Academic Research
You are currently viewing the Computer Science Lens. Click on the other lens to view it.
For AI to play a more integrated role in human tasks, it must align with human values and decision-making processes. Doing so, I believe, requires a radically different approach to conceptualizing and training AI; it requires an AI that understands, on a fundamental level, why humans make the decisions we do. My work illuminates this process through simply observing the creative outputs humans produce.
I imagine a society where professional journalists can produce more higher-quality information at lower costs, improving our information ecosystem, better informing the public and improving our democracy. My work has targeted different steps in the news gathering process, from finding story ideas, to sourcing, to structuring the finished story, to editing it, and I develop computational methods to improve each step.
I am currently developing a line of work exploring tacit knowledge: how to measure it, learn it and understand what role it has in human tasks.
Tacit knowledge, I believe, is how humans construct taste, meaning and values; our models do not currently acquire tacit knowledge in the same ways humans do, and I believe that is the root of safety and capability failures at the frontier.
I am currently asking how much human preference can be captured by rubrics and verifiers across seven creative domains (journalism, law, math, code, etc.) and find robust evidence of tacit knowledge, even in math and code. This builds from a long line of observations I have made in how human preference gets encoded in homepages, news coverage selections and source selections. Understanding how these qualitative shared values arise, and how they can be encoded, presents in my mind the biggest challenge to continued frontier model development.
I believe modern agentic systems benefit when considering more elements of planning. Goal-directed RL, constrained planning and latent learning are all directions that I am interested in considering.
My prior work in this domain has considered hierarchical approaches to mapping trajectories, learning from human planning processes and fair low-level plans.
Most recently, I have studied planning where the plan is a conversation: in NewsInterview, we built a simulated interviewing environment from 40,000 journalistic interviews and found that LLM interviewers rabbit-hole and fail to pivot, a long-horizon planning failure that no single-turn benchmark would show. Before that, I asked whether LLMs plan like human writers when choosing an angle on a press release (they match us better on angles than on sources, and are less creative at both), and treated a journalist's choice of sources as a latent plan that can be recovered from the finished article.
That work grew out of earlier attempts to impose structure on generation and to read structure off of human drafts: sequentially controlled text generation asks a model to follow a discourse plan sentence by sentence, and NewsEdits captures 4.6M article versions so we can watch a writer's plan unfold edit by edit. The earliest thread, actionable recourse, is planning at its most constrained: what is the cheapest sequence of changes a person can make to flip a classifier's decision?
Humans leave traces of our own planning and ideation in the artifacts we create: you or I could look at a scientific paper and reconstruct many steps in its execution (explicit and implicit), based on what we call discourse cues. Cognitive scientists call this emulation, or the ability to (a) work backwards and recreate missing steps in the state-and-action sequences of others and (b) recreate them.
I centered my thesis around this topic: that our ML systems can learn better from human actions if we take some of these processes better into account.
My thesis frames this as emulation learning: infer the actions and rewards behind a finished artifact, then learn to plan from them. The most recent pieces read discourse to recover intent: NewsEdits 2.0 separates factual from stylistic edits and shows an article's own language predicts which facts will change, and DiscoSum plans a discourse structure before it summarizes, so a story keeps its shape across platforms. The same move works outside news: which patent claims survive examination is predictable from textual entailment against the prior art.
The discourse machinery itself came first. LegalDiscourse asks when a law applies and to whom, and answers with a span-and-relation schema that journalists then used, through StateCensusLaws.org, to investigate 6,000 census-dependent state laws. Underneath that is the practical problem that discourse datasets are small and their schemas drift, which multitask learning helps with where the usual low-resource tricks do not, and a roadmap for why the structure of a news article is a scaffold for computational journalism in the first place.
I have been fascinated by sequence modeling both in and outside of language, and what this tells us about fundamental similarities between domains.
Most recently, with EleutherAI, we trained autoregressive transformers on 60,000 hours of transcribed solo piano; the models produce coherent continuations and MIDI embeddings that set the state of the art on music-information-retrieval classification, and Aria-Duet puts one on a Disklavier for a live, turn-taking duet with a human pianist. Before that, we showed that classifier-free guidance, borrowed from image diffusion, works for plain language models too: at inference time it buys the accuracy of a model twice the size.
Outside language and music, the same tools show up in physics and public health: masked autoregressive transformers predict plasma disruptions in fusion tokamaks and suggest the plasma carries a long memory, and, early in the pandemic, co-training on event-extraction views let classifiers transfer from COVID-19 policy announcements to Twitter discussion of them.
Journalists find stories by reading: leaks, bills, court filings, press releases, meeting minutes. Only a sliver of it becomes news, and the judgment of what deserves coverage is one of the most important, and least articulated, decisions in the newsroom.
Most recently, we showed that editors' judgments are legible in the layout of a homepage: from 3,000 homepages captured twice daily for five years, we learned stable, transferable preferences over what goes where. Before that, we linked San Francisco Chronicle coverage to city-government records and built newsworthiness predictors that expert journalists found helpful 84% of the time, extending an earlier idea of training on front pages and transferring to court cases, bills and council minutes.
Looking at the other end of the pipeline, we asked whether LLMs choose angles on press releases the way reporters do: they are closer to us on angles than on sources, and less creative at both. The newest piece, on tacit components of preference, treats newsworthiness as one of several human judgments that resist rubrics and verifiers.
Reporting is sourcing. Which people, documents and databases a journalist consults shapes the story more than any sentence they write, and sources tend to be used together in canonical ways.
NewsInterview is the most recent step: a simulated interviewing environment built from 40,000 journalistic interviews, in which LLM interviewers rabbit-hole, fail to pivot and leave information on the table. Alongside it, we recast source selection as multi-document retrieval and showed query planners do better when they plan over the discourse role each source will play, and we modeled a journalist's choice of sources as a latent plan recoverable from the finished article.
The foundation is the largest annotated dataset of informational sources in news, which grew out of first steps toward a source recommendation engine and, before that, a model of articles as mixtures of source types labeled by affiliation and role.
A news article has a shape: a main event, context, history, reactions, expectations. Journalists learn this structure by apprenticeship; I want machines to read it off finished stories and to write with it.
Most recently we took the discourse approach to law: LegalDiscourse parses when a law applies and to whom, and StateCensusLaws.org put those parsers in front of journalists investigating 6,000 census-dependent state laws. In news itself, sequentially controlled generation asks a model to follow a discourse plan one sentence at a time and gets closer to human-level writing the more structure it is given.
Underneath is the practical problem that discourse datasets are small and their schemas drift, which multitask learning helps with where the usual low-resource tricks do not, and the roadmap that argued news discourse structure is the scaffold for computational journalism tools in the first place.
Breaking news is written in versions. Each revision records a decision: a fact that changed, a frame that shifted, a quote that was cut. Those traces are a rare window into the writing process itself.
Most recently, NewsEdits 2.0 separates factual updates from stylistic and narrative ones and shows that an article's own language predicts which facts will change, well enough to tell an LLM when to abstain from answering. DiscoSum treats re-writing a story for each social platform as a structured edit, planning the discourse structure before summarizing.
The dataset behind this is NewsEdits, 4.6M versions of 1.2M articles, with tasks that ask models to predict which sentences will be added, deleted or edited next; the same edit-prediction framing carries over to patent claims, where textual entailment against prior art predicts which claims survive examination.