome was that when we built map with a smaller and more…

Full transcript with timestamps

Hi everyone and welcome to Microsoft Research Forum where we explore purposeful research and its real world impact from scientific discovery to innovation that shapes how people live and work. I'm Karen Easterbrook, senior director of the cryptography team here at Microsoft Research in Redmond. It's great to celebrate the incredible breadth of research happening across our labs with you. At Microsoft Research, we believe progress starts with curiosity with open exploration that grows into breakthroughs and products and ultimately drives positive change in the world. Today's talks reflect that full arc from foundational science to AI for social good. From brains inspired architectures to new tools that make data more human and intuitive. Joining us from our New York City lab, Ida Momenad is exploring how neuroscience can inform artificial intelligence. Her research on brain inspired agenic

architectures shows how multiple LLMs can collaborate like neurons in the brain, improving reasoning, planning, and even reducing hallucinations. This is the kind of curiositydriven science that lays the foundation for future breakthroughs. Over to you, Ida.

Hello, I'm Idam Manhad, principal researcher at Microsoft Research New York City. Today I will talk to you a little bit about how brain inspired agentic architectures built with LLMs can improve multi-step reasoning. Large language models demonstrate impressive performance on a variety of tasks like writing emails or answering questions. But LLMs and agentic AI systems often struggle with tasks that require multi-step reasoning or goal- directed planning which are necessary skills in many real world applications. Think about planning a trip, coordinating a project or enforcing safety rules in a sequence of steps. Those are problems that require multi-step reasoning and planning. You need to keep track of where you are, what's allowed, and what you're trying to achieve. And this way, you take actions and track your goals. These skills are crucial for users, organizations, and Microsoft customers

who wish to deploy generative AI in their work. In our earlier work, we have rigorously shown failure modes of large language models in solving simple multi-step reasoning tasks like navigating a building described in text or passing a message in a network of colleagues. While the models often sound confident, we found that they commonly propose illegal moves and hallucinate paths that get stuck in loops or wander off detours. For users and Microsoft customers, improving multi-step reasoning and planning is not just academic, it's a necessity.

This is why we propose a modular agentic planner or map inspired by the brain to solve these problems. To address these challenges, we took inspiration from the human brain in which planning is accomplished via component processes that are predominantly associated with specific brain regions. These processes include task decomposition, task coordination, conflict monitoring, state prediction, and evaluation. A key insight from our research was that when testing LLMs on individual functions or processes, we find that they are often capable of carrying out these functions in isolation. However, they struggle to autonomously coordinate them in the service of a goal. Inspired by these findings, we designed a brain inspired architecture with specific brain inspired modules.

Here is how map or our modular agentic planner works. It is designed with functional rules, communication protocols and iterative algorithmic steps for collaborative problem solving inspired by the brain. The brain inspired interactions include the actor proposing actions in response to a prompt to solve a task. The monitor checking whether those actions are valid or hallucinations and gating them. The predictor and evaluator then taking the actions that pass that first stage and then doing its research to predict its future and evaluate whether it's the right outcome. And the orchestrator determining when the goals are achieved.

This is implemented with specialized prompting of LLM instances where roles and interaction protocols are all inspired by the brain. What we found is that MAP yields significant improvement over both standard LLMs at zeroot and multi-shot examples and competitive agentic baselines. For instance, on Tower of Hanoi, we show that map improves performance from 11% for GPT4 zeroot to 74% for map that's been designed with all GPT4 agents. Graph traversal was improved from 50% for our best baseline to 95% on a fourstep path. in terms of errors and hallucinations. MAP had 0% invalid moves even in out of distribution tasks whereas the other methods had up to 31% hallucination. MAP also shows superior transfer learning between tasks and out of distribution. A notable outcome was that when we built map with a smaller and more costefficient LLM, meaning Llama 7TB as opposed to GPT4, we saw superior performance and transfer

across tasks. So as a summary, MAP can improve accuracy, reliability, and potentially safety in AI systems. In parallel work published in artificial life, we explored topologies of multi-agent architectures for collective innovation. Together, this line of research findings offers a blueprint to advance future end-to-end architectures beyond transformers that readily incorporate MAP's multi-level and multi-roll computations to improve performance, reliability, and safety. This has implications for customers real world needs and their trust way beyond our toy research tasks. Together, we can use human- centered and brain inspired methods in the service of mitigating inaccuracies and risks and improving performance of our AI systems with rigorous methods. Thank you.

