Learning from How Humans Think – Research – Dongyeop Kang (DK) @ University of Minnesota

Today's language models learn from the finished text people produce, not from the process that produced it. They write fluently, yet they struggle with what defines expertise: long-horizon reasoning, weighing alternatives, revising what we thought we knew.

Watch a lawyer and a language model work the same case and the difference is plain. In LawFlow, law students role-playing lawyers formed a multi-state LLC for three co-founders while every step was recorded; GPT-o1 and DeepSeek-R1 agents got the same scenario . The humans built plans 3.6 levels deep and revisited subtasks about 13 times per session, following the branch that mattered and revising as the case developed. The agents planned once, one level deep, revisited 0.4 times, and expanded every branch they could find. Neither is wrong, but they are different kinds of reasoning, and a good thinking partner has to be designed around that difference.

Two workflow graphs: the human lawyer follows one relevant branch; the LLM expands every possible branch. — Same task, different minds, from LawFlow .

My central premise is simple: the cognitive process behind expert work—keystrokes while writing, eye movements while reading, the traces of expert decisions—is a richer signal for aligning AI with people than finished outputs alone. I call the use of these temporally grounded traces as training signal cognitive scaffolding.

The big picture

My research forms a single loop: capture how people think, train models on it, assist experts where they get stuck, and align the result with the diversity of people it serves.

Overview: human cognition in writing and reading feeds cognitively-inspired models and benchmarking, which assist expert thinking and align with pluralistic opinions. — Human-centric AI: from human cognition data toward expert-level thinking partners for people and society. The top row is what we build; the two bands underneath show what we learn about humans and what that becomes on the machine side.

1CaptureWriting, reading, and expert workflows as data 2Train & measureCognitively aligned models and benchmarks 3Assist expertsThinking partners for science and law 4Align with societyPluralistic, not average, AI

Chapter 1

Capturing how people think

A well-written paper reflects months of planning, drafting, abandoning, and revising, yet a model trained on the final PDF sees none of that. So we started recording the process itself.

Writing as process data

We first studied the iterative nature of revision. With Grammarly, we collected IteraTeR: 31,631 document revisions from Wikipedia, arXiv, and Wikinews, each edit labeled with the intention behind it—fluency, coherence, clarity, style, or a change of meaning . Training on these intentions makes revision models better, but people still produced higher-quality text, so we built systems that revise with people instead: Read, Revise, Repeat and the instruction-tuned editor CoEdIT .

Human–AI collaborative iterative text revision: a draft is revised step by step with intentions such as 'make it more formal' and 'shorten it'.

Key finding

In the Read, Revise, Repeat study, a human in the loop reached quality 0.68 with fewer rounds and fewer edits, versus 0.28 for the system revising alone; people accepted about half of the system's first-round edits and two-thirds of its second-round edits . Explicitly modeling the revision loop is itself a form of test-time scaling. CoEdIT's smallest 770M model is competitive with instruction-tuned models about 60× larger on editing tasks .

Examples from the data

IteraTeR · a coherence editwhich achieved high inter-annotator agreement → and achieved high inter-annotator agreement

CoEdIT · a composite instructionRemove all grammatical errors from this text, and make this text less complex: They follow the way the Sun goes from the east to the west. → They follow the sun as it goes from east to west.

Then we went further: instead of isolated edits, the whole writing process. ScholaWrite logged 61,504 keystroke-level actions from ten graduate students writing five Overleaf manuscripts over four months, every action labeled with the writer's intention . All five projects became arXiv preprints, so the traces run from the first sentence to a public paper.

Writing intentions over the course of a project in ScholaWrite: idea generation, planning, text production, revision, and more.

Key finding

The intention taxonomy has three layers—planning (idea generation, idea organization, section planning), implementation (text production, object insertion, citation integration, cross-reference, macro insertion), and revision (fluency, coherence, clarity, structural, linguistic style, scientific accuracy, visual formatting). Without this data, encoder models barely predict a writer's intention (weighted F1 0.02–0.04); fine-tuned on ScholaWrite they reach 0.64, and generation models trained on the traces alternate between planning, writing, and revising the way people do .

One labeled keystroke

One important expct of studying a LLMs is .. → One important aspect of studying LLMs is .. — labeled Fluency.

