Abstract
In contemporary Natural Language Processing (NLP), evaluation is frequently treated as a neutral technical instrument—a lens through which we measure model performance. This paper argues that this perspective is insufficient. Evaluation is better understood as a sociopolitical "force" that exerts power over the research community by defining progress and coordinating collective effort. By shifting from an "evaluation-as-measurement" paradigm to an "evaluation-as-change" framework, the community can intentionally leverage benchmarks to promote pluralistic values—such as fairness, robustness, and efficiency—against the growing dominance of resource-heavy scaling.
Introduction: Beyond the Technical Lens
Traditionally, the NLP community views evaluation through a technical account: it is a tool used to derive scientific insight or document the state of technology. However, as argued by Rishi Bommasani, this view ignores the sociological reality of how research directions are set. Evaluation is not merely a measurement; it is a mechanism of coordination. A successful benchmark does more than provide a metric; it redistributes the community’s attention, labor, and capital.
Currently, the field faces a "waning power" of evaluation. Research trajectories are increasingly governed by the "force of resources"—the sheer availability of compute and funding—rather than scientific or social objectives. When scaling laws become the primary driver of research, the community converges on narrow metrics like perplexity, often sidelining broader values. This paper proposes a power-centric analysis to understand how evaluation can be reclaimed as a force for intentional change.
The Dual Accounts of Evaluation
To understand the role of benchmarks, we must distinguish between two conceptual frameworks:
- The Technical Account: Evaluation serves as a lens. Its success is measured by its psychometric validity, the reliability of its signals, and the depth of scientific insight it provides.
- The Political Account: Evaluation serves as a force. Under this view, an evaluation is successful only if it achieves a desired shift in the community's research trajectory. It is a tool for intervention, intended to steer the field toward specific goals.
By viewing evaluation as a force, we can analyze how benchmarks accrue authority through adoption, how they compete with the influence of resource-rich institutions, and whether the power they wield is legitimate.
The Dynamics of Power and Adoption
An evaluation’s power is directly proportional to its adoption. A technically superior benchmark that fails to gain traction among researchers has failed as a "force." Historically, standards like ImageNet, SQuAD, and WMT gained immense authority, creating "inertia." Once a benchmark becomes a standard, it is difficult to displace, even when its flaws become apparent. This leads to "value imposition," where the priorities of a few benchmark creators dictate the research agenda for the many.
However, the legitimacy of the "evaluation force" is currently threatened by its own rising costs. For example, a full evaluation using the Holistic Evaluation of Language Models (HELM) framework can cost approximately $40,000 USD and require 20,000 GPU hours. As evaluation itself becomes a resource-heavy endeavor, the barrier to entry for creating and running benchmarks rises, potentially centralizing the power to define progress within a few wealthy institutions.
Empirical Observations: Fragmentation and Scaling
The current state of Large Language Model (LLM) evaluation reveals a significant lack of coordination. A survey of 32 major LLM papers found that among the scenarios introduced in HELM, models were evaluated on average on only 17.9% of them. Furthermore, Recognizing Textual Entailment (RTE) was the only dataset used by more than 50% of the surveyed papers. This fragmentation suggests that the "force of evaluation" is currently fractured, unable to provide a unified counter-weight to the "force of resources."
While scaling laws offer predictability—relating compute to performance on specific tasks—they typically center on a single objective, such as perplexity. In contrast, holistic evaluations like HELM and BIG-bench are designed to promote pluralism. They highlight a diverse array of values, including fairness, uncertainty, and robustness, which scaling laws often ignore.
Conclusion: Harnessing Evaluation for Change
The rise of few-shot evaluation and prompting offers a glimmer of optimism; by reducing the need for massive, expensive training sets, it may lower the barrier for benchmark creators to introduce new tasks. To counter the narrowing effects of resource-driven research, the community must more aggressively harness evaluation as a tool for change. This involves building benchmarks specifically designed to drive the field toward the social and scientific outcomes we wish to see, while utilizing decentralized tools—such as open-source libraries and shared compute—to democratize the power of evaluation.
Summary updated