Thank you, Ida. A powerful example of how foundational ideas can reshape what's possible in AI. I'm happy to introduce a fellow team member, Josh. Josh Benalo, senior principal cryptographer here at Microsoft Research Redmond, has spent decades pioneering cryptographic tools that bring transparency and trust to elections. Today, he will share his work on publicly verifiable elections, making it easier for voters to confirm that their ballots are counted correctly. His latest research eliminates the need for cryptographic keys, making secure, transparent voting even more practical for election administrators. It's amazing to see impactful research grow into technology that empowers people globally while strengthening systems we all rely on. Let's hear from Josh.

Hello, I'm Josh Benalo, senior cryptographer at Microsoft Research in Redmond. We have a crisis of confidence in elections today. Both in the US and elsewhere, there are millions of voters who simply don't believe the results of recent elections. Now, regardless of where you stand on the issues, we should all be able to acknowledge that elections in the US or anywhere in the world look pretty much the same regardless of how much integrity they have. From the voter perspective, voters cast their votes, go home, wait for the results to be reported by the government, um, and are asked to simply trust the process.

We can do better than this. We have technologies available that allow voters to confirm for themselves that their votes have been accurately recorded and counted in an election without having to trust any election software, any election hardware, even any election personnel. Microsoft has built free open-source tools which allow voters to do this kind of confirmation and these tools can be integrated into new and existing systems so that elections can be made verifiable. In partnership with the nonprofit election technology initiative, Microsoft is deploying tools which have been used in at least five states thus far in public elections to achieve exactly this kind of verifiability in elections. Um I'm going to describe one such application of these tools, but they're flexible and can be used in a variety of ways. What I'm showing here is a uh precinct scanner uh produced by one of the major

uh voting equipment companies in the US, Hart Intercivic. Uh voters can use this device by marking a paper ballot either by hand or with an assisted device. They then take their paper ballot, put it into the scanner, and the scanner will display for them the contents of the ballot. Voters are able to review these contents. Um, if they like them, um, they can cast the ballot. If they don't, they can ask that the ballot be canceled and continue. At the same time, the election guard code is called by the device which encrypts the ballot and produces a confirmation code which is given to a voter. the voter can use to see that their vote has been correctly recorded and eventually that it's been correctly counted. The way the counting is done is using homorphic encryption. Um homorphic encryption allows the encrypted ballots to be combined into an encrypted tally without ever being individually

decrypted. and the encrypted tally and only the tally is then decrypted to to show what the results are without compromising the privacy of voters. Uh voters are able to independently assess and confirm that the steps done have all been correct, that the homamomorphic tallying is correct, that their votes have been properly recorded and all on their own or by proxying to other people that they choose to trust. Uh the one of the major differences with this technology is voters are not forced to trust certain people or or processes.

They can choose where to place their trust or do everything on their own. New research that that's taking place um here at Microsoft um is actually improving this process further by um removing the encryptions and replacing encryptions by what are called cryptographic commitments. These commitments um allow us to dispense with all of the cryptographic keys and all the associated management. They make the process much much simpler for election officials without encumbering voters in any way. We very much look forward to building this and deploying this and and trying to get more out into the real world. And if you'd like to learn more, please visit this URL or contact me.

