SFBench: The SciFy Scientific Feasibility Benchmark

TL;DR

SFBench is a benchmark dataset with 197 material science claims for evaluating scientific claim feasibility.

cs.AI 🔴 Advanced 2026-06-29 2 views
Cash Costello James Mayfield Elsbeth Turcan Christine Piatko Christina K. Pikas Justin Rokisky Sam Scheck Chris Ribaudo Ritwik Bose Alex Memory
scientific feasibility materials science benchmarking GPT models DARPA SciFy

Key Findings

Methodology

SFBench evaluates systems' ability to judge scientific claim feasibility using 197 expert-created material science claims, each with a -2 to 2 feasibility score and explanation, emphasizing complex reasoning and open-ended responses.

Key Results

  • Baseline tests with GPT models show scientific claim assessment is challenging, with limited performance in feasibility scoring.
  • Automatic scoring of explanations remains an open issue due to multiple valid explanations.
  • Experiments reveal existing models struggle with complex semantic comparisons in technical areas.

Significance

SFBench provides a new benchmark for scientific feasibility assessment, particularly in materials science, aiding researchers in developing systems that can automate or partially automate the evaluation process. It fills a gap in existing datasets for complex reasoning tasks.

Technical Contribution

SFBench introduces an original dataset created by domain experts, avoiding biases from existing scientific publications, and offers an open-ended explanation framework challenging current AI reasoning capabilities.

Novelty

SFBench is the first to provide an expert-created scientific claim feasibility assessment dataset, emphasizing open-ended explanations and complex reasoning, contrasting with existing multiple-choice or fixed-answer benchmarks.

Limitations

  • Automatic scoring of explanations is challenging due to multiple valid explanations, and models struggle with complex semantic comparisons in technical areas.
  • The dataset is relatively small and limited to materials science.

Future Work

Future work includes expanding the dataset's size and domain, improving model performance on complex reasoning tasks, and developing more effective automatic scoring methods.

AI Executive Summary

SFBench is a new benchmark dataset for evaluating systems' ability to judge the feasibility of scientific claims. Existing solutions often rely on claims extracted from scientific publications, which can lead to biases. SFBench introduces 197 material science claims created by domain experts, avoiding such biases and providing an open-ended explanation framework that challenges current AI reasoning capabilities.

Baseline tests with GPT models show that scientific claim assessment is a challenging task, especially in handling complex semantic comparisons. Models show limited performance in feasibility scoring, and automatic scoring of explanations remains an open issue due to the possibility of multiple valid explanations.

SFBench offers a new benchmark for scientific feasibility assessment, particularly in materials science, aiding researchers in developing systems that can automate or partially automate the evaluation process. Future work includes expanding the dataset's size and domain, improving model performance on complex reasoning tasks, and developing more effective automatic scoring methods.

Deep Analysis

Background

Scientific feasibility assessment is a crucial research area, especially in materials science. Existing solutions often rely on claims extracted from scientific publications, which can lead to biases. SFBench introduces an original dataset created by domain experts, avoiding such biases.

Core Problem

The core problem is accurately assessing the feasibility of scientific claims, particularly in the complex field of materials science. Existing methods show limited performance in handling complex reasoning tasks and providing open-ended explanations.

Innovation

SFBench's core innovation lies in introducing 197 material science claims created by domain experts, emphasizing open-ended explanations and complex reasoning. This contrasts with existing multiple-choice or fixed-answer benchmarks.

Methodology

  • �� 197 material science claims created by experts
  • �� Each claim has a -2 to 2 feasibility score
  • �� Provides open-ended explanations, emphasizing complex reasoning
  • �� Baseline tests using GPT models

Experiments

Experimental design includes baseline tests using GPT models on the SFBench dataset, evaluating model performance in feasibility scoring and explanations. Results show models struggle with complex semantic comparisons.

Results

Results indicate that baseline tests with GPT models show scientific claim assessment is a challenging task, with limited performance in feasibility scoring and automatic scoring of explanations remaining an open issue.

Applications

SFBench can be used to develop systems that automate or partially automate scientific feasibility assessments, particularly in materials science.

Limitations & Outlook

The dataset is relatively small and limited to materials science. Automatic scoring of explanations is challenging due to the possibility of multiple valid explanations.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. You have a recipe but aren't sure if it's feasible. SFBench is like a chef advisor, telling you if the recipe can succeed and explaining why. It not only tells you the outcome but details the feasibility of each step.

ELI14 Explained like you're 14

Imagine you're playing a game where you need to judge if a task can be completed. SFBench is like your game guide, telling you if the task is feasible and explaining why. It doesn't just give simple answers but details the likelihood of each step.

Glossary

SFBench

A benchmark dataset for evaluating the feasibility of scientific claims, containing 197 material science claims.

Used to assess AI systems' judgment of scientific claims.

Feasibility Score

A score from -2 to 2 used to evaluate the feasibility of scientific claims.

Each claim has a feasibility score and explanation.

Open-ended Explanation

A detailed explanation of scientific claim feasibility, not limited to fixed answers.

SFBench emphasizes open-ended explanations, challenging current AI reasoning capabilities.

Materials Science

The scientific field studying the properties, structure, and performance of materials.

SFBench's dataset focuses on materials science.

GPT Model

A large language model used for natural language processing tasks.

Used for SFBench's baseline tests, evaluating model performance in feasibility scoring.

Open Questions Unanswered questions from this research

  • 1 How to automatically score open-ended explanations remains a challenge due to multiple valid explanations.
  • 2 Existing models struggle with complex semantic comparisons in technical areas.

Applications

Immediate Applications

Scientific Research Evaluation

Researchers can use SFBench to evaluate the feasibility of scientific claims, aiding in research prioritization.

Long-term Vision

Automated Scientific Assessment

Develop systems that automate or partially automate scientific feasibility assessments, particularly in materials science.

Abstract

We present SFBench, a benchmark dataset for evaluating systems that assess the feasibility of scientific claims. SFBench includes 197 claims in materials science, each annotated with a ground-truth feasibility score on a five-point scale along with an explanation of that assessment. The collection differs from previous collections in several important ways: 1) it defines a complex task that requires reasoning over claims of varying scientific feasibility; 2) its claims are not extracted from existing scientific publications but are created de novo, greatly reducing the chances that LLMs have trained on them; 3) claims and ground truth are established by subject matter experts, not by artificial intelligence; and 4) unlike many benchmarks that ask about question/answer pairs, provide multiple choice answers, or ask questions requiring short, fixed answers, SFBench explanations are completely open-ended. We describe the benchmark design, data creation process, and evaluation metrics, and we report baseline results using recent GPT models.

cs.AI