Pith. sign in

REVIEW 4 major objections 5 minor 39 references

E.A.R.T.H.: Structuring Creative Evolution through Model Error in Generative AI

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The E.A.R.T.H. pipeline claims that generative model errors are creative raw material, reporting a 70.4% gain in its composite creativity score.

desk verdict E.A.R.T.H.'s headline creativity gains are selection-on-the-evaluation-metric; errors alone show nothing (p=0.482), so the causal claim is unsupported. read the letter →

arxiv 2507.18004 v2 pith:Z3JIFDZ3 submitted 2025-07-24 cs.AI

classification cs.AI
keywords generativeAIcreativityerror-drivengenerationnovelty-surprise-relevancesloganhuman-in-the-loopcross-modalalignmentLLMdecoding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the outputs we call model errors—low-probability, off-template, semantically odd generations—are not noise but creative raw material. It builds a five-stage pipeline that deliberately induces such deviations, filters the most promising ones with a novelty–surprise–relevance score, amplifies and rewrites them, and then uses human feedback to steer future rounds. On a corpus of advertising slogans, the pipeline reports its composite creativity score climbing from 1.179 to 2.010, a 70.4% improvement. The central claim is that creativity lives in the deviations that standard alignment procedures are designed to suppress, so a system that curates errors instead of erasing them can produce more creative output.

What carries the argument

The load-bearing object is the composite creativity score, $R = 0.4 imes ext{Novelty} + 0.4 imes ext{Surprise} + 0.2 imes ext{Relevance}$, where Novelty is $1$ minus the cosine similarity of sentence embeddings between prompt and output, Surprise is the average negative log-likelihood under the language model, and Relevance is BERTScore F1. The same three quantities are reweighted at different stages: the Amplify stage favors novelty and divergence for seed selection, the Refine stage uses $R$ to choose prototypes, and the Transform stage uses $0.7 imes ext{Novelty} + 0.3 imes ext{Relevance}$ to pick a final rewrite. The score supplies both the selection signal that filters candidates and the evaluation metric that measures the pipeline's improvement.

What would settle it

Run the Refine stage with a different selector—for example, blind human ratings of novelty and value instead of the R-score—and check whether the final outputs still beat standard sampling; if the 52.5% jump disappears, the reported improvement is an artifact of ranking by the evaluation metric.

Watch

Extended reading notes

Core claim

The paper's central claim is that the statistical excursions of a generative model—high-temperature, low-likelihood, semantically divergent outputs—carry creative potential, and that a structured selection pipeline can unlock it. The empirical demonstration is a slogan-generation run where standard sampling produced a mean creativity score of 1.179, error-induced sampling alone did not move the needle (1.244, p = 0.482), but applying the composite R-score at the Refine stage lifted the mean to 1.898 (52.5% improvement, t = -5.56, p < 0.001). Final rewritten slogans reached 2.010, a 70.4% gain over baseline (t = -6.99, p < 0.001), with slogans 48.4% shorter and 40.7% more novel while losing only 4.0% in relevance. The paper treats the small gain from Refine to final Transform (+5.8%, not significant) as showing that Transform contributes stylistic precision rather than further creativity.

Load-bearing premise

The load-bearing premise is that the composite score of novelty, surprise, and relevance is a valid and independent measure of creativity, even though the same score selects the outputs that are then measured.

