SFE Challenge

Advancing scientific discovery through artificial intelligence. Join researchers worldwide in pushing the boundaries of AI applications in scientific research.

Introduction

The SFE Challenge (Scientists’ First Exam Challenge) brings together the global research community to explore innovative frontiers of AI-empowered scientific discovery. This competition focuses on building and evaluating advanced multimodal large language models (MLLMs) capable of deep scientific understanding and cognitive reasoning across multiple scientific domains. By mirroring the complexity of real scientific workflows, the SFE Challenge offers a unique testbed to push the boundaries of automated scientific intelligence.

Participants are invited to develop models that not only excel within familiar scientific domains, but also demonstrate robust generalization to novel tasks and disciplines. With comprehensive resources, official baselines, and a supportive community, the SFE Challenge provides an inclusive platform for researchers and practitioners interested in AI for science to accelerate progress in automated reasoning and facilitate the next wave of breakthroughs in AI-driven scientific discovery.

Machine LearningAI for ScienceScientific CognitionMulti Discipline

Timeline

Warm-up Round Starts

October 20, 2025

Participants can start exploring the competition tracks and datasets

Official Round Starts

December 1, 2025

Official competition begins with full access to datasets and evaluation metrics

Official Round Ends

March 15, 2026

Submission deadline for all competition tracks

Results Announcement

April, 2026

Winners will be announced

Competition Tracks

🔬

SFE Benchmark Track

This track is built on the Scientists’ First Exam (SFE) benchmark, aiming to comprehensively evaluate and push forward the scientific cognitive abilities of multimodal large language models (MLLMs) in realistic scientific workflows. The SFE benchmark covers 66 expert-designed scientific tasks spanning five high-value scientific disciplines, containing 830 rigorously vetted multimodal VQA pairs. It systematically assesses models on three interconnected levels: scientific signal perception, scientific attribute understanding, and scientific comparative reasoning. Rule: Participants are provided with a training dataset that includes 25 out of the total 66 tasks in SFE, as well as a never-before-seen test set covering these same 25 tasks. For evaluation, the same LLM-as-a-Judge scoring method used in SFE will be adopted. We invite researchers and practitioners interested in multimodal models, scientific AI, and automated scientific reasoning to join us in advancing the frontier of AI-empowered scientific discovery!