Center for Decoding the Universe Quarterly Forum Recap - Fall 2025

By Steven Dillmann

The Autumn 2025 Quarterly Forum of the Center for Decoding the Universe brought together researchers from across Stanford to explore how modern advances in foundation models and AI agents are reshaping the scientific process, with a particular focus on astrophysics. The Forum consisted of several invited talks, contributed posters, and ample time for discussion. The invited talks are summarized below. 

Session 1: Forward Modeling the Universe

Towards an Astrophysical Foundation Model — Kate Storey-Fisher

Kate Storey-Fisher outlined a vision for astrophysical foundation models that integrate data across wavelengths, instruments, and scales. She highlighted the challenges of working with multi-modal data at vastly different scales from massive datasets like the Rubin Observatory.

Current work aims to predict one modality from another, perform super-resolution enhancement, estimate source properties, identify rare objects, and potentially learn physical laws directly from data. Open challenges include whether to train on raw versus processed data and how much instrument knowledge to embed versus learn from observations.

Foundation Models in Imaging — Sonia Minseo Kim

Sonia Minseo Kim from the Stanford Computational Imaging Lab discussed the mathematical backbone of imaging as an inverse problem — reconstructing physical reality from noisy measurements. She presented a framework for understanding physical image formation and reconstruction through inverse problems and end-to-end optimization, and outlined design decisions around reconstruction algorithms, instrument calibration, noise modeling, and prior selection, with applications spanning from black hole imaging to medical diagnostics.

Foundation Models in Particle Physics — Ben Nachman

Ben Nachman discussed foundation models in particle physics. Moving beyond token prediction approaches, the 500M parameter model OmniLearn was trained on 1 billion jets and generalizes across jet types, detectors, and collision systems for tasks including conditional generation, likelihood ratio estimation, and anomaly detection. He raised critical questions around leveraging simulations directly in training and inference, integrating metadata with data, and transferring knowledge across different experiments.

Session 2: Agentically Discovering the Universe

AI Agents in the Research Workflow — James Zou

James Zou highlighted the evolution from AI as a tool toward AI as a co-scientist as a paradigm shift for science. Drawing on the AI Agents for Science conference, the first open conference where AI serves as both primary authors and reviewers of research papers, he discussed how agents now assist across the scientific lifecycle from hypothesis formation to writing, but remain limited in context understanding and experimental grounding. A study on LLMs as reviewers found varying levels of critical assessment across models, with some too positive, others too conservative, and some striking a more balanced approach. He emphasized the complementary role of human creativity, domain intuition, and experimental validation alongside AI capabilities.

ReplicationBench: Possibilities and Challenges — Christine Ye

Christine Ye introduced ReplicationBench, a benchmark decomposing peer-reviewed astrophysics papers into structured subtasks to evaluate AI performance on real research tasks. Current models achieve around 20% task completion, with failures often due to lack of persistence, procedural errors, and technical execution issues. Her team's work aims to define reproducibility standards and design future benchmarks for rigorous expert-AI collaboration.

Social and Cultural Impact of AI Agents — Angèle Cristin

Angèle Cristin reflected on the societal implications of delegating scientific reasoning to AI systems, emphasizing ethical, creative, and epistemic dimensions of automated discovery.

Towards Automated Data-Driven Discovery in Astrophysics — Owen Queen

Owen Queen presented the Data Discovery Agent, a multi-agent system for open-ended exploration of astrophysical datasets. Designed to mimic long-horizon, trial-and-error scientific reasoning, it integrates hypothesis, execution, and review agents. Early benchmarks show the system outperforming existing models on astrophysics benchmarks, with a domain-expert feedback further improving discovery outcomes.

alphaXiv — Steven Corley

Steven Corley concluded the session with alphaXiv, a community platform that enhances how researchers engage with open-access scientific literature. By enabling discussion, annotation, and exploration of papers across disciplines, alphaXiv aims to make scientific discovery more interactive and collaborative.