Editorial extensions

If this is right

  • If the reported gains hold, alignment-as-error-suppression is not the only viable strategy: deliberately sampling the long tail and then ranking by novelty, surprise, and relevance can raise perceived creativity while keeping relevance nearly intact.
  • Because the surprise component is computed as negative log-likelihood, it can be read directly off the model's token probabilities, making the selection mechanism portable to any autoregressive generator that exposes likelihoods.
  • Because the T-stage final outputs are 48.4% shorter while staying 40.7% more novel, the framework implies that creative and concise versions of an idea can be derived from verbose, divergent prototypes through prompt-driven rewriting.
  • The cross-modal results (CLIPScore 0.249, BERTScore F1 0.816) suggest that refined slogans retain enough concrete imagery to guide text-to-image generation, so the error-driven selection need not sacrifice grounding.
  • The H-stage human scores, with metaphorical slogans rated 4.09 versus 3.99 for literal ones, imply that future generations can bias sampling toward metaphor-rich structures without retraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's design leaves open the possibility that the largest reported gain comes from selection on the same metric used for evaluation; a drop-in test with an external creativity rubric or blind human comparison would show whether the 52.5% Refine-stage jump reflects creativity or optimizing the yardstick.
  • Implicit in the method is that the base model need not be retrained: the entire creativity gain is attributed to decoding, scoring, and rewriting, which suggests the same recipe could be applied to non-slogan domains such as product names, headlines, or visual concepts.
  • A natural next experiment is to run E.A.R.T.H. on a different corpus or a different language model and check whether the Refine-stage jump reproduces; if it does, the effect is a property of the selection policy rather than of the specific slogan theme set.
  • The paper's framing of errors is statistical rather than semantic: it operationalizes error as low token likelihood and semantic distance, so the framework does not yet distinguish productive deviations from genuinely incoherent or harmful ones; safety thresholds would be needed before this becomes a general creativity engine.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces E.A.R.T.H., a five-stage generative pipeline (Error generation, Amplification, Refine selection, Transform, Harness feedback) that deliberately induces and exploits model 'errors' to generate creative advertising slogans and cross-modal images. Using LLaMA-2-7B-Chat with high-temperature sampling, SBERT-based novelty, negative log-likelihood surprise, and BERTScore relevance, the authors report a 52.5% increase in a composite creativity score after the Refine stage and a 70.4% improvement from initial baseline to final refined outputs. The paper argues that error-centered, feedback-driven generation enhances creativity and can serve as a path toward self-evolving creative AI.

Significance. The manuscript is clearly written and offers a plausible theoretical framing that connects predictive coding, compression progress, and surprise search with a concrete generation pipeline. The qualitative examples are evocative, and the authors honestly report the non-significant effect of error generation alone and the absence of an implemented RLHF loop. However, the central quantitative claims are undermined by a fundamental circularity: the composite score used to select outputs at the Refine stage is identical to the score used to evaluate final outputs, and no control condition applies the same selection/refinement pipeline to standard outputs. The reported gains therefore do not demonstrate that error-driven generation enhances creativity. The paper also contains an explicit grid-search tuning of evaluation weights against human ratings, further compromising the independence of the final evaluation. If the claims were supported, the framework would be a useful contribution to computational creativity, but as it stands the evidence does not carry that conclusion.

