Source-linked AI summary
Measuring Progress Toward AGI: A Cognitive Framework
Ryan Burnell, Yumeya Yamamori, Orhan Firat, Kate Olszewska, Steph Hughes-Fitt, Oran Kelly, Isaac R. Galatzer-Levy, Meredith Ringel Morris, Allan Dafoe, Alison M. Snyder, Noah D. Goodman, Matthew Botvinick, Shane Legg
TL;DR
The paper addresses the lack of a clear, empirical framework for measuring progress toward AGI. It introduces a 10-faculty Cognitive Taxonomy and an evaluation protocol using targeted, held-out tasks and human comparisons to produce cognitive profiles. The framework is a practical starting point for tracking system strengths and weaknesses, while requiring caution about system-level evaluation and the taxonomy’s scope.
Problem
Existing efforts do not cover the full breadth of human cognition or provide robust comparisons with human performance, leaving progress toward AGI difficult to measure.
Method
The paper draws on research in psychology, neuroscience, and cognitive science to define 10 faculties and evaluates systems across targeted, held-out cognitive tasks against human baselines.
Results
Systems can be represented by cognitive profiles that show strengths and weaknesses relative to human performance across the 10 faculties.
Takeaways & Limitations
The framework offers a practical way to move AGI discussion toward empirical measurement and track progress across a jagged landscape of capabilities.
Takeaways & Limitations
Cognitive profiles can be affected by the system harness and tools, which complicate attribution of intelligence and interpretation of targeted tests.
Abstract
from arXiv · showhide
Despite widespread discussion of AGI, there is no clear framework for measuring progress toward it. This ambiguity fuels subjective claims, makes it difficult to track progress, and risks hindering responsible governance. As a starting point to address this gap, we present a framework for understanding system capabilities in relation to human cognitive abilities. Drawing from decades of research in psychology, neuroscience, and cognitive science, we introduce a Cognitive Taxonomy that deconstructs general intelligence into 10 key cognitive faculties. We then propose a rigorous evaluation protocol in which a system's performance is measured across a suite of targeted, held-out cognitive tasks, generating a 'cognitive profile' that can be used to understand a system's strengths and weaknesses. We hope this framework will provide a practical roadmap and an initial step toward more rigorous, empirical evaluation of AGI.
1 Introduction
The paper addresses the lack of clear, empirically grounded ways to define and measure progress toward AGI. It proposes a Cognitive Taxonomy and evaluation framework to compare AI capabilities with human cognition.
- Clearer measurement is intended to support communication about progress and help policymakers develop more effective governance for increasingly capable systems.
- The paper frames human cognitive capabilities as a key reference point for assessing whether systems are both highly capable and highly general.
- Existing AGI frameworks and benchmarks do not fully cover human cognition or provide robust comparisons with human performance.The paper identifies this as the central measurement gap.
- The proposed framework has two parts: a Cognitive Taxonomy of important human cognitive abilities and an evaluation framework spanning that cognitive space.
2 Cognitive Taxonomy
The Cognitive Taxonomy organizes general intelligence into 10 faculties grounded in research on human cognition. It is designed as a practical, mechanism-agnostic framework rather than a definitive account of all intelligence.
- Scope: The framework is explicitly a starting point because artificial systems may lack some human cognitive features and develop capabilities that do not map neatly onto the taxonomy.
- Cognitive faculties: The taxonomy identifies 10 cognitive faculties that research suggests are important for intelligent behavior, with specific abilities and sub-abilities under each faculty.
- Design principle: The taxonomy focuses on what systems accomplish while remaining agnostic about their underlying mechanisms and modeling approaches.
- Cognitive faculties: Eight basic faculties include perception, generation, attention, learning, memory, reasoning, metacognition, and executive functions.
- Composite faculties: Problem solving and social cognition are composite faculties because cognitive abilities interact and operate together in important psychological contexts.
