AI researcher Jennifer Neville reveals why language models stumble in real-world conversations

Jennifer Neville, a partner research manager at Microsoft Research and professor of computer science and statistics at Purdue University, has spent her career understanding where artificial intelligence systems fall short in practice—and how to fix them.

Through a winding academic path that took her from mathematics to physics to cognitive science before landing in computer science, Neville has become a leading voice on AI evaluation and real-world performance. Her journey reflects the broader challenge she now tackles: getting AI systems to work reliably when they encounter the messy, unpredictable conditions of actual user workflows.

The Frontier Between Theory and Practice#

Neville's research spans the spectrum from theoretical algorithms to applied systems. After earning tenure at Purdue, she spent her first sabbatical exploring advanced theory at the Simons Institute in Berkeley. Her second sabbatical, however, redirected her career. Joining Microsoft Research allowed her to witness something crucial: watching algorithms operate not in textbooks or controlled experiments, but in real systems with genuine users and actual data.

"Being able to see how algorithms in theory hit the road with respect to how they behave in practice—that's hard to turn back from," she explains. This experience prompted her to remain at Microsoft while maintaining her academic appointment, positioning her to bridge the gap between theoretical innovation and practical deployment.

Evaluating Beyond the Benchmarks#

The AI Interaction and Learning team at Microsoft Research, which Neville leads, focuses on mapping the boundaries of current AI performance in realistic work environments. They've identified a critical problem: standard machine learning benchmarks—the tests used to validate AI systems—often bear little resemblance to how people actually use these tools.

Researchers typically evaluate models on relatively simple, self-contained tasks. But users operate differently. They work across multiturn conversations where requests evolve, collaborative workflows where humans and AI iterate together, and long-horizon tasks that unfold over time. A system that scores perfectly on a benchmark can fail dramatically in these scenarios.

When Models Lose the Thread#

Two recent research projects illustrate these hidden vulnerabilities. In studies of multiturn conversations, Neville's team discovered that language models perform significantly worse when users clarify and refine their requests across multiple exchanges. Single-turn performance often appears nearly perfect—because models are optimized specifically for that test format. But introduce realistic back-and-forth interaction, and accuracy drops noticeably.