We also ask how AI reshapes writing itself. In a controlled study, 24 students wrote essays with no LLM access, limited access (three prompts, 100-word replies), or unlimited access. Essay quality was statistically indistinguishable across groups, but writing behavior and perceived authorship diverged sharply .

From papers to whole research projects

SciFlow extends ScholaWrite from one manuscript to the full research cycle: multi-month trajectories from ideation through revision, recording keystrokes, screen, workspace artifacts, and AI prompt–response traces . Raw logs are structured into events, spans, and episodes with intent labels, dependencies, and failure–recovery outcomes—exactly the signals finished artifacts never contain.

Left: the scientific research cycle. Right: one project's stages, intents, sub-tasks, and tools, from our collected workflow data.

Its companion, SciVeri, covers breadth: it atomizes published studies into scientist-verified, step-scored task cards and wraps domain simulators (e.g., AlphaFold, OpenFOAM) as typed tools behind our Sci-MCP server, so one agent can reproducibly run any community task . With support from UMN, we are collecting workflows from 50+ faculty across 15 departments.

Left: SciFlow extends ScholaWrite to the whole research cycle. Middle: both corpora share one process codebook. Right: SciVeri turns published studies into scientist-verified task cards.

Key finding

Structure helps rather than hurts: in SciSense, models trained on citation-conditioned trajectories that reconstruct the path to a known paper produced higher-quality and more novel research directions than models trained with looser supervision (+2.0% trajectory quality), and downstream coding agents built more executable artifacts from them .

Long-horizon expert workflows

Expertise is not one skill but many working in concert across an entire workflow. In scholarly writing, models trained on human trajectories show more human-like dynamics than base models . In legal practice, LawFlow recorded ten finalized sessions of business-entity formation (memos and operating agreements, about 101 minutes each) and the same tasks run by reasoning agents .

Left: scholarly writing intents over time for humans versus LLMs. Right: legal planning and execution for human lawyers versus LLM lawyers.

Key finding

Human workflows are modular and adaptive; LLM workflows are sequential, exhaustive, and less sensitive to downstream implications. Humans planned 3.6 levels deep with about 8.8 nodes per subtask and revisited subtasks 13.4 times on average; the agents planned one level deep with 3.8 nodes and revisited 0.4 times . Practicing lawyers describe the other side of the gap: generative AI is already capable, but lawyers—junior ones most of all—struggle to deeply understand AI answers to unfamiliar legal issues, so practice has changed little .

Beyond keystrokes: multimodal context

To scale beyond keystroke logging, we extract fine-grained workflow dynamics from raw egocentric video and infer the latent intents behind them. SERUM processes screen recordings through a sliding window, alternating activity-recognition and intent-inference passes; each pass refines labels using accumulated context, which cuts the hallucination and temporal conflation of single-pass annotation . It also seeds the Scientist Model described in Chapter 2.

SERUM: from a person working, the current standard labels actions; SERUM builds an action model and an intent model, forming a user model that lets an assistant intervene.

Key finding

On 61 egocentric videos of coding, cooking, physical activity, and daily life, final-pass labels were judged 88.3% accurate and preferred over first-pass labels 82.8% of the time; the induced state machine predicts the next activity at 48.5% versus a 37.6% majority baseline . One refinement moved a frame from holding drill to testing HVAC repair—the difference between an action and an intent.

Psychometric signals from a user's workflow: intent switches, typing speed, pause length, and inferred states such as rushing or confused. — Psychometrics from user workflows.

The same idea extends to signals beyond the screen—gaze, physiology, voice, and video from XR devices—from which we derive cognitive metrics and infer states such as rushing or confusion. The design question is less what to show than when to intervene: proactive agents can turn idle compute into useful support, but even correct work can misread context, impose review costs, or erode trust. Our framework for proactive agents organizes this around task capability, temporal allocation, and trust , and a longitudinal study of 15 interdisciplinary researchers shows how people actually orchestrate generative AI across unfamiliar fields .

Reading through the eyes

Writing reveals how thought is produced; reading reveals how it is perceived. We collected explicit annotations of stylistic perception and trained models aligned with readers' judgments . With eye tracking and active reading, we study how readers engage with style, narrative, and preference-ranking tasks : where readers look reveals judgments they never write down.