major comments (4)
  1. [Sections 3.3.2 (Eq. 6) and 3.4.6 (Eq. 7)] The R-stage selection score and the final evaluation score are the same expression, Rscore = 0.4*Novelty + 0.4*Surprise + 0.2*Relevance, despite the paper's statement in Section 3.4.6 that the two scoring approaches differ. Because the R-stage selects the top variants using this metric and the paper reports the mean of the selected subset computed with the same metric, the 52.5% Err-to-Rewarded gain is the expected mathematical result of selection on the evaluation criterion, not evidence about the creative value of model errors. Table 7 confirms this: the only significant gains are the selection step and the baseline-to-final comparison, while error generation alone is not significant (p=0.482) and the Rewarded-to-Final change is not significant (p=0.272).
  2. [Table 7 and Sections 3.1-3.4] The pipeline evaluation lacks the critical control of applying the same Amplify-Refine-Transform selection and rewriting procedure to standard (temperature 0.7) outputs. Without such a control, the 70.4% Std-to-Final improvement cannot be attributed to error-based generation; it may simply reflect selection and refinement under the same scoring function. The paper's central claim that 'error-centered, feedback-driven generation enhances creativity' requires showing that error-seeded outputs improve over standard-seeded outputs when both are processed through identical downstream stages.
  3. [Section 3.4.3] The T-stage weights (0.7 for Novelty, 0.3 for Relevance) were chosen by grid search against human ratings of 'overall impressiveness.' Consequently, the final evaluation of T-stage outputs is not independent of the tuning procedure: the reported final score of 2.010 corresponds to the configuration that maximized human preference among the tested alternatives. The paper does not report scores for the other grid-search configurations (0.5/0.5, 0.8/0.2, 0.6/0.4) or apply any hold-out validation, so the final 70.4% improvement cannot be taken as an unbiased estimate of creative quality.
  4. [Sections 3.5.1 and 3.5.3] The human evaluation uses only five evaluators and fifty slogans, reports no inter-rater reliability (e.g., Krippendorff's alpha), and gives the metaphorical-versus-literal comparison (4.09 vs 3.99) without a significance test. Moreover, Section 3.5.3 explicitly states that 'this study has not yet implemented a full Reinforcement Learning with Human Feedback (RLHF) loop,' so the 'Harness feedback' stage and the paper's 'self-evolving' claim describe a proposal rather than an implemented component of the reported results.
minor comments (5)
  1. [Section 3.1.3] The 'placeholder baseline methods (CAN and DQD)' are mentioned but never defined, implemented, or used in any reported experiment; they should be described and included in the analysis or removed.
  2. [Section 3.4.6 and Table 7] The paper states 'over 500 slogan generations' but the described pipeline yields 100 slogans at Stage E and 75 at Stage A; the sample sizes for the Std and Final groups used in the t-tests are not stated. Please clarify sample sizes, whether comparisons are paired on the same themes, and how many distinct final slogans were included.
  3. [Sections 3.3.4 and 3.4.6] Figures are numbered inconsistently: Figure 2 is cited for both the length-delta histogram and the Novelty-Surprise distribution map, and Figure 3 is described as both a linguistic style profile and a Novelty-Surprise landscape. Renumber the figures so each has a unique caption and reference.
  4. [References] The reference list contains entries not cited in the text (e.g., Qi et al., 2021), and the in-text citation 'Wei et al., 2023' corresponds to a 2022 arXiv preprint with a 2023 page number; harmonize the citation style and remove uncited references.
  5. [Section 3.2.1] The Stage A creativity score in Eq. (4) includes a 'Divergence' component, but Divergence is not defined in the A-stage subsection; define it or state explicitly how it is computed.

Circularity Check

3 steps flagged · score 8.0 of 10

R-stage selection and final evaluation use the identical Rscore (Eqs. 6-7); with error-only gains insignificant (p=0.482), the headline 52.5% and 70.4% improvements are selection artifacts, not evidence for error-driven creativity.

  1. self definitional [Sec. 3.3.2 Eq. (6), Sec. 3.3.3, Sec. 3.4.6 Eq. (7)]
    "To combine these signals into a single metric, we define Rscore = 0.4Novelty+ 0.4Surprise+ 0.2Relevance(6) ... Selection: Choose the top variants to form a refined corpus for Stage T. ... These components were integrated into a unified Creativity Score using weighted aggregation: Rscore = 0.4Novelty+ 0.4Surprise+ 0.2Relevance(7)"

    Eq. (7), used to report the headline 52.5% and 70.4% improvements, is the same weighted formula as Eq. (6), which is used at the R-stage to rank and retain the 'top variants'. Recomputing the mean of a top-scoring subset under the very metric used to select it guarantees an increase relative to the unselected pool, regardless of any property of 'errors'. The final score 2.010 is therefore the selection objective, not an independent measure of error-driven creativity. The paper's own Table 7 confirms this: Std→Err is insignificant (p=0.482), while the large jump occurs exactly at the Err→Rewarded selection step.

  2. fitted input called prediction [Sec. 3.4.3 and Sec. 3.4.5]
    "We assign weights of 0.7 for Novelty and 0.3 for Relevance ... We tested multiple configurations via grid search—[0.5, 0.5], [0.8, 0.2], [0.6, 0.4]—and found that 0.7:0.3 achieved the highest 'overall impressiveness' scores in human evaluation. ... The average Novelty score increased by 40.7%"

    The T-stage chooses 'the best' rewritten candidate using Tscore = 0.7*Novelty + 0.3*Relevance, where the 0.7/0.3 weights were fitted by grid search against human ratings on this pipeline's outputs. Reporting a 40.7% novelty increase on outputs selected by a score that weights novelty at 0.7 is a selection artifact: a subset ranked by a novelty-heavy objective will have higher mean novelty by construction. The fitted weights are an input, not an independent test, so the gain cannot be cited as evidence for error-driven generation.

