Rethinking Model Evaluation as Narrowing the Socio-Technical Gap

One-size-fits-all models create a "socio-technical gap": The paper argues that using a single general-purpose LLM across many different applications risks failing to meet the specific human needs of diverse use cases, and current evaluation methods don't adequately assess whether models actually work for real people in real contexts.

Current evaluation practices are disconnected from real-world usage: The authors contend that standard benchmarks and tests don't capture how models perform in actual downstream applications with genuine user needs, drawing on lessons from social sciences and human-computer interaction (HCI) to show that evaluation must be grounded in real-world contexts rather than abstract metrics.

Evaluation requires trade-offs between realism and practicality: The paper acknowledges that perfectly realistic evaluation is expensive and time-consuming, so the research community needs to consciously choose evaluation methods that balance how closely they reflect real conditions against the practical costs of conducting them.

Diverse evaluation methods are necessary: Rather than relying on a single standardized benchmark, the authors advocate for multiple evaluation approaches tailored to different use cases and user populations, recognizing that different applications have different requirements and failure modes.

Interdisciplinary approaches are essential: The paper calls for evaluation methods informed by HCI, social sciences, and explainable AI (XAI) research, suggesting that responsible AI evaluation requires insights beyond traditional computer science to understand human needs and social context.