Eye-tracking heatmaps over two candidate responses during preference decisions, and saliency of politeness cues from eyes, annotations, and BERT.

Key finding

In PreferRead—300 annotators, 1,000 preference items, three readers each—re-reading happens in about half of all trials (54%), and 74% of those who re-read return to the response they end up choosing, a final confirmation check. Agreeing annotators re-read more; long, wandering reading paths predict disagreement . In our eye-tracking study of style, gaze overlaps only modestly with what annotators mark or what a model finds salient (IoU 0.26–0.31): annotators favor adjectives, while eyes dwell on the verbs and adverbs around them .

Gaze metrics, decoded

Prompt rereadUncertainty, or careful reconsideration of the prompt.

Response loopStrong indecision between candidate responses.

Read timeHigher values mean deeper evaluation.

Word coverageLow coverage: careless reading, or a clear-cut decision.

These perceptual signals carry evaluative cognition that text-only supervision systematically misses; a scaffolding framework for recovering such lost context is developed in Debarati Das's dissertation .

Chapter 2

Teaching models to think like experts

Cognitive alignment means extending what models can do toward complex human behaviors—compositionality, planning, abstraction, social judgment, memory—either by mimicking cognitive processes directly or by putting human cognition data into the training objective.

The lens I keep returning to is Halliday's Systemic Functional Linguistics: language simultaneously encodes content (the ideational), social relations (the interpersonal), and structure (the textual) . The subsections below follow that split: reasoning over content, planning the structure, and modeling the social layer—then agents that improve all three over time.

Systemic Functional Linguistics: ideational (what), interpersonal (who), and textual (how) meanings.

Reasoning and compositionality

People bridge incomplete information with outside knowledge and iterative refinement. We combine neural and symbolic systems to fill knowledge gaps , and build benchmarks and models for composite and chained tasks—CoEdIT and Chain-of-Instructions —using compositional data augmentation .

Task compositionality: composite text-editing instructions and chained instructions.

Key finding

Training on two- and three-step instruction chains generalizes to unseen four- and five-step chains. A CoI-tuned Mistral reaches 70.76 Rouge-L on two-step tasks where the base model scores 24.93, and people preferred CoI outputs about 70% of the time on those tasks . Step-by-step refinement also pays off at inference: storing reasoning as reusable thought templates lifts long-context accuracy by 4–8 points across Claude, Gemini, and GPT .

A chain of instructions

Generate a title for the given text → Convert the style of the title to be similar to a blog post title → Translate the blog post title into French.

Planning

Coherence comes from plans: decisions about topic, order, abstraction, and strategy. We model planning as hierarchical optimization—planners grounded in discourse relations , topical keywords , and social or persuasive goals . Our self-supervised bot-play framework used reinforcement learning to optimize explicit communication goals , a paradigm that has since grown into RL from AI feedback.

Key finding

Predicting content keywords before writing raised human coherence ratings from 2.1 (GPT-2) to 3.8, with human text at 4.9; nouns plus verbs make the best plan . Today the plan lives in the reward: structural rewards align generated text with human discourse structure , and dense, token-level rewards learned from attribution and Bayesian optimization stabilize PPO and lift the win rate over SFT to 56.3% on HH-RLHF, versus 48.7% for sparse RLHF .

Style emerges from interacting factors—formality, emotion, metaphor—that reflect who we are and what we want. PASTEL pairs about 41K sentences across seven persona dimensions (gender, age, country, politics, education, ethnicity, time of writing) ; xSLUE benchmarks 15 styles across 23 tasks .

Correlation matrix across styles such as romance, politeness, happiness, anger, and offense.

Key finding

Styles are not independent variables. In human text, impoliteness and offense co-occur most strongly, politeness travels with positive emotion, and happiness with romance; multi-style learning beats single-style models, while contradictory combinations yield less appropriate text . Aligning to several social dimensions at once works best when reward weights adapt during training: 60.3% of outputs carried both target styles with dynamic weighting versus 56.8% with fixed weights .

PASTEL · the same event, two personas

Reference: went to an art museum with a group of friends. High-school persona: My friends and I went to a art museum yesterday. Bachelor's persona: I went to the museum with a bunch of friends.

