Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Measure of a Model: From Intelligence to Generality

Ruchira Dhar, Ninell Oldenburg, Anders Soegaard

TL;DR

This work questions the use of intelligence as the central lens for evaluating AI systems, arguing that intelligence is ill-defined and benchmarks often fail to predict real-world utility. It formalizes three notions—generality, stability, and realism—and shows that only generality yields a robust, transferable evaluation framework, supported by multitask learning theory. The authors introduce the Generality Score (G-Score) to operationalize breadth and consistency across a diverse task set, and demonstrate how evaluating across many tasks reduces generalization error by a factor roughly proportional to $\sqrt{n}$, where $n$ is the number of tasks. The result is a practical reorientation of model assessment toward broad, reliable performance in open-ended environments, with implications for benchmarking, deployment, and future research on task distributions and evaluation design.

Abstract

Benchmarks such as ARC, Raven-inspired tests, and the Blackbird Task are widely used to evaluate the intelligence of large language models (LLMs). Yet, the concept of intelligence remains elusive- lacking a stable definition and failing to predict performance on practical tasks such as question answering, summarization, or coding. Optimizing for such benchmarks risks misaligning evaluation with real-world utility. Our perspective is that evaluation should be grounded in generality rather than abstract notions of intelligence. We identify three assumptions that often underpin intelligence-focused evaluation: generality, stability, and realism. Through conceptual and formal analysis, we show that only generality withstands conceptual and empirical scrutiny. Intelligence is not what enables generality; generality is best understood as a multitask learning problem that directly links evaluation to measurable performance breadth and reliability. This perspective reframes how progress in AI should be assessed and proposes generality as a more stable foundation for evaluating capability across diverse and evolving tasks.

On the Measure of a Model: From Intelligence to Generality

TL;DR

This work questions the use of intelligence as the central lens for evaluating AI systems, arguing that intelligence is ill-defined and benchmarks often fail to predict real-world utility. It formalizes three notions—generality, stability, and realism—and shows that only generality yields a robust, transferable evaluation framework, supported by multitask learning theory. The authors introduce the Generality Score (G-Score) to operationalize breadth and consistency across a diverse task set, and demonstrate how evaluating across many tasks reduces generalization error by a factor roughly proportional to , where is the number of tasks. The result is a practical reorientation of model assessment toward broad, reliable performance in open-ended environments, with implications for benchmarking, deployment, and future research on task distributions and evaluation design.

Abstract

Benchmarks such as ARC, Raven-inspired tests, and the Blackbird Task are widely used to evaluate the intelligence of large language models (LLMs). Yet, the concept of intelligence remains elusive- lacking a stable definition and failing to predict performance on practical tasks such as question answering, summarization, or coding. Optimizing for such benchmarks risks misaligning evaluation with real-world utility. Our perspective is that evaluation should be grounded in generality rather than abstract notions of intelligence. We identify three assumptions that often underpin intelligence-focused evaluation: generality, stability, and realism. Through conceptual and formal analysis, we show that only generality withstands conceptual and empirical scrutiny. Intelligence is not what enables generality; generality is best understood as a multitask learning problem that directly links evaluation to measurable performance breadth and reliability. This perspective reframes how progress in AI should be assessed and proposes generality as a more stable foundation for evaluating capability across diverse and evolving tasks.

Paper Structure

This paper contains 21 sections, 2 theorems, 23 equations, 2 figures.

Table of Contents

  1. Introduction
  2. Why Intelligence is Problematic
  3. Conceptual Instability
  4. Illusion of Competence
  5. Rethinking our Ground
  6. Unpacking Intelligence
  7. Generality
  8. Stability
  9. Realism
  10. Generality is Independent
  11. Formal setup.
  12. (Only) Generality is Necessary
  13. From Intelligence to Generality: What it Offers
  14. Conceptual Stability
  15. Theoretical Grounding
  16. ...and 6 more sections

Key Result

Theorem 1

Consider an environment $\mathcal{E}$ consisting of a distribution $Q$ over tasks, where each task $P \sim Q$ is a distribution over data in a learning problem. Let $\mathcal{H}$ be a hypothesis class, $L_P(h)$ be the loss on task $P$, and $L_Q(h)$ be the model's environment average error. Then, for

Figures (2)

  • Figure 1: Comparing performance of LLMs on AGI benchmarks ARC-AGI-1, ARC-AGI-2 and preference-based benchmark like LMArena.
  • Figure 2: The performance of LLMs on task-specific benchmarks OpenBookQA, Entity Extraction, and StackUnseen.

Theorems & Definitions (7)

  • Definition 3.1: Generality
  • Definition 3.2: Stability
  • Definition 3.3: Realism
  • Theorem 1
  • proof
  • Theorem 2
  • proof