Learning Taste to Augment Scientific Discovery

A research lab generates more good ideas than it can pursue. AI systems can process more information than any human, and execute experiments with increasing autonomy. The scarce resource is not ideas or execution, but judgment: the taste that selects which question is worth asking, which approach is promising, and which surprising result matters. This project asks whether that judgment can be learned and applied by AI systems to accelerate collective discovery.

Two questions organize the work. First, how should a research community's taste—its sense of what is worth pursuing and why—be represented so that a machine can learn it? Our working hypothesis is that taste compresses the reasons behind a community's preferences rather than the objects it has preferred. Second, which learning methods can recover and update such a representation as the community itself changes?

We will study these questions inside Chorus, an existing AI research infrastructure embedded in the University of Chicago's Knowledge Lab, and seeded with the lab's accumulated papers, projects, data, methods, open questions, and work in progress. The system is designed to surface connections across siloed teams and knowledge.

The first phase of the project built datasets and pipelines for contextualizing the lab including internal people, knowledge and resources as well as candidate questions from lab meetings, notebooks, papers, and proposals; and building interfaces between humans, AI, and resources.

This project is focused on the second phase: learning and feedback. We will begin with a controlled task in which an AI system receives a pool of candidate research questions, predicts which ones a scientific community would find most promising, and explains why. Independent AI judges provide a cheap initial signal so that alternative representations and learning strategies can be tested rapidly. We then replace this synthetic signal with corrections from scientists: the system commits to a ranking and explanation first, researchers identify where and why it is wrong, and those corrections become training data.

The central research problem is to characterize which types of internal representations allow for the expression of preferences over research choices and the associated reasons for those preferences, such that these representations can be updated over time through continuous human feedback. We will explore and characterize the behavior of a variety of representations to encode reusable principles of scientific judgment in language, along with algorithms that update those representations. We will compare approaches using prediction accuracy, the quality of stated reasons, and their ability to improve with feedback.

We will not train a single system toward an averaged consensus. Different researchers and communities may share standards of rigor while valuing different questions. We will therefore preserve multiple learned perspectives and test whether locally learned systems retain meaningful differences in attention rather than converging on one global ranking.

Evaluation proceeds in stages: inexpensive model-based tests; judgments from participating scientists; and independent expert assessment of whether surfaced questions are rigorous, surprising, and important. We will also compare what can be inferred from the lab's archival record with what is learned through direct correction, testing the project's central claim that records of past attention do not fully reveal the reasons behind scientific judgment.

The project will validate this process end to end at the level of research-question selection. Generating fuller research proposals is a stretch goal. Autonomous experimentation is outside the scope: greater initiative must be earned by evidence that the system has learned to exercise judgment at the preceding stage.

AI is rapidly increasing scientists' capacity to generate and pursue ideas, but generation is not the same as judgment. Hao et al. (Nature, 2026), analyzing 41.3 million papers, find that scientists using AI publish more and receive more citations while scientific attention collectively narrows toward data-rich topics. The problem is therefore not simply whether AI can accelerate discovery, but whether it can help decide what is worth discovering without making inquiry more uniform.

Scientific taste—the tacit judgment of what is promising, surprising, or important—is largely transmitted through participation in a community. Published and archival records capture what received attention, but omit negative evidence that shapes judgment: ideas considered and rejected, projects abandoned, and reasons one direction was preferred over another.

Two developments now make this missing layer of judgment empirically tractable. Modern language models can state a prediction about why something matters and learn from correction; Chorus provides a laboratory in which those predictions can be tested against practicing scientists. We will test whether a system that guesses and is corrected can learn a community's tacit judgment while preserving differences within and between communities.

Mentees will build the scaffolding/harness layer around existing LLM API infrastructure to represent the preferences of the humans on the team, and their reasons for those preferences. The work will adapt assistance games (Russell, 2019) as the underlying decision-making framework, and leverage probabilistic dependency graphs (Richardson, 2024) as the belief reconciliation framework.

Mentees will have a moderate-to-high level of autonomy in how they go about their daily tasks. They are expected to attend a 1-2 weekly project meetings (more can be scheduled as needed) and coordinate with mentors and other researchers working on the same project.

• ability to implement and debug reinforcement learning algorithms
• understanding of multi-agent RL, recommender systems, and/or human-robot interaction
• understanding of Bayesian networks and probability theory
• experience with fine-tuning LLMs (RL fine-tuning and traditional)
• 2+ years experience writing high-quality, extensible code

Ideally local to either UC Berkeley or University of Chicago (but one or more shorter visits would also be acceptable)