3 Evaluating Cognitive Capabilities
The evaluation protocol measures systems across targeted cognitive tasks, compares them with human baselines, and represents strengths and weaknesses in cognitive profiles. The protocol emphasizes held-out, varied, independently verified tasks and uncertainty-aware comparisons.
- Evaluation protocol: The protocol has three stages: assess systems across cognitive tasks, collect human baselines on the same tasks, and build profiles relative to human performance.
- Task design: Tasks should target specific abilities, use held-out test sets, receive independent verification, vary in human difficulty, and differ in structure and format.
- Human baselines: Human baselines should use the same task conditions and sample broadly from the adult population to support meaningful comparisons with system performance.
- Cognitive profiles: Profiles place each system along human performance distributions by estimating how much of the human sample it outperforms across the 10 faculties.
- Cognitive profiles: Above the human-sample median across all 10 faculties indicates performance matching at least 50% of the sample, while the 99th percentile across all faculties matches almost anyone in that sample.
- Uncertainty and limitations: Interpretation requires uncertainty estimates because task quality, construct validity, and stochasticity can make observed differences between systems or humans difficult to interpret.
4 Discussion
The discussion situates cognitive benchmarking as one component of broader AI evaluation. It highlights gaps involving speed, propensities, creativity, system-level deployment, and the taxonomy’s scope.
- Processing speed: Response speed matters for real-world utility but reflects both cognitive processes and non-cognitive factors such as hardware and network speed.
- System propensities: System propensities concern what systems tend to do, including risk-taking, value alignment, problem-solving strategies, and interaction patterns, but their full evaluation is beyond this paper’s scope.
- Creativity: Creativity is difficult to evaluate objectively because novelty and quality are subjective and domain-specific, so the taxonomy instead captures related processes such as flexibility, knowledge, and problem solving.
- Deployment evaluation: Cognitive benchmarking should complement, rather than replace, applied end-to-end evaluations of important deployment workflows.
- Model versus system evaluation: The authors favor evaluating complete systems because isolating the core model is increasingly impractical and may not represent deployed performance.
- Model versus system evaluation: System-level evaluation creates a trade-off: harnesses and tools affect attributed intelligence while potentially muddying interpretation of targeted cognitive tests.
5 Conclusion
The framework aims to move AGI discussion from subjective claims and speculation toward a grounded, measurable scientific endeavor by mapping AI capabilities and tracking progress toward general intelligence.
- The Cognitive Taxonomy and evaluation protocol provide an empirical framework for mapping AI capabilities and tracking progress toward general intelligence.
7 Appendix: Cognitive taxonomy
The taxonomy organizes perception across visual, auditory, textual, and multisensory abilities, while recognizing that modality importance and human comparability remain unsettled.
- Perception covers the extraction and processing of sensory information across visual, auditory, and textual modalities.
- Visual perception: Visual perception ranges from low-level features such as light, color, contrast, depth, motion, and shape to high-level scene understanding.
- Auditory perception: Auditory perception includes detecting sound features and understanding speech, speakers, sounds, auditory scenes, and spoken language.
- Text perception: Text perception is divided into low-level extraction of letters and words and high-level understanding of language and code.
- Scope: The taxonomy excludes touch, smell, and temperature from detailed treatment because their relative importance for intelligent behavior remains unclear.
- Multisensory integration: Multisensory integration combines complementary modalities to support joint processing, reasoning, and planning, and should be measured for each modality combination.
7.2 Generation
Generation covers the production of text, speech, environmental actions, and internal thoughts, distinguishing execution quality from decisions about what to do.
- Generation includes outputs such as text, speech, motor movements, computer control actions, and tool-use calls needed to communicate and complete tasks.
- Output generation measures execution ability, which can be partly decoupled from reasoning and planning about which output to attempt.
- Text generation: Text generation includes natural-language production with grammatical correctness and lexical selection, as well as structurally and syntactically correct code generation.
- Speech generation: Speech generation includes clarity, grammatical correctness, lexical selection, prosody control, and emotional expression.
