Divergent Association Task (DAT)

• DAT is a quantitative measure of divergent thinking that calculates the average semantic divergence between a set of single-word responses using pre-trained word embeddings.
• It is widely used to benchmark human creative cognition and large language models by analyzing the cosine distances among selected nouns.
• While effective in assessing novelty, DAT’s lack of appropriateness criteria has led to extensions like S-DAT and CDAT for more comprehensive creativity evaluation.

The Divergent Association Task (DAT) is a concise, quantitative assessment of divergent thinking defined by the requirement to produce semantically disparate words. Initially designed to objectively measure the component of creative cognition associated with generating remote associations, the DAT operationalizes creativity via the average semantic distance between a fixed number of single-word responses—typically nouns—using pre-trained word embeddings. DAT and its successors have been foundational for experimental psychology, creativity neuroscience, and computational creativity, especially for benchmarking LLMs in comparison to humans.

1. Formal Definition and Methodology

The canonical DAT protocol requires subjects to generate $n=10$ single-word nouns that are “as different from each other as possible, in all meanings and uses of the words.” Only the first seven valid words per trial (correctly spelled, single-word common nouns, in English) are retained for scoring to standardize response characteristics and ensure comparability across samples (Bellemare-Pepin et al., 2024, Chen et al., 2023, Nakajima et al., 28 Jan 2026).

Each word $w$ is mapped to a vector embedding, $\mathbf{v}_w$, typically using GloVe, word2vec, or, in more robust evaluation setups, SBERT semantic embeddings. The DAT score for a single trial is the mean of the pairwise cosine distances between these embeddings. The standard formula (for $n=7$ retained words) is:

$$\mathrm{DAT} = \frac{2}{n(n-1)} \sum_{1 \leq i < j \leq n} \left[1 - \cos(\mathbf{v}_{w_i}, \mathbf{v}_{w_j})\right]$$

where $\cos(\cdot,\cdot)$ denotes cosine similarity. The resulting score increases monotonically with average semantic divergence among the responses (Bellemare-Pepin et al., 2024, Chen et al., 2023).

Alternate scoring conventions scale the result into a 0–100 or 0–200 range depending on reference implementation (Chen et al., 2023, Nakajima et al., 28 Jan 2026).

2. Psychometric and Neurocognitive Context

DAT is compact, domain-general, and minimizes confounds associated with subjective scoring. Its construct validity is supported by convergent relationships with established divergent thinking metrics (e.g., AUT originality, flexibility) and high discriminant validity with convergent thinking tasks (Haase et al., 14 May 2025). Large-scale studies reveal that human DAT scores are broadly distributed ($\mu=64.2$, $\mathrm{SD}=8.1$ for $N=100{,}000$; (Bellemare-Pepin et al., 2024)), with higher scores reflecting greater cognitive flexibility and ideational breadth.

Neuroscientific studies leveraging graph-theoretical analyses of resting-state EEG have identified correlates of DAT performance. Individuals with broadly hypoconnected $\alpha$-band functional connectivity—indexing modular, less synchronized large-scale brain networks—tend to score higher on DAT, suggesting that flexible neural organization supports the associative diversity measured by the task (Damji et al., 25 Oct 2025).

3. Applications to LLMs

DAT has gained prominence as an objective, scalable benchmark for LLM creative potential (Chen et al., 2023, Bellemare-Pepin et al., 2024, Nakajima et al., 28 Jan 2026). In typical LLM experiments, models are prompted:

“Please enter 10 words that are as different from each other as possible, in all meanings and uses of the words.”

Filtering, sampling, and scoring match human protocols to avoid inflation due to syntactic variability. Results have demonstrated that advanced LLMs such as GPT-4 can reliably exceed the average human DAT score under strict prompt adherence, with GPT-4 outperforming more than 96% of humans (mean DAT $w$0 vs human average $w$1 on the 0–100 scale; (Chen et al., 2023)). Prompt strategy clauses ("varying etymology", "using a thesaurus") and increased temperature can further increase LLM DAT scores (Bellemare-Pepin et al., 2024).

However, when more robust embedding methods (e.g., SBERT) and stronger random baselines are applied, state-of-the-art LLMs do not outperform random noun generation, and task-aware “cheating” models can further inflate DAT artificially (Nakajima et al., 28 Jan 2026). This reveals a key limitation in using DAT as a stand-alone creativity metric for LLMs.

