Research spotlight · four works
A closer look at four recent machine learning and AI works—and the shared instinct behind them: use context carefully, keep systems practical, and make decisions more useful.
The through-line
Read together, these four works trace a path from understanding what a model should forget, to scaling how it organizes information, to using context to decide what a person sees next.
A research arc
Each work addresses a different pressure point in the journey from raw data to a better outcome.
Shape model behavior at inference time, with uncertainty made explicit.
Organize massive collections without giving up the quality of the structure.
Turn a query and its conversational history into a clearer search task.
Use context to make the content around a result more helpful.
The four works
Two papers establish methods for learning and scale. Two patents carry that same attention to context into search experiences.
Machine unlearning usually means retraining. This work moves the operation to inference time: a verifier checks generated responses, then conformal prediction helps guide refinement without updating the model’s parameters.
Why it matters It offers a more flexible route to removing targeted information while preserving general capabilities, with distribution-free guarantees.
Hierarchical agglomerative clustering produces useful structure, but traditional approaches are difficult to run at web scale. This work introduces a scalable method designed to preserve cluster quality, with theoretical recovery and approximation guarantees alongside strong empirical results.
Why it matters It brings high-quality hierarchical clustering to billions of data points, including a demonstration on a dataset of 30 billion queries.
In conversational search, the latest query may only make sense in light of earlier turns. This work uses the current query and multi-turn history to infer intent, rewrite the query with missing context, and use intent-aware representations to improve search result determination.
Why it matters Search can respond to what a user means across a conversation—not just the last string they typed.
This patent uses a language model to condition a resource description on the current query, related search history, and the resource itself. The system can generate clearer headlines, bullets, summaries, or explanations for the moment at hand.
Why it matters Relevance is not only which result appears—it is also how that result is explained.
Take the longer view
The details differ—unlearning, clustering, query understanding, and generated descriptions—but the goal is consistent: build learning systems that respect uncertainty, scale with the data, and meet people where they are.
See every publicationSummaries are based on the linked paper abstracts and patent descriptions. View the full citations.