Thank you very much. It's inspiring to see how cryptographic innovation can strengthen institutions that matter to all of us. Thanks, Josh. As we progress in all directions of research at MSR, we stay true to a core part of our mission, advancing AI responsibly by understanding not just what these systems can do, but how and why they sometimes fail. Tyler Payne, a senior research software engineer with Microsoft Research AI Frontiers in New York City, is investigating how AI agents perform when they're given access to multiple tools, from calculators to code interpreters. Surprisingly, his findings show that adding more tools can sometimes hurt performance, introducing tool space interference. Over to you, Tyler.

Hi, my name is Tyler and I'm a research engineer at Microsoft Research AI Frontiers. Today I'm going to be talking about an emerging problem for LLM agents that we call toolbas interference. This was an exploration done over the summer of 2025 in collaboration with my colleagues here at AI Frontiers. AI agents powered by LLMs have become a popular topic in both research and industry. In general, an agent is a system that can sense and affect its environment in pursuit of a goal. LLM agents are usually software systems that equip LLM with tools they can use to understand and manipulate their environment to complete tasks on behalf of their users. Often, these agents act in computer environments where they can browse the web, write code, and manipulate the file system.

For example, Magent 1 is a popular generalist agent developed by my collaborators here at MSR. It is designed as a multi- aent system which is a useful programming abstraction that delegates certain capabilities to sub agents. Specifically in Magent, these sub aents are the coder, terminal, web surfer, and file surfer. All of which are coordinated by a top level orchestrator agent. Now let's imagine you ask Magentic 1 to solve a git related task. First, the orchestrator must decide whether to delegate that task to the terminal agent or the web server agent. But when building a system like Magentic 1, we can evaluate its behavior on such tasks and fix issues by adjusting any part of the system. So for example, we can provide in context examples to the orchestrator if it decides to delegate to the wrong sub agent. Likewise, we can adjust the tools and prompts of these sub aents directly.

In this way, Magent 1 is a vertically integrated system. But in the past year, the model context protocol or MCP has exploded in popularity. MCP enables developers to bundle their tools into a server that can be easily shared and consumed by LLM agents. Most popular LLM agents like Cloud Code, Cursor, and GitHub Copilot already support MCP servers. This lets any user extend their agent at runtime, breaking the assumptions of vertical integration. Now while this horizontal extensibility is exciting in principle in practice we observed that it can actually reduce LLM agents performance we call this phenomenon tool space interference.

In order to study toolspace interference we developed MCP interviewer a CLI tool that automatically analyzes MCP servers collecting descriptive statistics like the number of tools they provide the depth and length of those tool schemas and many more features. It can also use an LLM to generate a functional test plan that invokes each of the servers tools to test that they behave as expected. You can also use MCP interviewer to do qualitative LLM as a judge evaluation of the server. We're excited that MSR enables us to share these tools with the world and we've open sourced the MCP interviewer on GitHub. Back to the research, we collected nearly 1500 real MCP servers from public registries including smithy.ai and Docker MCPhub. We then ran the MCP interviewer on each of these servers and analyzed the results which we lay out in detail in our blog post on the MSR blog. To recap our main findings, we identified a few common

issues that can cause tool space interference. First of all is tool name collisions. Two tools cannot have the same name and LLM provider APIs will reject requests if there are name collisions between tools. MCB provides no formal guidance on namespacing and so clients have had to develop each develop their own strategies like prefixing the server name before the tool name. Beyond exact collisions though, tool names can also have significant semantic overlap like search, web search, Bing search, and Google search. This can also confuse agents. Next, we identified servers that expose too many tools. OpenAI's API accepts a maximum of 128 tools, and their documentation recommends keeping that number well below 20, but we observe many servers above this 20 tool threshold. Long contexts can also degrade LLM tool calling performance, and MCP provides no limit on the length of tool responses. We identified some

tools that returned more than 128,000 tokens in a single response, overflowing the available context of models like GPT40 and reducing the number of possible tool calls for other long context models like Gemini. Finally, different models need to be prompted differently. For example, OpenAI recommends providing in context examples of tool calls for chat completion models, but discourages them for reasoning models. An MCP server generally does not know what model is connected to its client. And so its tool descriptions may work better for some models than others. So what can you do?