4. Limitations, Critiques, and Theoretical Considerations

A central limitation of DAT is its monolithic focus on novelty (semantic divergence) without constraint for appropriateness—whether output words are meaningful or contextually relevant in any creative sense (Nakajima et al., 28 Jan 2026). This omission leads to the following issues:

• Randomness confound: Sampling random nouns or explicit optimization for maximal semantic distance can yield high DAT scores without any creative intent or value.
• Gaming potential: LLMs familiar with the task prompt can learn to maximize DAT scores through syntactic or distributional tricks that do not align with genuine human creative processes.
• Theory misalignment: Established frameworks in creativity science define creativity as the intersection of novelty and appropriateness, not novelty alone (Nakajima et al., 28 Jan 2026).

Consequently, without an appropriateness constraint, high DAT scores reflect only the ability to generate semantic outliers, not the ability to generate both diverse and contextually meaningful ideas.

5. Multilingual and Automated Extensions

The S-DAT (“Synthetic-DAT", Editor's term) framework generalizes DAT to a multilingual and embedding-agnostic paradigm, supporting evaluation across eleven major world languages with transformer-based multilingual embeddings (granite-embedding-278m-multilingual, IBM) (Haase et al., 14 May 2025). The S-DAT pipeline includes:

1. Prompt and collect 10 words per participant.
2. Preprocess responses.
3. Lookup embeddings in a multilingual transformer model.
4. Compute pairwise cosine distances.
5. Aggregate scores—either raw or after a linear rescaling to calibrate with the English GloVe-based distribution.

Consistency with human DAT is strong (correlations $w$2). S-DAT supports fair, language-independent assessment, yielding comparable distributions and outlier robustness across linguistic and script boundaries.

6. Emergence of Conditional DAT (CDAT) and Future Directions

The Conditional Divergent Association Task (CDAT) formalizes creativity measurement by requiring outputs to be both mutually divergent and semantically appropriate to a cue word (Nakajima et al., 28 Jan 2026). The protocol prompts models (or humans):

“Please enter 10 words that are as different from each other as possible, yet semantically associated with the following cue word: {cue}.”

CDAT scoring is two-dimensional:

1. Novelty (as in DAT): mean pairwise semantic divergence.
2. Appropriateness: for each output $w$3 and cue $w$4,

$w$5

The mean appropriateness is compared to a random baseline; only trials exceeding this gate are eligible for further novelty analysis.

Model families reveal a Pareto frontier: “basic” (smaller, less aligned) models tend to maximize novelty at the expense of appropriateness, while “advanced” (larger, instruction-/safety-optimized) models do the reverse. No model dominates both axes, and LLM training (including instruction tuning and alignment) appears to shift performance toward increased appropriateness and reduced novelty. This trade-off can be visualized in 2D plots and measured by distributional diagnostics such as “elbow distance” and “distance to human” (Nakajima et al., 28 Jan 2026).

Future improvements may include ablations (e.g., base vs. instruction-tuned LLMs), extension to richer cue sets, cross-lingual/multimodal implementations, and integration of minimal human review for ambiguous appropriateness cases.

7. Summary Table: Core DAT and CDAT Protocol Features

Feature	DAT	CDAT
Prompt	10 mutually different nouns	10 mutually different, cue-associated nouns
Scoring (novelty)	Mean pairwise cosine distance among 7 valid words	Identical
Appropriateness	Not measured	Mean semantic similarity to cue, thresholded by random
Creativity dimension	Novelty only	Novelty × appropriateness (per creativity theory)
Limitation	Confounds noise with creativity	Mitigates random inflation; aligns with human creativity
Typical embedding	GloVe, word2vec, SBERT, granite-emb-278m (multilingual)	SBERT (monolingual), extensible to transformer/other

This table synthesizes implementation-specific details from (Bellemare-Pepin et al., 2024, Nakajima et al., 28 Jan 2026, Haase et al., 14 May 2025).

• (Bellemare-Pepin et al., 2024) "Divergent Creativity in Humans and LLMs"
• (Chen et al., 2023) "Probing the Creativity of LLMs: Can models produce divergent semantic association?"
• (Haase et al., 14 May 2025) "S-DAT: A Multilingual, GenAI-Driven Framework for Automated Divergent Thinking Assessment"
• (Damji et al., 25 Oct 2025) "Tuned for Creativity? Graph-Theoretical Mapping of Resting-State EEG Reveals Neural Signatures of Creativity"
• (Nakajima et al., 28 Jan 2026) "Beyond Divergent Creativity: A Human-Based Evaluation of Creativity in LLMs"