New benchmark tests AI scientists for real discoveries, not just coding

Original: TruthInsightBench: An Evidence-Grounded Benchmark for Automated Evaluation of Open-Ended Scientific Discovery Agents

Source: arxiv.org ↗