As a user of MCP servers, you can use the MCP interviewer tool to test servers before using them. As the developer of an MCP client, you can intercept long tool responses before submitting them to your LLM provider. As an MCP server developer, you should expose as few tools as possible, have short tool responses, unique and descriptive tool names, and you should report what models and clients you tested your server with. MCP marketplaces should also test uploaded servers and report their findings and even reject servers not meeting certain minimum criteria like exceeding maximum tool counts. To learn more, please read our blog post and check out the MCP interviewer on GitHub.

This is amazing work. Understanding how AI systems succeed and sometimes fail is essential to making them truly dependable. That is especially key for our next talk. One of MSR's many priorities is research focused on applying technology to advance human health, safety, and well-being, and to deliver tangible societal impact. To share how AI and cloud computing can make health care faster, more efficient, and more equitable, we'll hear from Ilana Rosenberg, principal product manager in Microsoft Research Health Futures. Eliana will introduce how the Tiger framework enables faster, more accessible MRI imaging by streaming raw data to the cloud for real-time reconstruction, helping reduce patient weight times, and making advanced medical imaging available to more people. Let's check it out.

My name is Iana Rosenberg, principal product manager in Health Futures, where our mission is to empower every person on the planet to live a healthier future. Today, I'll explain how our open-source Tiger remote signal processing framework is transforming MRI imaging by accelerating reconstruction, decoupling programming from the instrument for rapid prototyping, and making MRIs more accessible across the globe. I want to acknowledge the team behind this technology who continue to push boundaries with AI in medical imaging, led by senior director of health futures, Michael Hansen. Let me start with a simple reality for three individuals who sit in the imaging workflow. The first, the developer.

Developers build applications that can be used to advance MRI scans including image reconstruction and denoising techniques that aid in motion correction. Today, code is developed in isolation from the scanner and prototype testing can take weeks if not longer as the code is needed to be first deployed onto the scanner. Second, the patient. A typical MRI exam can take 20 to 60 minutes with complex exams even longer causing not only patient discomfort but also long wait times to get an MRI. And third, the radiologist. There is an increasing shortage of radiologists and burnout is common. Radiologists review on average 20 to 40 exams per day, each with many individual scans. Review on complex cases can take even longer.

Tiger, developed in Microsoft Research, alleviates these challenges. Tiger is an open-source and cloud-based remote signal processing framework that enables online MRI reconstructions. It enables rapid prototyping of algorithms in minutes rather than weeks due to flexibility of where the code can be run. It improves patient experience by accelerating reconstruction and thus reducing scan time. and it optimizes image interpretation due to the ease of leveraging advanced denoising and other modern algorithms due to the MRI scan.

Here's how. As the MRI scan is being completed, images are streamed to the cloud real time. Data from the scanner is divided into blocks and uploaded as buffers to Azure blob storage, enabling reliable, high throughput and asynchronous data transfer between the scanner and the cloud. Think of buffers as a sequence of files that can be tagged for easy retrieval and organization. Next, Kubernetes orchestrates the compute drops by launching containers that read from the buffer, processing the data, and writing results back to the output buffers.

Tiger acts as an abstraction layer over Kubernetes for tasks such as job scheduling, resource allocation, and job execution per predefined specifications. The resulting output buffers are stored as files on Azure blob storage. Output buffers are flexible and can hold any data type produced by the job. For an image reconstruction job, the reconstructed images are now sent back to the scanner to be viewed on the MRI console for interpretation. We are partnering with Seaman's health and ears to scale and integrate the tiger technology into seaman's health and ears MRI scanners through the open recon prototype framework. Tiger is currently available as a work in progress package for customers to evaluate. One common application of Tiger is for accelerated and scalable image reconstruction. Scale opportunity is huge. On average, there are 100,000 MRI scans completed per hour in the United States. Tiger alone can do

250,000 reconstructions per hour, way more than what is needed across the entire US for that same time frame. Connecting these powerful cloud-based computing resources that allow scale to the scanner also enables MRIs to be usable and more accessible across the globe by shifting the complex MRI image reconstruction from local scanners to the cloud. Tiger also makes items such as advanced image correction and enhancement feasible even in low resource environments. In addition to scale, Tiger is transforming MRI development by removing the bottleneck of scanner dependency.

