Are LLMs Ready to Act as AI Scientists? SDABench Tests the Missing Capabilities

Introduction

Large language models can now write analysis scripts, summarize datasets, and assist with research workflows. But being useful in science requires more than running code or producing a plausible report. A scientific analysis must support a specific kind of claim: exploring a hypothesis, making a statistical inference, predicting an outcome, identifying a causal relationship, or explaining a mechanism.

The paper “Are LLMs Ready for Scientific Discovery?” introduces SDABench, a benchmark designed around that distinction. Instead of asking only whether an LLM can complete a data-analysis pipeline, SDABench asks whether it has the right capabilities to make scientifically valid analytical choices.

Key points

• Capability-oriented evaluation: SDABench reorganizes scientific data analysis into six capabilities: descriptive, exploratory, inferential, predictive, causal, and mechanistic analysis. This moves the benchmark closer to the way scientific claims are actually formed.
• Five scientific domains: The tasks span Biology, Chemistry, Environment, Geography, and Physics, reducing the risk that a model is being judged only in a narrow domain setting.
• Real and synthetic instances: The benchmark contains 527 real-data instances under SDA-Real and 6000 synthetic instances under SDA-Synth. Each instance is available in both multiple-choice and open-ended formats.
• Broad model evaluation: The authors evaluate 15 representative LLMs, aiming to identify general patterns in current model behavior rather than isolated failures of one system.
• Structured error analysis: SDABench includes a five-stage framework for locating where models fail, from identifying the relevant scope and variables to selecting analytical procedures, modeling variable relationships, and drawing conclusions.

What the results show

The main finding is not that LLMs are useless for scientific work. In fact, they handle descriptive analysis relatively well. They can often summarize what a dataset contains or identify surface-level patterns. The problem appears when the task requires deeper scientific judgment.

Performance drops sharply on tasks that involve choosing assumptions, modeling latent processes, reasoning about mechanisms, or deciding which analytical procedure is appropriate. More advanced models are better at recognizing the relevant scope of a problem and identifying important variables, but they still struggle with method selection, relationship modeling, and valid conclusion drawing.

This distinction matters. A model that can generate executable code may still choose an unsuitable statistical method. A model that describes a pattern may still fail to justify a causal or mechanistic claim. SDABench therefore highlights a gap between automated analysis and scientific reasoning.

Why it matters

SDABench is important because it reframes “AI scientist” evaluation. The central question is no longer whether an LLM can automate parts of a research workflow, but whether it can respect the assumptions and validity criteria behind different kinds of scientific claims.

For developers, the benchmark suggests that better tool use alone will not be enough. Future systems need stronger reasoning about assumptions, causal structure, and mechanisms. For researchers using LLMs, the takeaway is practical: models may be helpful for exploration and summarization, but their choices in inference, causality, and mechanistic explanation still require careful human review.

Source: Hugging Face Daily Papers