SDSignal Desk

Microsoft researcher Jennifer Neville on how AI degrades over multi-turn chats and long edits

· Updated · Signal Desk newsroom

Analysis of reporting by Microsoft Research. How we work

Image: Microsoft Research

On the Microsoft Research Podcast, Jennifer Neville described findings that models perform significantly worse when a task is spread across several clarifying turns, and that repeated document edits accumulate small errors that subtly erode meaning, failures single-turn benchmarks rarely catch.

Key takeaways

  • Neville leads Microsoft Research’s AI Interaction and Learning team and holds a chaired professorship at Purdue.
  • Tasks handled well in one fully specified prompt degrade significantly across multiple turns.
  • Long workflows such as repeated document edits accumulate subtle, uncaught errors.
  • She advises checking answers and restarting derailed conversations with one complete prompt.
  • Her team uses the gaps it finds to improve models, including via reinforcement learning.

What the research found

On the Microsoft Research Podcast, partner research manager Jennifer Neville, who also holds a chaired professorship at Purdue, described her team’s work on how AI behaves in realistic, multi-step work. One finding: when a task that models handle well in a single, fully specified prompt is instead spread across several turns of clarification, performance degrades significantly.

Her team also studies long workflows such as repeated document edits, where small uncaught errors pile up and leave a subtle loss of meaning. Neville leads the AI Interaction and Learning team, which designs evaluations around multi-turn, collaborative and long-horizon tasks and then uses the gaps it finds to improve models, including through reinforcement learning.

Why these failures matter

The research targets the gap between how models are tested and how people use them. Real users rarely spell everything out in the first message; they work out what they want as they go. If models get worse exactly as people behave naturally, that is a design problem rather than a user problem, and single-turn benchmarks will overstate real-world performance.

Document drift may be the more consequential finding. A hallucinated citation can be checked. Content that slowly loses meaning across dozens of agent edits is not dramatic enough to notice and may never be flagged. As agents become a common way to revise contracts, reports and code, this kind of drift could degrade work before anyone measures it. Products moving in that direction include Google’s workplace agent (Gemini work agent). Microsoft has also published conduct rules for its models (code of conduct), but reliability over long tasks is a separate engineering problem.

Practical advice and its limits

Neville’s guidance is practical: check answers, do not treat these systems as ready for full delegation, and when a conversation goes off the rails, start over with one complete prompt. She also encourages people to give detailed feedback when things go wrong, which helps improve models but asks users to do quality work on vendors’ behalf. She says her team analyzes usage only at scale, in privacy-preserving ways, without viewing individual data; documentation of those practices would help users assess the claim.

The broader takeaway is that AI is useful at work today with a human in the loop, and evaluation research of this kind is how it becomes more trustworthy. Neville’s own lesson for agent builders is to look at the data. Agent products would benefit from rollback and self-checking for long tasks, and benchmarks that measure gradual degradation. More in our Microsoft coverage, research coverage, agents coverage and future of work coverage.

Who feels it

Knowledge workers
Restarting with one complete prompt is a practical fix; checking outputs still matters.
Agent builders
Long workflows need checks for gradual drift, not only obvious errors.
Evaluation researchers
Single-turn benchmarks overstate performance in messy real use.

What to watch

  1. Published results from Microsoft’s reinforcement learning work on multi-turn reliability.
  2. Benchmarks for gradual document degradation.
  3. Agent products adding rollback or self-checks.

FAQ

Who is Jennifer Neville?

A partner research manager at Microsoft Research leading the AI Interaction and Learning team, and a chaired professor at Purdue.

What happens in multi-turn conversations?

Performance degrades significantly compared with a single fully specified prompt, her team found.

What is document drift?

Small uncaught errors that accumulate over repeated edits and subtly erode meaning.

What does she recommend?

Check answers, avoid full delegation and restart derailed conversations with one complete prompt.

Read the original

Continue at the source.

Microsoft Research

Companies: Microsoft

More on Microsoft

All Microsoft coverage →