Instead of waiting weeks to deploy code, developers can prototype reconstruction and denoising algorithms in minutes using Tiger's cloud-based compute. This flexibility running the same APIs on a laptop, local cluster or Azure means faster innovation and quicker implementation of advanced techniques like motion correction and AIdriven denoising. One such application is the advanced denoising model for MRI images that was recently developed and published by the diagnostic imaging team in Microsoft Health Futures. On the left of this image, you see a late GD enhancement cardiac MRI. It shows significant noise where the heart is not even visible, something that is unfortunately common, especially in low resource imaging scenarios. On the right side is the same scan after applying our denoising model using Tiger. The heart is now visible for cardiologists to review. This is just one example of how Tiger can run an

application on an MRI scan without needing to first deploy the code onto the scanner itself. Tiger is designed as a general framework for scalable remote signal processing and is not limited to MRIs. It can be applied to any instrument or modality where data needs to be transmitted and processed remotely. This includes but is not limited to CTS, ultrasounds, emergency response with drones, and even transportation infrastructure. The only requirements are the ability to read and write files and build a Docker container that runs on Linux. Thank you for joining the Microsoft Research Forum. Please check out these links to learn more about Tiger and our health futures MRI image denoising model.

Wow, it's always incredible to see how AI can directly improve the quality of care and life. At Microsoft Research, we see fresh innovations emerge out of the labs, ready to reshape how people and organizations harness technology in their everyday work. A fellow Redmondbased colleague, Cheng Lang Wong, principal researcher, works with his team to reimagine how analysts interact with data by combining natural language and intuitive interface design. They introduced data formulator. This tool helps users guide AI collaboratively to visualize, explore, and truly vibe with data in control. Over to you, Chang Wong.

Hello everyone. I'm Chunlong from Microsoft research relement. Today I'd like to introduce you data formulator. It's an interactive AI powered visualization tool for you to explore data with AI agents in control. Why do we build this tool? For example, given a data set about US product price from 2005 to 20 2025, we may be curious about how much does a product price grow over the years as everything around us has become more expensive lately, right? The best way to answer this question is to explore the data and show insights with some visualizations.

But how can we explore effectively? To get this insights, we need to go through an iterative data exploration process. We first need to formulate plans based on our data. Then we need programming skills to transform and implement the visualizations. We often don't solve it in one step. We may need to go back to take up another branch to backtrack and revisit some steps. After a few iterations, we should find some insights and we are ready to share. This process however can be challenging. It requires a user to have both data science expertise to formulate good questions and also require good programming skills to implement this designs.

How can we make this accessible to everyone who is interested in data? Luckily, with AI agents that can generate code from user instructions, coding is no longer a big barrier. However, to make data analysis truly accessible, we still need to lower the interaction friction between user and AI. The first challenge is that natural language can be quite universal but it can be verbose for describing the visualization intent and it may not be very precise. The second is that chat interface is a way how we go with interact with agents but it doesn't support iteration naturally as we want for data analysis. Third agent can be difficult to control if they are working in a black box. The user may easily get lose control over what agent's doing. In order to solve this problem, we introduce data formulator. It's an interactive data analysis tool for people to work with agent effectively to explore data. It first features a

multimodel interaction. The multimodel UI combines UI interaction and natural language so that user can specify their design both precisely and concisely. Second, to support iteration, we organize exploration into threads, allowing user to backtrack, follow up, and revisit. On top of that, we develop agent mode. Instead of running in a black box, it operates on top of data thread so that user can take control whenever they want. Let me show you the experience of using data formulator to explore data. Let's explore how much do US product price increase over the decades.