1 more flagged steps
  1. renaming known result [Sec. 3.4.6 and Abstract/Conclusion]
    "Incorporating deliberate errors (Err) resulted in a modest and statistically insignificant increase ($1.244$, $t=-0.71$, $p=0.482$), suggesting errors alone are insufficient for meaningful creativity improvement. ... These results demonstrate that error-centered, feedback-driven generation enhances creativity."

    The paper explicitly admits that the error-generation stage alone produces no significant creativity gain; all significant improvements occur at the reward-selection step using Eq. (6)/(7). Attributing the final 70.4% improvement to 'error-centered ... generation' renames the effect of selecting on the evaluation metric as an effect of generating from errors. No control pipeline applies the same A-R-T selection and refinement to standard (non-error) outputs, so the causal claim is not separated from the selection artifact.

full rationale

The derivation chain for the paper's central claim collapses at the point where the evaluation metric is also the selection criterion. Eq. 6 (R-stage) and Eq. 7 (final 'Creativity Score') are identical: 0.4*Novelty + 0.4*Surprise + 0.2*Relevance. The R-stage workflow states 'Choose the top variants' using this score, so the reported Err→Rewarded jump from 1.244 to 1.898 is the arithmetic consequence of averaging a top-ranked subset under the ranking function, not a measurement of creative improvement caused by error generation. The same logic contaminates the Std→Final comparison: the final outputs are selected by a pipeline whose R-stage filter is exactly the final evaluation score, and whose T-stage selector is a novelty-weighted score whose weights were grid-searched against human ratings on the same pipeline. The 40.7% novelty gain is similarly forced by selecting on a 0.7-novelty-weighted objective. The paper's own Table 7 supplies the decisive control: Std→Err (the only pure test of error generation) is insignificant (p=0.482). Because there is no control condition in which the same selection/refinement is applied to standard outputs, the central claim 'error-centered, feedback-driven generation enhances creativity' reduces to selection-on-the-evaluation-metric. This is a partial but central circularity, not a matter of author intent: the improvements are built into the scoring construction. I therefore score it 8 rather than 10 because the framework does contain independently meaningful components (e.g., cross-modal alignment scores and human preference ratings for metaphor), but the headline creativity improvements are not independent evidence for error-driven generation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a metric that is both the selection objective and the evaluation outcome, on hand-set weights and arbitrary cutoffs, and on unvalidated operationalizations of creativity. The framework introduces no new physical or conceptual entities beyond a named pipeline.

free parameters (5)
  • R-score weights (0.4, 0.4, 0.2) = 0.4, 0.4, 0.2
    Chosen by hand; used for R-stage selection (Eq. 6) and final evaluation (Eq. 7); no external justification is provided.
  • T-score weights (0.7, 0.3) = 0.7, 0.3
    Picked from grid search over [0.5,0.5], [0.8,0.2], [0.6,0.4] by 'overall impressiveness' in human evaluation (Sec. 3.4.3), i.e., fit to the authors' own evaluators.
  • Sampling hyperparameters = std: temperature=0.7, top-p=0.9; err: temperature=1.3, top-p=0.9; amplify: temperature=1.5, top-p=0.95
    Chosen by hand; the paper claims these induce 'errors' but provides no systematic search or principled justification.
  • Selection counts = 15 seeds, 5 variants per seed, top 20 variants, 50 evaluated slogans
    Arbitrary cutoffs that determine which outputs are scored and reported; no sensitivity analysis is given.
  • Stage A weights (1.0, 0.5, 0.5, 0.2) = 1.0, 0.5, 0.5, 0.2
    Hand-set composite for seed pre-screening in Eq. (1)/(4); no validation of the weighting.
assumptions (4)
  • domain assumption Creativity is measurable as a weighted sum of Novelty (SBERT 1 - cos), Surprise (LLaMA NLL), and Relevance (BERTScore F1).
    Stated in Sec. 3.3.1 and Sec. 3.4.6; no validation that these metrics track human creativity judgments, and the selection and evaluation share this same premise.
  • domain assumption Error-induced (high temperature) samples contain creative seeds worth amplifying.
    Motivates stages E and A; the paper's own data show Err alone does not raise creativity (p=0.482), so the assumption is unsupported.
  • domain assumption SBERT cosine distance, LLaMA NLL, and BERTScore F1 are valid operationalizations of novelty, surprise, and relevance.
    The entire scoring and evaluation machinery depends on this; the paper offers no calibration against human judgments for these specific slogan texts.
  • domain assumption Rewriting with a short-slogan prompt is a valid creative transform that preserves semantic intent.
    Used in the T stage; the reported length reduction and novelty gain are partly the direct effect of the instruction 'Return exactly one concise sentence'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of E.A.R.T.H.: Structuring Creative Evolution through Model Error in Generative AI." pith.science (2026). https://pith.science/paper/Z3JIFDZ3

