Toward decision-aware AI for LSST-scale time-domain astronomy

This paper argues that to manage the massive data volume of the Vera C. Rubin Observatory's LSST, time-domain astronomy must shift from static classification to a decision-aware AI framework that integrates foundation models with decision-theoretic policies to dynamically optimize follow-up actions and scientific value within an operational inference loop.

The Gist

The Big Problem: Too Many Fireflies, Too Few Flashlights

Imagine the Vera C. Rubin Observatory (LSST) is a giant camera taking a picture of the entire night sky every few nights. It's so powerful that it will spot about 10 million new "blips" or alerts every single night. These blips are things like exploding stars, black holes eating gas, or distant galaxies flaring up.

In the past, astronomers had a few blips a night. They could look at each one, decide if it was interesting, and grab a telescope to get a closer look (like taking a photo with a flashlight).

But with 10 million blips a night, human astronomers cannot look at them all. If they try to treat every blip as a simple "Yes/No" question (e.g., "Is this a supernova? Yes or No?"), they will miss the most important things. It's like trying to sort a million letters by hand; you'll only read the ones with the clearest stamps and miss the handwritten notes that might contain a life-changing message.

The Old Way vs. The New Way

The Old Way (The Static Classifier):
Currently, computers act like a multiple-choice quiz. They look at a blip and say, "I am 60% sure this is a Type Ia Supernova."

• The Flaw: This doesn't tell you what to do. Even if the computer is 60% sure, that specific blip might be the only chance to catch a rare event before it fades away. The old system treats every blip as an isolated fact, ignoring the fact that we have limited time and resources to investigate them.

The New Way (Decision-Aware AI):
The authors propose a system that acts less like a quiz-taker and more like a strategic game player or a triage nurse.

• Instead of just asking "What is this?", the AI asks: "What is the best thing we can do with our limited resources right now?"
• It understands that some mistakes are worse than others. Missing a rare, fast-fading explosion is a huge loss. Delaying a common, slow-moving star is a small loss. The AI learns to prioritize the "high stakes" situations.

The Three Key Tools

To make this work, the paper suggests combining three specific AI tools:

1. The "Foundation Model" (The Experienced Librarian)
Instead of training a computer to recognize specific types of stars one by one, we train a "Foundation Model" on all the light curves (brightness over time) from history.

• Analogy: Think of this as a librarian who has read every book in the library. When a new, strange book arrives, the librarian doesn't just check a list of titles. They understand the story inside. They can say, "This looks like a mix of a mystery and a sci-fi novel, and it's evolving in a way we haven't seen before."
• This gives the AI a deep "intuition" about what the object is and how it might change, even with very little data.

2. The "Agentic System" (The Smart Manager)
This is the part that makes decisions. It takes the librarian's intuition and asks: "We have 10 million alerts, but only 5 telescopes available for follow-up. Who gets the spotlight?"

• Analogy: Imagine a busy emergency room. The AI is the head nurse. It doesn't just diagnose patients; it decides who gets the operating room right now based on how critical the situation is and what we can learn from treating them. It might say, "Skip the common cold; let's operate on this rare, fast-moving patient because if we wait, we'll lose the chance to save them."

3. The "World Model" (The Simulator)
Before the AI spends a real telescope on a target, it runs a simulation in its head.

• Analogy: It's like a chess player thinking, "If I move my knight here, what will my opponent do next?" The AI simulates: "If we take a spectroscopic picture of this star tonight, what will we learn? If we wait until tomorrow, will the information be lost?" This helps the AI choose the action that gives the most scientific value.

Why This Matters for Science (and People)

The paper argues that this shift changes who gets to do science and what gets discovered.

• The Risk of Automation: If we just let the AI decide based on what it was taught, it might only look for things that fit its training (like common supernovas) and ignore weird, rare things that don't fit the pattern.
• The Human Role: The paper insists that humans must stay in the loop. We need to define the "goals" (e.g., "Find rare black holes" vs. "Study dark energy"). The AI is the tool that executes these goals efficiently, but humans must set the rules.
• Transparency: The AI shouldn't just say "Go look at this." It needs to explain why. "I am suggesting this because it is rare, it is changing fast, and it could help us answer a big question." This allows scientists to check the AI's reasoning and trust it.

The Bottom Line

The LSST telescope will generate a "firehose" of data. We cannot drink from a firehose with a cup (human hands). We need a new kind of AI system that doesn't just classify the water, but decides how to catch the most valuable drops.

By combining deep learning (to understand the data) with decision-making logic (to manage resources), we can turn this massive data stream into a "genuinely intelligent observatory" that not only finds what we are looking for but also notices the strange, unexpected things we didn't even know to ask about.