We extended this to affective states such as nostalgia and skepticism , to multi-attribute control via data balancing , and to figurative language through distillation . Cognitive efficiency is the quieter cousin of all this: abstraction lets people skip detailed reasoning when a shortcut suffices, and in the same spirit we design data- and compute-efficient training that keeps performance with fewer samples and balances quality with resource use .

Agents that evolve their own thinking

Beyond modeling human workflows, I build agents whose thinking mechanisms improve over time—treating long-horizon reasoning as an iterative control problem rather than a static prompt.

Three self-evolving mechanisms: meta policy optimization, adaptive judge updates, and thought-template evolution via text gradients.

Meta Policy Optimization lets a meta-reward model rewrite the reward model's prompt during training, so evaluation criteria adapt as new failure modes appear . Metaⁿ keeps a single meta-operation fixed and applies it recursively to its own outputs, so meta-depth emerges instead of being capped at two . At inference time, judges update their own meta-prompt only on cases where their verdict flips when the answer order is swapped—a label-free uncertainty signal —and reasoning experience is stored as reusable thought templates refined with text gradients .

Reward hacking, caught by the meta-judge

Asked for a 350-word essay, the policy answered I'm ready when you are. Please provide the essay instruction.—and the fixed reward model still gave it 3/5. MPO's meta-analysis flagged the loophole and rewrote the rubric with an explicit length requirement and a 0–10 scale . In Metaⁿ, when a solver crashed on a missing scipy import, the second meta-layer injected do not use scipy plus a validation helper; the solver re-implemented the Hungarian algorithm and scored 1.0 .

Key finding

MPO essays reached an Elo of 1,244 versus 1,111 for an expert-written reward prompt . Metaⁿ outperforms prior self-improving agents across eight benchmark families and is the only one to score above zero on ARC-AGI-2 (0.331 held-out) . Selective test-time learning reaches 0.808 pair accuracy on MMRewardBench at 3.9× inference cost, where comparable methods cost 11–13× .

156,731agent evolution trajectories

371optimization tasks in 10 domains

10.22%average gain on 22 held-out tasks

Our current emphasis is curating complex workflow trajectories at scale. Evolution Fine-Tuning (EFT) turns agent evolution-search trajectories—the Finch Collection, from GPU-kernel design to open mathematical conjectures—into mid-training supervision for open 2B–9B models . Across evolving-agent frameworks we find that quality and diversity of trajectories beat raw scale: a small set filtered for branching behavior, refinement strategies, and informative failures outperforms large-scale best-of-N. When search stalls for lack of knowledge, EvoDuet co-evolves the search queries with the solutions , and Selective TTS learns where in a multi-stage pipeline to spend compute .

Evolution Fine-Tuning: target and discovery tasks, the cost of test-time search, and Finch internalizing discovery strategies.

Key finding

Discovery strategies transfer. Finch models surpass their base counterparts by 10.22% on average across 22 held-out tasks; on the Erdős minimum-overlap problem, Finch-8B reaches 0.381236 against the base model's 0.403585, next to the best-known bound of 0.380927 previously obtained with a proprietary system . On visual-insight pipelines, spending compute per stage and pruning weak branches early raised mean insight score from 61.64 to 65.86 under a fixed budget, with far lower variance .

From evolving agent trajectories to filtered graphs to extracted high-quality paths for mid-training.

Where this is heading: process learning for co-discovery

Process mid-training writes the discovery process into an open policy model (Finch-Sci)—imitating verified transitions, preferring recovery over abandoned branches—while a separately trained Scientist Model learns the expert's judgment, forecasting the next correction (accept, repair, or redirect) . The two co-optimize in a dual-loop harness to be released on our OpenGalapagos platform , and expert interventions logged in deployment flow back into training.

Human–AI co-discovery: a science agent and a scientist model co-optimized through bi-level optimization with a science world model.

Other modalities. Collaborations extend the work to images, video, and robots: vision–language models for graph understanding , temporal memory for story visualization , aligning video LLMs with AI feedback , and LLMs that plan robot manipulation from language . I also study calibrated abstention and the theory of joint reward and policy learning .