@misc{pith2026250718004,
  author       = {Pith},
  title        = {Pith review of: E.A.R.T.H.: Structuring Creative Evolution through Model Error in Generative AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3JIFDZ3}},
  note         = {Machine review of arXiv:2507.18004}
}
read the original abstract

How can AI move beyond imitation toward genuine creativity? This paper proposes the E.A.R.T.H. framework, a five-stage generative pipeline that transforms model-generated errors into creative assets through Error generation, Amplification, Refine selection, Transform, and Harness feedback. Drawing on cognitive science and generative modeling, we posit that "creative potential hides in failure" and operationalize this via structured prompts, semantic scoring, and human-in-the-loop evaluation. Implemented using LLaMA-2-7B-Chat, SBERT, BERTScore, CLIP, BLIP-2, and Stable Diffusion, the pipeline employs a composite reward function based on novelty, surprise, and relevance. At the Refine stage, creativity scores increase by 52.5% (1.179 to 1.898, t = -5.56, p < 0.001), with final outputs reaching 2.010 - a 70.4% improvement. Refined slogans are 48.4% shorter, 40.7% more novel, with only a 4.0% drop in relevance. Cross-modal tests show strong slogan-to-image alignment (CLIPScore: 0.249; BERTScore F1: 0.816). In human evaluations, the generated outputs were consistently rated highly, demonstrating strong creative quality and expressive clarity. Feedback highlights stylistic precision and emotional resonance. These results demonstrate that error-centered, feedback-driven generation enhances creativity, offering a scalable path toward self-evolving, human-aligned creative AI.

Figures

Figures reproduced from arXiv: 2507.18004 by the authors.

Figure 1
Figure 1. The E.A.R.T.H. framework: A feedback-driven model for error [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Histogram of length differences between seed and amplified variants [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Novelty–Surprise landscape of top-ranked variants [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Comparison of novelty and relevance before and after refinement [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Novelty–relevance scatter plot of refined slogans [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]
Figure 6
Figure 6. Figure 6: Creativity score distribution across standard, error-induced, and [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Comparison of mean creativity scores: standard vs pipeline-refined [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 8
Figure 8. Figure 8: Creativity comparison between rewarded variants and final edited [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Distribution of human scores. As shown in [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]
Figure 10
Figure 10. Figure 10: Proportion of metaphorical expressions. We classified all slogans into two categories based on the presence of metaphorical imagery: • Metaphorical (label 1): e.g., “The stars light the path ahead for you.” • Non-Metaphorical (label 0): e.g., “Green energy makes the f…
Figure 11
Figure 11. Figure 11: Word cloud of human suggestions. We extracted keywords from open-ended reviewer comments to create the word cloud in [PITH_FULL_IMAGE:figures/full_fig_p033_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 24 canonical work pages

  1. [1]

    Why the standard definition of creativity fails to capture the creative act

    Abraham, A.\ (2025). Why the standard definition of creativity fails to capture the creative act. Theory & Psychology, 35(1), 40--60. https://doi.org/10.1177/09593543241290232

  2. [2]

    A.\ (2024)

    Boden, M. A.\ (2024). Creativity: definitions and computability. Journal of Cognitive Psychology, 1--11. https://doi.org/10.1080/20445911.2024.2449028

  3. [3]

    B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al.\ (2020)

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al.\ (2020). Language models are few‑shot learners. arXiv preprint. https://doi.org/10.48550/arXiv.2005.14165

  4. [4]