- Action generation: Action generation includes manipulating environments through motor control, computer control, and tool use.
- Thought generation: Thought generation produces internal language, images, or abstract representations that can guide decisions, but may be difficult or impossible to evaluate.
7.3 Attention
Attention comprises processes that allocate cognitive resources to relevant information while balancing focused control with responsiveness to unexpected changes. The section also situates attention within broader adaptive cognition, including learning and abstraction.
- Attention focuses cognitive resources on selected perceptual stimuli, information, or thoughts, especially when cognitive resources are limited.
- Effective attention balances narrow focus on current goals with monitoring the wider environment for unexpected changes.
- Attentional control: Selective attention prioritizes goal-relevant information and ignores irrelevant information to support goal-driven behavior.
- Attentional control: Attention includes sustained focus, inhibition of distraction, and shifting between locations or information sources.
- Stimulus-driven attention: Stimulus-driven attention redirects focus toward new stimuli or environmental changes that may require a rapid response.
- The broader taxonomy also treats learning as acquiring knowledge or skills through experience and abstraction as extracting key features to support generalization.
7.5 Memory
Memory is the ability to maintain and retrieve information over time, encompassing general, episodic, sensory, temporal, spatial, procedural, and prospective information. The section distinguishes memory from learning while emphasizing their close relationship and the importance of forgetting for efficient storage and retrieval.
- Memory keeps track of information over time, whereas learning focuses on acquiring new knowledge; evaluating memory includes pre-existing and newly learned information.
- Memory evaluation should remain agnostic about implementation and focus on how well a system retains and retrieves information.
- General memory includes facts and information, while episodic memory covers specific events and associated sensory information.
- The taxonomy separates sensory, temporal, and spatial memory according to sensory details, event sequences, and relationships among objects.
- Memory also includes retaining action patterns, remembering planned actions when cues arise, and removing outdated or irrelevant information.
- Reasoning includes deduction from premises to certain conclusions, induction from observations to probabilistic generalizations, abduction to likely explanations, analogy, and mathematical operations.
7.7 Metacognition
Metacognition concerns a system’s knowledge of and control over its own cognitive processes. The taxonomy divides it into self-knowledge, monitoring, and adjustment of strategies or actions.
- Metacognition combines knowledge about a system’s own cognitive processes with the ability to monitor and control them.
- Metacognitive knowledge: Metacognitive knowledge covers abilities, limitations, learning processes, stored information, and behavioral tendencies.
- Metacognitive monitoring: Metacognitive monitoring evaluates cognitive states such as learning progress and current performance.
- Metacognitive monitoring: Monitoring includes confidence calibration, judgments of learning, error monitoring, and source judgments.
- Metacognitive control: Metacognitive control uses insights from knowledge and monitoring to adjust cognitive processes or strategies.
- Metacognitive control: Control abilities include error correction, learning-strategy selection, goal maintenance, planning, inhibition, task switching, conflict resolution, and working memory.
7.9 Problem solving
Problem solving is a composite ability for overcoming obstacles by representing problems, retrieving relevant knowledge, decomposing goals, planning actions, and executing plans. The taxonomy spans novel-pattern problems, mathematical and logical problems, real-world reasoning, temporal, spatial, causal, and intuitive-physics problems, as well as scientific creativity and social cognition.
- Problem solving relies heavily on planning, reasoning, and in-context learning to overcome obstacles.
- Its process includes representing the problem, retrieving relevant knowledge, breaking it into sub-goals, planning actions, and executing the plan.
- The taxonomy includes identifying patterns to solve novel problems and applying mathematical concepts and techniques.
- Problem-solving domains include logical, real-world, temporal, spatial, causal, and intuitive-physics reasoning.
- Scientific problem solving involves generating novel hypotheses, experiments, and solutions to scientific questions.
- Social cognition includes interpreting social information, reading cues, reasoning about others’ mental states, following norms, cooperating, negotiating, and potentially deceiving or persuading others.