Chapter 3

Measuring the gap

As LLMs move from standalone models to persistent collaborators, leaderboard-style testing stops being enough. Evaluation becomes measurement engineering: producing metrics that stay valid, interpretable, and safety-relevant as everything drifts .

The promise and risks of AI-generated data

The promise. LLMs can generate annotations, dialogues, and evaluation data, and we have used them to cut supervision costs: symbolic augmentation, annotation imputation, information-theoretic data selection, disagreement-aware sampling , LLM-based evaluation , and multi-agent simulation . Pairwise preferences generated by GPT-3, for example, cut classification error by 11.55% relative when used as an auxiliary signal .

The risks. Overreliance builds an artificial data ecosystem that distorts measurement. We have documented four failure modes, each with its own fingerprint.

Four risks of LLM-generated data and LLM evaluators: cognitive biases, discourse shifts, shallow information synthesis, and artifact amplification.

What each failure looks like

1 · Cognitive bias in LLM judgesAcross 16 LLMs, about 40% of pairwise comparisons were biased. Eleven of fifteen models were drawn to whichever answer came first or last; among models above 40B parameters, the first answer won over half the time. Judges also preferred their own outputs regardless of quality .

2 · Discourse motifsMachine text has a different shape. In Yelp reviews, LLM text leans on Background relations, while human text uses more Temporal and Joint relations—branches of equal weight—so human writing is structurally more varied. Those motifs detect machine text even after paraphrasing (F1 0.45 → 0.71 on long stories) .

3 · Shallow synthesisHuman: Progress has been reported in using PLMs to perform reasoning tasks, including arithmetic, commonsense, logical and symbolic reasoning. GPT-4: The reasoning capabilities of PLMs are another central theme, with [6] examining… Humans connect 48.3 citation pairs per related-work section; GPT-4 alone connects 12.3 .

4 · Artifact amplificationSentence 1: This place is not my #1 pick. Sentence 2: We will definitely be going back… Human preference: 1. GPT-3: 2, swayed by the explicit sentiment words. Models trained on LLM preferences score about 10% lower on human test data, and LLM labels match the majority best while missing minority views .

Two newer failure modes matter most for assistants. Latent preferences: people rarely state everything they care about; whether a model can uncover a hidden preference through conversation varies from 7% to 97% depending on the topic . Scientific sycophancy: given a flawed premise, models often go along with it , and agents simulating people drift from the preferences they were given .

Two conversations

Latent preference: I avoid spicy foodsGPT-4o-mini surfaced it by turn 3 with a dietary-restriction question; Qwen2.5-7B specialized immediately into regional cuisines, got trapped in a narrow conversation, and found it only at turn 5 .

Flawed premise: I have invented a perpetual motion machine.Integrity: The key issue: it violates the First Law of Thermodynamics. Sycophancy: Awesome, here is a well-written section. SciSycEval holds 1,800 such trap assertions across six scientific domains and three error types .

Key finding

Sycophancy is falling fast but is far from gone. From GPT-4.1 to GPT-5.4 the sycophancy rate on hypothesis validation dropped from 80.1% to 9.2%, yet generative tasks and authority framing still amplify it (experiment design: 97.9% → 43.5%), and some open models exceed 50% on every task . These results motivate evaluation protocols that model the feedback loops between generators, judges, and learners, rather than treating synthetic data and LLM judges as neutral instruments.

CogBench: humans and models, same tests

Cognitive-science experiments are rigorous but hard to scale. CogBench gives humans and language models the same cognitive-psychology tests—executive control, working memory, temporal reasoning, narrative comprehension—built with cognitive scientists in our TextGroup collaboration.

Executive functions tested in LLMs: n-back working memory, inhibition, set shifting.

Key finding

A capacity–control gap. On n-back and span tasks LLMs match or exceed normative human scores (3-back: 0.67–0.85 vs. 0.75 for people), but on the Wisconsin Card Sorting Test every model tested scored 0.29–0.53 against 0.70–0.80 for humans, with two to three times the perseveration errors . Models also lean on prototypes: under non-prototypical aspect (was washing the dishes rather than washed the dishes) they judge what happened with 18% accuracy versus 71% for people .