Technical Summary

Technical Summary: Toward Decision-Aware AI for LSST-Scale Time-Domain Astronomy

Problem Statement
The Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) will generate approximately $10^7$ alerts per night, a scale that renders manual intervention and static classification pipelines infeasible. The paper argues that the prevailing paradigm in time-domain astronomy—treating alerts as independent, static classification problems optimized for metrics like accuracy or purity—fails to address the dynamic nature of the scientific challenge. In the LSST era, scientific return depends not merely on labeling an event correctly, but on making optimal follow-up decisions under uncertainty with finite observational resources. Early light curves are often sparse and ambiguous (e.g., distinguishing between a young Type Ia supernova, a core-collapse event, or a tidal disruption event), and the cost of incorrect actions is asymmetric; missing a rare, rapidly evolving transient is irreversible, whereas delaying a well-understood event may carry little penalty. Current systems do not account for the fact that follow-up decisions reshape the information available for future inference.

Methodology and Proposed Framework
The authors propose a shift from modular, classification-centric pipelines to an integrated, decision-aware AI system embedded within a human-supervised agentic loop. The methodology relies on three core components:

1. Foundation Models as Scientific Priors: Instead of training task-specific classifiers on incomplete labeled datasets, the authors advocate for foundation models trained on heterogeneous, self-supervised time-domain data. These models learn general representations of variability, encoding a source's latent state (evolutionary phase, timescales, color evolution, and uncertainties) rather than a single class probability. This allows the system to represent "what is known" and "what remains uncertain," supporting counterfactual reasoning about what alternative observations would reveal.
2. Agentic Orchestration and Decision Theory: The paper suggests framing alert handling as a Partially Observed Markov Decision Process (POMDP) or utilizing agentic systems that combine probabilistic inference with memory and context-sensitive utility evaluation. These agents evaluate candidate actions (e.g., spectroscopy, multi-band photometry, deferral) against an explicit utility function. The goal is to maximize expected scientific utility (information gain) rather than classification accuracy.
3. Human-in-the-Loop Governance: The system is not designed for unrestricted autonomy. Instead, it operates within bounded agency, where researchers define scientific goals and constraints, and the agentic system proposes interventions. The architecture emphasizes transparency, requiring auditable decision logs, interpretable utility functions, and mechanisms for human override to ensure alignment with community intent.

Key Contributions

• Conceptual Shift: The paper redefines the LSST challenge from a "labeling problem" to a "partially observed dynamical environment" where inference and intervention are coupled.
• Architectural Proposal: It outlines a transition from current broker-centric workflows (Selection $\to$ Classification $\to$ Scheduling) to a decision-aware pipeline where a foundation model maps data to a latent source-state, which an AI agent then uses to optimize resource allocation based on scientific utility.
• Governance Framework: The authors identify the epistemic and institutional risks of centralizing decision-making in AI systems. They propose design principles to prevent the concentration of scientific influence, advocating for open scrutiny of utility functions, training data, and policy objectives to ensure diverse participation in discovery.
• Simulation Pathway: The paper identifies existing infrastructure, such as the OpenUniverse2024 simulation suite and Rubin's survey-strategy framework, as the necessary substrate for benchmarking these decision policies before operational deployment.

Results and Claims
As a perspective paper, this work does not present new experimental results or performance metrics from a deployed system. Instead, its "results" are theoretical and argumentative:

• It demonstrates that static classification metrics (e.g., ROC curves) are insufficient for LSST-scale science because they ignore the asymmetric costs of actions and the sequential nature of information gathering.
• It argues that foundation models can recover meaningful latent structures from photometric data without relying on incomplete labeled sets, providing a more actionable substrate for decision-making than class probabilities.
• It posits that active learning and agentic orchestration are operationally necessary to manage the trade-offs between high-cost interventions (like 8-meter spectroscopy) and lower-cost observations, particularly for early-time, ambiguous events.

Significance
The paper claims that the architectural decisions made in the coming years will fundamentally shape the "epistemic character" of LSST science. By moving toward decision-aware AI, the community can transform the observatory from a "data firehose" into an "intelligent observatory" that learns from its own discoveries and allocates attention where it matters most. The significance lies not just in efficiency, but in preserving the scientific agency of the community: ensuring that the systems built can "notice what we are not" looking for, while maintaining human oversight over the definition of scientific value. The authors emphasize that without transparent governance and human engagement with uncertainty, the scale of LSST could constrain rather than expand the scope of astronomical discovery.