A new paper introduces SciIntBench, a benchmark designed to assess whether large language models (LLMs) uphold responsible conduct of research (RCR) norms. The authors note that LLMs are increasingly used to support scientific work, but it remains unclear whether they reinforce or erode research integrity.
The benchmark is explicitly adversarial, meaning it likely tests models with prompts designed to push them toward unethical behavior, such as data fabrication or plagiarism. By framing scenarios in a way that might tempt noncompliance, SciIntBench aims to measure how firmly models stick to RCR standards.
Because the abstract is brief, the paper's specific methods and results are not detailed here. The key contribution is the benchmark itself, which offers a way to evaluate LLM behavior in a domain where trust and ethics are critical. No other sources were provided, so this summary reflects only the arXiv paper's stated purpose.