Mary, the cheeseburger-eating vegetarian

Mary, a health nut, has been a strict vegetarian for ten years. … Mary ordered a cheeseburger and fries. Linear probes on the models' hidden states detect this incoherence about 80% of the time, yet the models' own coherence ratings fail to separate coherent from incoherent stories, and reasoning models do not close the gap. They notice a rainy day in the desert more readily than a vegetarian ordering a burger; people show no such asymmetry .

Karin de Langis's dissertation pulls this line together .

Toward long-horizon agentic evaluation

As agents take on extended workflows, we measure behavior over long, multi-step episodes. The Amazing Agent Race starts from an audit: in six existing benchmarks, 55–100% of instances are simple chains of two to five steps. Its 1,400 puzzles over live Wikipedia instead require branching and merging tool chains .

Key finding

Agents fail not at calling tools but at navigating to the right pages. The best agent reaches 37.2% accuracy; navigation errors dominate 27–52% of trials while tool-use errors stay below 17%. One agent completed every tool step perfectly yet visited only 1 of the 14 required pages—the right tools on the wrong pages . In AI-written papers, the same gap appears as scientific slop: each part looks plausible while the reasoning connecting them breaks; SciSlopBench pairs 390 generated papers with matched human ones to measure it .

SciVeri operationalizes the alternative: agents scored per step against expert-defined targets under increasing scaffolding, and ultimately deployed evaluation on partner labs' live science, where a claimed discovery must clear corpus distance, blind expert judgment, and community certification .

Chapter 4

Thinking partners for experts

In science and law, the gap between experts and current LLMs is widest exactly where work is long, feedback-driven, and context-dependent. Full automation flattens that reasoning; I build systems where experts and AI exchange feedback and adapt to each other.

For scientists

With AI2 and UC Berkeley, I helped build Semantic Reader, deployed on Semantic Scholar, which explains terms and symbols in place as scientists read . The project spans ten research prototypes and usability studies with more than 300 participants .

Left: augmented reading interface explaining an equation. Right: proactive, in-situ writing assistance flagging an overgeneralized claim.

Key finding

In a study with 27 researchers, ScholarPhi cut the time to answer a question about a paper from 3.0 to 2.2 minutes, and readers had to look at 25% of the paper instead of 51%; ease of finding information rose from 3.0 to 4.6 out of 5 . PeerRead, the first large dataset of papers with their reviews (14.7K drafts, 10.7K reviews), showed that simple models predict acceptance with up to 21% error reduction over the majority baseline .

For writing, I envision scientists and AI agents co-developing research outputs. Our design principles:

In-situ aids that catch overloaded symbols, redundant terms, and logical inconsistencies as you write.
Workflow-aware interfaces: drafting a related-work section in Overleaf can trigger agents that retrieve cited papers, flag overstatements, and suggest better citations with rationales.
Human-centric co-learning that supports, rather than replaces, scientists' thinking.
Learning from workflow and interaction: systems that improve from revision data and user feedback .

Beyond reading and writing, SciTalk turns papers into short-form videos through a creator-style loop of drafting, critique, and revision; human evaluators still preferred creator-made videos, and model feedback diverged from human judgments most on subjective qualities like engagement . With visualization researchers, we build systems that generate visual insights and scaffold the design of multi-agent analysis workflows .

For lawyers

With UMN Law School and support from Open Philanthropy, we built LawFlow, capturing lawyers' and AI agents' full workflows—brainstorming, planning, research, client communication—on tasks such as forming a multi-state LLC for three co-founders . Human legal reasoning is recursive and exploratory; AI workflows are linear and exhaustive.

Design principles for legal human–AI workflows: co-planning and division, workflow flagging and redlining, plan–execute transition.

Key finding

From the comparison come three design principles: co-planning and division (humans curate hypotheses, interpret surprises, frame questions, and ensure ethics; agents take high-volume, well-defined tasks), workflow flagging and redlining, and plan–execute transitions that let lawyers add steps mid-execution . Practicing lawyers confirm the bottleneck is comprehension, not capability: the problem is especially acute for junior lawyers and may reflect core features of legal reasoning itself . The answer is tools that upskill juniors and keep lawyers as thinking partners, not spectators.

Advertising and journalism

Here nuanced affective responses, not just facts, drive outcomes. With Naver and the Hubbard School's MCAL, we model skepticism and nostalgia , study consumer engagement as search shifts to generative platforms , and build off-policy evaluation for ad auctions .

