publications

2026

1. AgentBayes: Open-Ended Scientific Model Discovery

   Farhang, A. R., Erickson, A. L., Sehgal, A.,  and Yue, Y.

   preprint 2026

   Abs PDF

   Scientific data modeling often requires more than predicting observations; it must recover the probabilistic data-generating structure: how observations are nested, how variation arises at each level, and how uncertainty propagates through to predictions. We introduce AgentBayes, an agentic system for Bayesian model discovery on real scientific datasets. AgentBayes automates the iterative Bayesian workflow through two roles: an Interactor that explores raw data and fitted posteriors through executable analysis, and a Modeler that converts these findings into hierarchical probabilistic programs. Because data and posterior samples remain in an executable sandbox rather than being serialized into LLM context, AgentBayes can critique and revise models at dataset scales where prior agentic Bayesian systems exhaust the context window. Across five posteriordb benchmarks and three scientific case studies, AgentBayes matches or improves on expert-written Bayesian models on most datasets, scales to larger datasets than prior agentic Bayesian systems, and outperforms symbolic regression baselines on data with experimental hierarchy. AgentBayes also surfaces overlooked data structure in existing benchmarks, and adapts its probabilistic programs accordingly. Together, these results show that LLM agents can aid scientists by performing the full Bayesian workflow, from exploratory data analysis to open-ended posterior predictive checks, on real-world, large-scale scientific data, broadening LLM-based scientific discovery from equation search to full probabilistic modeling.

2025

1. Optimized Semantic Steering via Sparse Autoencoder Adapters

   Farhang, A. R., Erickson, A. L.,  and Yue, Y.

   preprint 2025

   Abs PDF

   We study the problem of steering LLMs, where the goal is to intervene on hidden layer activations in order to improve specific behavioral properties. We are particularly interested in settings where one can optimize over a large candidate set of steering interventions to find valuable, task-related perturbations. We take the approach of using sparse autoencoder adapters coupled with natural language feature descriptions to identify disentangled latent dimensions, after which we select and optimize a subset of the the latent codes that are relevant for the target downstream behavior. A key benefit of our approach is the ability to leverage LLM priors to guide feature selection without manual inspection of relevant features. We empirically demonstrate that our approach can generate steered LLM variants that outperform unsteered LLMs on natural language tasks. Furthermore, the selected steering variants can exhibit cross-lingual transfer, providing task improvements on other languages, unseen during selection or tuning. Our method enables tractable, optimized LLM steering by decomposing the problem into discrete feature selection and continuous optimization. This work demonstrates how tools from mechanistic interpretability can be leveraged to improve model capabilities.

2022

1. 

   The United States COVID-19 Forecast Hub dataset

   Cramer, Estee Y., Huang, Yuxin, Wang, ...,  and US COVID-19 Forecast Hub Consortium,

   Scientific Data 2022

   Abs HTML PDF

   Academic researchers, government agencies, industry groups, and individuals have produced forecasts at an unprecedented scale during the COVID-19 pandemic. To leverage these forecasts, the United States Centers for Disease Control and Prevention (CDC) partnered with an academic research lab at the University of Massachusetts Amherst to create the US COVID-19 Forecast Hub. Launched in April 2020, the Forecast Hub is a dataset with point and probabilistic forecasts of incident cases, incident hospitalizations, incident deaths, and cumulative deaths due to COVID-19 at county, state, and national, levels in the United States. Included forecasts represent a variety of modeling approaches, data sources, and assumptions regarding the spread of COVID-19. The goal of this dataset is to establish a standardized and comparable set of short-term forecasts from modeling teams. These data can be used to develop ensemble models, communicate forecasts to the public, create visualizations, compare models, and inform policies regarding COVID-19 mitigation. These open-source data are available via download from GitHub, through an online API, and through R packages.