On the use of foundation models in cognitive science

Technical Summary: On the use of foundation models in cognitive science

Problem Statement
Recent studies have demonstrated that Foundation Models (FMs)—including large language models, multi-modal models, and constrained variants like BabyLMs—can reproduce behavioral signatures of human cognition across domains such as mathematical reasoning, language comprehension, and analogical reasoning. However, the authors argue that using FMs as explanatory cognitive models poses significant methodological and conceptual challenges. The core problem is that behavioral alignment (the statistical correspondence between model outputs and human data) does not inherently identify the mechanisms generating those behaviors. Without a structured inferential framework, observed similarities risk being incidental, and treating FMs as cognitive models without rigorous constraints may lead to "behavioral proxies" rather than explanatory models. The paper asks: Under what conditions does behavioral alignment justify treating FMs as explanatory models of cognition?

Methodology: A Four-Stage Inferential Framework
The paper proposes a structured, four-stage framework to evaluate FMs as cognitive and developmental models. This framework separates task adaptation, linking hypotheses, behavioral evaluation, and contrastive model comparison to constrain the interpretation of alignment.

1. Adaptation of Experimental Stimuli and Tasks: Human experimental materials must be translated into input-output modalities compatible with FMs. This ensures the model and human participants perform functionally equivalent tasks. For example, visual Raven's Progressive Matrices are converted into symbolic "digit matrices" to preserve relational structure while accommodating the textual modality of FMs.
2. Specification of Linking Hypotheses: Researchers must explicitly define how model outputs map to measurable aspects of human behavior. This stage is critical as it determines what counts as evidence of alignment. The paper analyzes four common linking hypotheses:
   • Similarity: Mapping the geometry of a model's representational space (e.g., embedding distances) to human judgments of typicality or response times.
   • Surprisal: Using the negative log-probability of tokens to predict human response times or processing difficulty.
   • Prompting: Instructing the model to follow the same protocols as human participants to directly compare generation distributions or choice probabilities.
   • Process-Trace Linking: Mapping properties of intermediate generation traces (e.g., reasoning token counts, step length) to human processing dynamics or cognitive effort.
3. Evaluation of Correspondence: Model performance is compared to human data using statistical goodness-of-fit measures. Crucially, alignment must be assessed not just in aggregate but across theoretically diagnostic variations (e.g., across conditions, individuals, or stimuli) to ensure the model captures specific cognitive mechanisms rather than general trends.
4. Cross-Model Comparison: Demonstrating alignment in a single model provides only "proof of possibility." Explanatory value is derived from contrastive evaluation: comparing different architectures, training regimes, scales, or ablated variants to identify which computational components are necessary to reproduce a specific behavioral signature.

Key Contributions
The paper makes three primary contributions to the field of cognitive science:

• Structural Delineation: It presents a four-stage framework (Figure 1) that separates the inner alignment loop (Stages 1–3: adaptation, linking, evaluation) from the outer contrastive loop (Stage 4: comparison). This structure clarifies how each stage constrains the interpretation of alignment.
• Theoretical Role of Linking Hypotheses: It articulates the interpretive role of linking hypotheses, emphasizing that they are not neutral mappings but theoretical commitments that define how model behavior is interpreted. The paper demonstrates that a strong fit under one linking hypothesis (e.g., prompting) may weaken or vanish under another (e.g., representational similarity), suggesting alignment is a property of the model evaluated under specific assumptions, not the model alone.
• Guidelines for Explanatory Evaluation: It distills the framework into operational principles for designing evaluations that prioritize explanatory value over descriptive similarity. These guidelines emphasize diagnostic measurement contrasts, the preservation of task structure, triangulation across linking assumptions, and comparative evaluation against baselines and manipulated variants.

Results and Analysis of Challenges
The paper does not present new empirical results but synthesizes existing literature to identify four key inferential challenges that arise when using FMs as cognitive models:

1. Theoretical Underdetermination: High aggregate goodness-of-fit can mask failures on critical diagnostic contrasts. Alignment claims are strengthened only when models explain a broad set of theoretically related phenomena and diagnostic contrasts, rather than isolated empirical regularities.
2. Mechanistic Opacity: FMs may solve tasks at the computational level while diverging at the algorithmic level (relying on different representations or search strategies). The paper notes that internal model components (layers, attention heads) do not map cleanly onto cognitive constructs (e.g., working memory), making cross-level inference difficult. Mechanistic interpretability and causal interventions (Stage 4) are proposed as necessary routes to validate whether a model implements specific cognitive constructs.
3. Training and Developmental Mismatch: Differences between FM training (next-token prediction on static corpora) and human learning (interactive, scaffolded, biological) complicate claims of developmental alignment. While intermediate training checkpoints offer opportunities to track developmental trajectories, such correspondences remain correlational unless causal manipulations of training regimes or data ordering are tested.
4. Variability and Population-Level Alignment: FMs are typically trained on aggregate data, often ignoring individual differences. The paper argues that alignment claims risk overstating generality if they ignore structured variation across individuals, groups, or contexts. Variability should be treated as a target for explanation rather than noise.

Significance and Claims
The paper claims that behavioral fit alone is insufficient for establishing FMs as explanatory models of cognition. Alignment becomes scientifically meaningful only when embedded within:

• Explicit theoretical commitments.
• Theory-diagnostic tasks.
• Systematic contrastive evaluation across candidate models.

The authors do not advocate for the uncritical adoption of FMs as cognitive models. Instead, they argue that their scientific value lies in their potential to illuminate which computational principles are sufficient (and perhaps necessary) for reproducing characteristic patterns of human cognition. The proposed framework is offered not as a fixed prescription but as a necessary structure to ensure that claims of alignment are developed with appropriate theoretical and empirical guardrails, ultimately sharpening the understanding of the structure, limits, and development of human thought.