AI and the Everything in the Whole Wide World Benchmark

By Inioluwa Deborah Raji, Emily M. Bender, Amandalynne Paullada, Emily L. Denton, A. Hanna. 2021

502 citations recorded by the source corpus.

There is a tendency across different subfields in AI to valorize a small collection of influential benchmarks. These benchmarks operate as stand-ins for a range of anointed common problems that are frequently framed as foundational milestones on the path towards flexible and generalizable AI systems. State-of-the-art performance on these benchmarks is widely understood as indicative of progress towards these long-term goals. In this position paper, we explore the limits of such benchmarks in order to reveal the construct validity issues in their framing as the functionally"general"broad measures of progress they are set up to be.

Subfields: Artificial Intelligence, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Evaluation & Benchmarks

Citations in the corpus

PaLM: Scaling Language Modeling with Pathways (2022, background): “As Raji et al. (2021) point out, benchmark evaluations are often deficient in construct validity, where they may not capture the full complexity of language understanding capabilities and gaps, and may not live up to their claims of general applicability.”
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model (2022, neutral): “We therefore expect a distinction between our findings using CrowS-Pairs and wider model use (for a more detailed exploration on such differences, see Raji et al., 2021).”
Evaluating the Social Impact of Generative AI Systems in Systems and Society (2023, background): “There is no universal evaluation by which to evaluate a model’s performance, and any evaluation metrics should be used with deployment context in mind [210, 191].”
The Capacity for Moral Self-Correction in Large Language Models (2023, background): “4) we drew inspiration from the study of fairness in real-world automated (8)See [45] for a compelling criticism on the use of benchmarks in machine learning in general.”
What Do NLP Researchers Believe? Results of the NLP Community Metasurvey (2022, methodology): “We choose research agendas to ask about based on criticisms, debates, or findings in the literature and public sphere, for example regarding current practice in benchmarking (Bowman and Dahl, 2021; Raji et al., 2021), the relative value of advances in model architectures (Narang et al.”
PaLM 2 Technical Report (2023, background): “These measures are a limited approximation, and are by no means exhaustive or holistically complete (Raji et al., 2021; Selbst et al., 2019).”
Data Governance in the Age of Large-Scale Data-Driven Language Technology (2022, methodology): “For LLMs in particular, the data used to train them are one further step removed from the task-specific models built from them, so the link between data and ML progress is even more abstracted [87, 116].”
You reap what you sow: On the Challenges of Bias Evaluation Under Multilingual Settings (2022, background): “Machine learning (ML) systems, especially large lan-021 guage models (LLMs), are prone to (re)produce harmful outcomes and social biases (Bender et al., 2021; Raji et al., 2021; Blodgett et al., 2020; Aguera y Arcas et al., 2018).”
Taxonomy of Risks posed by Language Models (2022, background): “ing the methodological toolkit for LM analysis beyond benchmarks [151, 152, 180].”
SafetyKit: First Aid for Measuring Safety in Open-domain Conversational Systems (2022, background): “In order to discourage hill-climbing on a benchmark and the negative effects which can stem from it (Raji et al., 2021), we do not provide an aggregate score, but discuss possible uses of the tools under investigation and their advantage and disadvantages.”
GEMv2: Multilingual NLG Benchmarking in a Single Line of Code (2022, background): “metrics (Raji et al., 2021), and many evaluation advancements are thus put aside.”
Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (2024, unknown): “Their knowledge of the labeling process came only from our written documentation of labeling criteria, fully described in Appendix C.”
Underspecification in Scene Description-to-Depiction Tasks (2022, background): “Underspeciﬁcation in task formulation is a major challenge for machine learning and artiﬁcial intelligence disciplines as a whole (D’Amour et al., 2022; Raji et al., 2021).”
Harms from Increasingly Agentic Algorithmic Systems (2023, background): “Raji et al. [2021] discuss several issues with benchmarking, including construct invalidity and limitations in scope.”
Can Large Language Models Transform Computational Social Science? (2023, background): “As zero-shot models approach or outperform the quality of the gold-reference generations, reference-based scoring becomes an invalid construct for measuring models’ true utility (Raji et al., 2021), even if we assume the semantic similarity metrics are ideal.”
The Fallacy of AI Functionality (2022, background): “[149] argue that AI makes claims to generality while modeling behaviour that is determined by highly constrained and context-specific data.”