    Creativity

    Chen, H.\ and Ding, N.\ (2023). Probing the “Creativity” of Large Language Models: Can models produce divergent semantic association? In Findings of the Association for Computational Linguistics: EMNLP 2023, 12881--12888. https://doi.org/10.18653/v1/2023.findings-emnlp.858

  5. [5]

    R.\ and Hauser, O

    Doshi, A. R.\ and Hauser, O. P.\ (2024). Generative artificial intelligence enhances creativity but reduces the diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290

  6. [6]

    On the antibacterial action of cultures of a Penicillium

    Fleming, A.\ (1929). On the antibacterial action of cultures of a Penicillium. British Journal of Experimental Pathology, 10(3), 226--236

  7. [7]

    Predictive coding under the free‑energy principle

    Friston, K.\ and Kiebel, S.\ (2009). Predictive coding under the free‑energy principle. Philosophical Transactions of the Royal Society B, 364(1521), 1211--1221. https://doi.org/10.1098/rstb.2008.0300

  8. [8]

    L.\ and Lewis, S

    Guzman, A. L.\ and Lewis, S. C.\ (2020). Artificial intelligence and communication: A human–machine communication research agenda. New Media & Society, 22(1), 70--86. https://doi.org/10.1177/1461444819858691

Show all 39 references
  1. [9]

    Haase, J.\ and Hanel, P. H. P.\ (2023). Artificial muses: Generative AI chatbots have risen to human‑level creativity. Journal of Creativity, 33(3), 100066. https://doi.org/10.1016/j.yjoc.2023.100066

  2. [10]

    The curious case of neural text degeneration

    Holtzman, A., Buys, J., Du, L., Forbes, M., and Choi, Y.\ (2020). The curious case of neural text degeneration. arXiv preprint. https://doi.org/10.48550/arXiv.1904.09751

  3. [11]

    T.\ (2024)

    Huang, M.-H.\ and Rust, R. T.\ (2024). Automating creativity. arXiv preprint. https://doi.org/10.48550/arXiv.2405.06915

  4. [12]

    Comparison of diverse decoding methods from conditional language models

    Ippolito, D., Kriz, R., Sedoc, J., et al.\ (2019). Comparison of diverse decoding methods from conditional language models. In Proceedings of ACL 2019, 1--11. https://doi.org/10.18653/v1/p19-1365

  5. [13]

    B., Wu, C.-T., and Chao, Z

    Kern, F. B., Wu, C.-T., and Chao, Z. C.\ (2024). Assessing novelty, feasibility and value of creative ideas with an unsupervised approach using GPT‑4. British Journal of Psychology. https://doi.org/10.1111/bjop.12720

  6. [14]

    Neural Glitch / Mistaken Identity

    Klingemann, M.\ (2018). Neural Glitch / Mistaken Identity. Quasimondo. https://quasimondo.com/2018/10/28/neural-glitch/ (Accessed: 23 Apr 2025)

  7. [15]

    Metaphors We Live By

    Lakoff, G.\ and Johnson, M.\ (2003). Metaphors We Live By. University of Chicago Press

  8. [16]

    The effect of sampling temperature on problem solving in LLMs

    Lee, J.\ (2023). The effect of sampling temperature on problem solving in LLMs. arXiv preprint. https://doi.org/10.48550/arXiv.2303.07831

  9. [17]

    A diversity‑promoting objective function for neural conversation models

    Li, J., Galley, M., Brockett, C., et al.\ (2016). A diversity‑promoting objective function for neural conversation models. In NAACL 2016, 1--14. https://doi.org/10.18653/v1/n16-1014

  10. [18]

    Leveraging hallucinations in large language models to enhance drug discovery

    MarkTechPost\ (2025). Leveraging hallucinations in large language models to enhance drug discovery. MarkTechPost, January 28. https://www.marktechpost.com/2025/01/27/leveraging-hallucinations-in-large-language-models-to-enhance-drug-discovery/ (Accessed: 23 Apr 2025)

  11. [19]

    Training language models to follow instructions with human feedback

    Ouyang, L., Wu, J., Jiang, X., et al.\ (2022). Training language models to follow instructions with human feedback. arXiv preprint. https://doi.org/10.48550/arXiv.2203.02155

