PaperlayerView exact revision on arXiv

Source-linked AI summary

Probing the Creativity of Large Language Models: Can models produce divergent semantic association?

Honghua Chen, Nai Ding

arXiv:2310.11158v1cs.CLcs.LG

TL;DR

The paper addresses whether large language models can generate creative content and whether decoding strategies support it. It evaluates models with the divergent association task, comparing their scores across decoding methods and with human results. Under greedy search, GPT-4 surpasses 96.1% of humans, while stochastic decoding can raise other models’ scores but trades off stability.

  • Problem

    It is unclear whether LLMs genuinely exhibit creativity during modeling or whether decoding strategies help, since sampling can produce creative and nonsensical content alike.

    Source: §1 · p. 2

  • Method

    The study applies the divergent association task, asking models to generate unrelated nouns and scoring their pairwise semantic distances, then compares models and decoding strategies.

    Source: §1 · p. 2, §2 · p. 2

  • Results

    Under greedy search, GPT-4 scores 89.1 DAT, surpassing 96.1% of humans; GPT-3.5-Turbo scores 80.8, above average, while stochastic decoding raises other models’ scores but destabilizes outputs.

    Source: §4 · pp. 4–5, §5 · p. 5

  • Takeaways & Limitations

    The results indicate that LLMs can generate divergent content with instruction, but stochastic decoding does not reliably distinguish creative from nonsensical content.

    Source: §4 · p. 5, §5 · p. 5

    The study evaluates little-C creativity and self-contained creativity, so its findings cannot be directly generalized to all language-generation tasks or human–AI co-creation.

    Source: Limitation · p. 5

Abstract

from arXiv · show

Large language models possess remarkable capacity for processing language, but it remains unclear whether these models can further generate creative content. The present study aims to investigate the creative thinking of large language models through a cognitive perspective. We utilize the divergent association task (DAT), an objective measurement of creativity that asks models to generate unrelated words and calculates the semantic distance between them. We compare the results across different models and decoding strategies. Our findings indicate that: (1) When using the greedy search strategy, GPT-4 outperforms 96% of humans, while GPT-3.5-turbo exceeds the average human level. (2) Stochastic sampling and temperature scaling are effective to obtain higher DAT scores for models except GPT-4, but face a trade-off between creativity and stability. These results imply that advanced large language models have divergent semantic associations, which is a fundamental process underlying creativity.

1 Introduction

The paper asks whether language models can generate creative content despite their next-word prediction objective, and whether decoding choices affect that capacity. It approaches the question through divergent associations and comparison with human results.

1 Introduction

  • The motivation spans LLMs’ language mastery and reported reasoning abilities, alongside the importance of creative language and creativity in human life.The introduction notes that LLM creativity may be relevant to art, science, and everyday life.Source: §1 · p. 1
  • The paper frames creativity as producing something novel and valuable through remote association while inhibiting common ideas.It contrasts this account with the view that models predict within existing language distributions rather than create new ones.Source: §1 · pp. 1–2
  • Whether LLMs genuinely have creativity remains unclear because stochastic decoding can produce both creative and nonsensical content, which next-word prediction does not distinguish.Greedy decoding may produce dull, repetitive text, while stochastic strategies have received better human preferences.Source: §1 · p. 2
  • The study tests whether LLMs can generate creative content and whether decoding strategies help, using the divergent association task and comparison with humans.The task asks models to generate unrelated nouns and measures their pairwise semantic distance.Source: §1 · p. 2

2 Measuring Creativity

The section motivates an objective, semantic approach to creativity assessment, while explaining why expert ratings and alternate-use tasks have drawbacks. It presents DAT as a reproducible measure but notes that its applicability to LLMs requires examination.

2 Measuring Creativity

  • Expert judgments of creative quality are subjective because novelty and value are relative, and raters differ with their surprise and domain knowledge.The study therefore considers an automatically assessed semantic approach.Source: §2 · p. 2
  • Alternate-use tasks depend on the chosen object, require human ratings, and risk data leakage from uses recorded in training data.The task asks participants to generate unusual uses for an object.Source: §2 · p. 2
  • DAT prompts people to reject related associations and generate unrelated nouns, then measures their average pairwise cosine distance in semantic space.The authors describe DAT as valid and reliable and report that it correlates with other creativity metrics.Source: §2 · pp. 2–3
  • Applying DAT to LLMs requires checking its applicability: similar semantic networks do not establish equal ability to produce remote associations.Emotion and motivation are not considered in this study.Source: §2 · pp. 2–3

3 Experiment Setup

The experiment compares five language models and decoding settings using DAT, with human and random-word reference conditions. It also examines semantic-distance measurement choices and sets up comparisons to validate the task for models.