Uh let's first load the data. This data we can start with agent mode. Ask AI agent to automatically explore the data on our behalf. The agent first generated a plan and then is trying to generate code to execute the plan by transforming the data and visualize it. This is the first chart recommended by the agent. It's a line chart for temporal price trends. Wow, everything has been more expensive now. It now recommended a second chart. It's a bar chart to show the product volatility. It seems act price is most volatile. Now we can take in control. We can ask a follow-up question using natural language. How does COVID and 2008 crisis affect prices?

It now generates a stack bar chart. Not something the best way to visualize it. We can use UI to change it to a group bar chart. Now it's clear. It seems that CO affect prices more significantly. That's very interesting. We can now go back to the previous branch to explore something different. Hm. What should we explore here? We can ask the AI agent for some ideas. It seems it's interesting to see the price correlation with gas price. Let's go with that. Now we get a bar chart to show the price correlation with gas price. It seems that most products prices are correlated to the gas except tomato.

We can continue to explore but now we can also write a report to share our findings. We can ask AI agents to generate a report grounded on our data. It will leverage the computation data and charts to compose a report. As you can see with AI agents automating the implementations we can easily vibr our data. The blended UI plus natural lang interaction approach allow us to easily work with agent. While agent is automating both steps we can easily take control when we need to go to the exploration path we like. So this is data formulator. It is interactive tool for people to explore and visualize data. It features a multimodel interface that make it easy for user to specify visualization ideas.

It provide a new design data threads so that user can perform nonlinear interaction with AI agent in exploration task. Thirdly, the agent plus interactive modes allow user to vibe with data but still having full control. We invite you to try data formulator on data formul.ai. Explore with some data and let us know what you find. And that brings us to the close of our final forum episode of 2025. A huge thank you to Ida, Josh, Tyler, Ilana, and Shanglong for sharing their incredible work spanning the full spectrum of Microsoft research from open explorations to societal impact.

Together, these talks highlight what defines our organization. Multidisciplinary missiondriven research that pushes boundaries while staying grounded in purpose. Whether it's advancing AI, improving health, securing trust, or creating intuitive tools, the common thread is impact. Research that moves the world forward. And if you'd like to stay connected with what's next for Microsoft Research, new discoveries, conferences, and future forum episodes, be sure to follow us on social, and to subscribe to our newsletter. Thank you for joining us, and from all of us at Microsoft Research, happy holidays, and we'll see you next year.

baby. Baby, daddy.

Questions about this episode

Who speaks in "Microsoft Research Forum | Season 2, Episode 2"?

Karen Easterbrook (Host), Ida Momenad (Guest), Josh Benalo (Guest), Tyler Payne (Guest), Ilana Rosenberg (Guest), Chunlong Wong (Guest).

What is "Microsoft Research Forum | Season 2, Episode 2" about?

The content delves into the innovative research at Microsoft focused on the intersection of neuroscience and artificial intelligence, specifically exploring how brain-inspired architectures can enhance the multi-step reasoning capabilities of large language models. Researcher Ida Momenad discusses how collaborative mechanisms akin to neuron interactions can improve planning and reduce…

What is mentioned in this Microsoft Research episode?

13 people, works and ideas are mentioned, including Microsoft Research, Josh Benalo, Ida Momenad, Tyler Payne, Tiger, Karen Easterbrook. Each one on this page links to its own page with every episode that mentions it.

How long is "Microsoft Research Forum | Season 2, Episode 2"?

The episode runs 34 min; the transcript is about 4,785 words in 42 timestamped sections.

More from Microsoft Research

About this transcript: the episode's spoken words as captioned on YouTube, split into timestamped sections. Reclip.Me locates every person, work and idea in the text, keeps the time it is first named, and quotes the words around it verbatim. Last updated January 24, 2026.