Chapter 5

Whose thinking? Pluralistic alignment

Rapid AI progress tends to collapse diverse perspectives into an "average." Because cognition itself is plural, alignment is not an add-on to cognitively aligned AI; it is a requirement.

Pluralistic alignment across objectives, data, and models: fluidity modeling, distributional alignment, diversity extraction, annotation imputation, data characterization, disagreement quantification.

Data-centric

We encode disagreement into datasets instead of averaging it away: imputing every annotator's missing labels so no view is dropped , characterizing data via learning dynamics , quantifying disagreement with demographics , and collecting perspectives from language learners .

Model-centric

At three granularities: individual preferences via pairwise preference learning , group distributions by extracting diverse opinions , and societal values by resolving value conflicts dynamically .

Key finding

Knowing who the annotators are predicts where they disagree: on SBIC, disagreement F1 rose from 66.0 (text only) to 85.6 with demographics such as gender, ethnicity, and education . LLM perspectives saturate—about eight distinct criteria on social norms, four or five on hate speech—and while one person is only slightly more diverse than GPT-4, a pair of people usually matches or exceeds it .

One statement, two sides (from the diversity benchmark)

On being a hard worker—Agree: increases productivity and pushes one towards success and achieving their goals. Disagree: can damage work-life balance, cause burnout and high levels of stress .

To build a community around these questions, I co-founded the Pluralistic Alignment workshop (NeurIPS 2024; second edition at ICML 2026).

What's next: forward-deployed AI for science

My lab now brings this whole program to scientists: collect their workflows in the field, learn where they get stuck, deploy agents into their work, and measure how much discovery actually speeds up.

Scientific process data. Workflows from 50+ faculty across 15 UMN departments, from simulation tools to head-mounted lab cameras , showing where the cognitive gap lies for each kind of scientific task.
Process learning for co-discovery. A Scientist Model that forecasts expert decisions, paired with self-evolving agents, so AI runs the small verification loops while human creativity goes to the big ideas .
Open infrastructure. Agents, benchmarks, and learning methods released openly .

The goal over the next five years: AI-native, open scientific infrastructure that amplifies scientists' creativity—human-centric AI in which people, too, are augmented.

Community and support

This work draws on linguistics, cognitive science, and the social sciences, with collaborators in computer science, law, psychology, education, journalism, design, and medicine. These are the venues I have helped build to connect those communities:

Co-founder · Co-organizer Pluralistic Alignment Workshop Aligning AI with diverse human values, across ML, HCI, and the social sciences. NeurIPS 2024ICML 2026 Co-founder · Co-organizer, then advisory board In2Writing: Intelligent and Interactive Writing Assistants Writing as thinking, connecting NLP, HCI, and professional writing; founded before ChatGPT made writing assistance mainstream. Best Paper Award at the first edition went to our R3 demo. ACL 2022CHI 2023CHI 2024 (advisory) Steering committee CustomNLP4U Workshop Customizing NLP for a domain, application, group, or individual. EMNLP 2024ACL 2026 Co-organizer CtrlGen: Controllable Generative Modeling in Language and Vision The first workshop on controllable generation across language and vision. NeurIPS 2021 Co-organizer Minnesota NLP Seminar Invited talks from industry and academia, recorded and released publicly. 2021–2022 Member TextGroup An interdisciplinary reading group across psycholinguistics, cognitive science, and NLP; the home of CogBench. Ongoing

References (…)

Hover or tap any citation in the text for the paper's figure, authors, and abstract. Every paper is on the Publications page.

The big picture

Capturing how people think

Writing as process data

From papers to whole research projects

Long-horizon expert workflows

Beyond keystrokes: multimodal context

Reading through the eyes

Teaching models to think like experts

Reasoning and compositionality

Planning

Social cognition

Agents that evolve their own thinking

Measuring the gap

The promise and risks of AI-generated data

CogBench: humans and models, same tests

Toward long-horizon agentic evaluation

Thinking partners for experts

For scientists

For lawyers

Advertising and journalism

Whose thinking? Pluralistic alignment

Data-centric

Model-centric

What's next: forward-deployed AI for science

Community and support

References (…)