3.1 Models

  • Five models were studied: GPT-4, GPT-3.5-Turbo, Oasst-Llama-30B, Vicuna-13B, and ChatGLM-6B.The latter three were fine-tuned from Llama or based on GLM; GPT-4 and GPT-3.5-Turbo used pre-training and RLHF.Source: §3.1 · p. 3

3.2 Decoding strategy

  • The study compares greedy search with top-p sampling at p = 0.9 and temperature t = 0.7, and varies temperature to examine its effect.It collects enough samples for results to converge.Source: §3.2 · p. 3

3.3 DAT

  • Models generate 10 unrelated nouns; after invalid words are filtered, the first seven valid words are scored using GLoVe-based semantic distances.Word2Vec and FastText results were similar, with correlation coefficients of 0.82 and 0.91, respectively; average surprisal was also measured to examine word-frequency effects. Human comparisons use 8,572 participants, with randomly sampled WordNet nouns as a reference.Source: §3.3 · p. 3

3.4 Validating DAT

  • The DAT validation setup contrasts DAT-prompt responses with base and random prompts; Figure 3 presents model word-distance matrices, model and human DAT, and human percentiles.The authors hypothesize that models produce associated words without DAT instructions but can make divergent associations under the DAT prompt.Source: §3.4 · p. 3, Figure 3 · p. 4

4 Result

Greedy search puts GPT-4 above most humans and GPT-3.5-Turbo above average human performance, while stochastic decoding and temperature tuning can raise DAT scores but reduce stability. Analyses also show that word rarity affects some scores and that instructions can shift LLM outputs toward more divergent words.

4 Result

  • GPT-4 scores 89.1 on DAT, surpassing 96.1% of humans, while GPT-3.5-Turbo scores 80.8, above average human performance.Other models perform less well, with scores roughly proportional to model size.Source: §4 · p. 4
  • Top-p sampling raises DAT above greedy search for models other than GPT-4, but results become unstable and may include low-scoring answers.The authors report higher scores alongside this instability.Source: §4 · p. 4
  • At the same surprisal, GPT-4 and GPT-3.5-Turbo exceed average human DAT, while other models fall short; controlling surprisal, both still outperform average human performance, though GPT-4's superiority is attenuated.Source: §4 · p. 4
  • Temperature from 0.1 to 1 has a limited positive effect on models except GPT-4, while higher settings bring instability and invalid answers.The temperature-tuning figure's bands indicate standard deviations.Source: §4 · p. 5, Figure 5 · p. 5
  • WordNet noun sampling can outperform most humans and models, while LLMs produce more related words without instruction and shift toward divergence when prompted.The authors attribute WordNet's performance to its lack of constraints from the language distribution, which limits its ability to perform general language tasks.Source: §4 · p. 5, Figure 6 · p. 5
  • The authors conclude that instructions can elicit divergent content, but models do not inhibit common words; stochastic decoding helps remote association yet may yield nonsense or degrade advanced-model performance.They note that advanced LLMs can implicitly modulate probability distributions and stably generate divergent answers.Source: §4 · p. 5

5 Conclusion

The study evaluates creativity with a divergent semantic task and finds different results across LLMs and decoding strategies. GPT-4 shows stable advanced human-level creativity, GPT-3.5-Turbo exceeds average human level, and stochastic decoding is effective but insufficient for creative generation.

5 Conclusion

  • GPT-4 demonstrates advanced human-level creativity stably, while GPT-3.5-Turbo exceeds the average human level.Source: §5 · p. 5
  • The divergent semantic task reveals distinguishable results across LLMs and decoding strategies.Source: §5 · p. 5
  • Stochastic decoding is effective, but not enough for creative generation.Source: §5 · p. 5

Limitation

The study evaluates everyday creative capacity rather than rare, marvelous creative products. Its findings are bounded to self-contained creativity and cannot be directly generalized to all language-generation tasks.

Limitation

  • The study evaluates little-C, or everyday creative capacity, rather than rare big-C creative products.Source: Limitation · p. 5
  • Because creativity measurement is controversial and requires multiple perspectives, the results cannot be directly generalized to all language-generation tasks.Source: Limitation · p. 5
  • The study is limited to self-contained creativity; human-AI co-creation is left for future work.Source: Limitation · p. 5

A Selecting sample size for each model

The appendix describes how sample sizes were selected to obtain stable top-p sampling results. It reports that GPT-4 and GPT-3.5-Turbo generate answers more stably.

A Selecting sample size for each model

B Results of the DAT and surprisal on human and random and baselines

The appendix compares DAT and surprisal for human and random baselines. Some people surpass the random baseline and approach ceiling DAT at particular surprisal levels, while DAT and surprisal are positively related.

B Results of the DAT and surprisal on human and random and baselines

C Result of the DAT and surprisal for more models using greedy search

Figure 9 reports DAT and surprisal results for more LLMs using greedy search, with contours indicating 95% confidence intervals.

C Result of the DAT and surprisal for more models using greedy search

D Controlling surprisal as a confounding variable

After controlling surprisal as a potential influence on semantic-distance measurement, GPT-4 and GPT-3.5-Turbo remain above average human performance, while other models are below it. GPT-4 and Oasst-Llama-30B lose their superiorities because their high DAT scores partially depend on generating rarer words.

D Controlling surprisal as a confounding variable

Loading 2310.11158v1…