CLINB: A Climate Intelligence Benchmark for Foundational Models

Frontier LLMs show strong knowledge synthesis but critical grounding failures: While advanced models demonstrate PhD-level understanding of climate topics and outperform hybrid expert-assisted answers, they frequently hallucinate references and images, undermining the reliability needed for scientific applications.

Verifiable attribution is essential for responsible AI deployment: The paper identifies a "critical dichotomy" where impressive knowledge quality is undermined by poor evidence grounding, highlighting that trustworthy AI systems must prioritize not just what models know, but whether they can prove it with accurate citations.

Domain-specific benchmarks with expert curation are necessary for accountability: CLINB's reliance on real user questions and evaluation rubrics from leading climate scientists demonstrates that responsible AI evaluation requires specialized expertise rather than generic benchmarks, enabling more meaningful assessment of model reliability in high-stakes domains.

Hallucination of references poses real risks in scientific contexts: The substantial hallucination rates for citations and images documented in the benchmark reveal a concrete failure mode that could mislead researchers and policymakers if models are deployed without proper safeguards, underscoring the need for interpretable evaluation before real-world use.