  12. [20]

    Driven by compression progress: A simple theory of curiosity, creativity, art, science, and jokes

    Schmidhuber, J.\ (2009). Driven by compression progress: A simple theory of curiosity, creativity, art, science, and jokes. In G. Pezzulo et al.\ (eds), Anticipatory Behavior in Adaptive Learning Systems, 48--76. Springer. https://doi.org/10.1007/978-3-642-02565-5_4

  13. [21]

    Simple algorithmic principles of discovery, subjective beauty, selective attention, curiosity and creativity

    Schmidhuber, J.\ (2007). Simple algorithmic principles of discovery, subjective beauty, selective attention, curiosity and creativity. arXiv preprint. https://doi.org/10.48550/arXiv.0709.0674

  14. [22]

    Instances need more care: Rewriting prompts for instances with LLMs in the loop yields better zero‑shot performance

    Srivastava, S., Mishra, A., and Jain, P.\ (2024). Instances need more care: Rewriting prompts for instances with LLMs in the loop yields better zero‑shot performance. In Findings of ACL 2024, 6211--6232. https://doi.org/10.18653/v1/2024.findings-acl.371

  15. [23]

    N.\ and Liapis, A.\ (2016)

    Yannakakis, G. N.\ and Liapis, A.\ (2016). Searching for surprise. In Proceedings of ICCC 2016, 1--8

  16. [24]

    Amabile, T. M. (1996). Creativity in context. Boulder, Colo: Westview Press

  17. [25]

    Amabile, T. M. and Pillemer, J. (2012). Perspectives on the social psychology of creativity. The Journal of Creative Behavior, 46(1), 3--15. https://doi.org/10.1002/jocb.001

  18. [26]

    and Johnson-Laird, P

    Oatley, K. and Johnson-Laird, P. N. (2014). Cognitive approaches to emotions. Trends in Cognitive Sciences, 18(3), 134--140. https://doi.org/10.1016/j.tics.2013.12.004

  19. [27]

    Amabile, T. M. (1983). The social psychology of creativity. New York, NY: Springer. https://doi.org/10.1007/978-1-4612-5533-8

  20. [28]

    Foss, S. (2004). Rhetorical Criticism: Exploration & Practice

  21. [29]

    Li, J., Li, X., Xia, L., Jiang, Y.-G., Han, S., Sun, M., and Wang, X. (2022). BLIP: Bootstrapping Language–Image Pre‑training for Unified Vision‑Language Understanding and Generation. arXiv preprint. https://doi.org/10.48550/arXiv.2201.12086

  22. [30]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I. (2021). Learning Transferable Visual Models From Natural Language Supervision. arXiv preprint arXiv:2103.00020. https://doi.org/...

  23. [31]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 10674--10685. https://doi.org/10.1109/CVPR5268...

  24. [32]

    Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., & Christiano, P. (2022). Learning to Summarize from Human Feedback. Advances in Neural Information Processing Systems. arXiv:2009.01325. https://doi.org/10.48550/arXiv.2009.01325

  25. [33]

    Ul Haq, I., Pifarré, M., & Fraca, E. (2024). Novelty Evaluation using Sentence Embedding Models in Open-ended Cocreative Problem-solving. International Journal of Artificial Intelligence in Education, 34(4), 1599--1626. https://doi.org/10.1007/s40593-024-00392-3

  26. [34]

    Q., & Artzi, Y

    Zhang, T., Kishore, V., Wu, F., Weinberger, K. Q., & Artzi, Y. (2020). BERTScore: Evaluating Text Generation with BERT. arXiv preprint arXiv:1904.09675. https://doi.org/10.48550/arXiv.1904.09675

  27. [35]

    S., Demeter, D., Downey, D., & Birnbaum, L

    Bursztyn, V. S., Demeter, D., Downey, D., & Birnbaum, L. (2022). Learning to Perform Complex Tasks through Compositional Fine‑Tuning of Language Models. arXiv preprint arXiv:2210.12607. Available at: https://doi.org/10.48550/arXiv.2210.12607

  28. [36]

    L., & Liang, P

    Li, X. L., & Liang, P. (2021). Prefix‑Tuning: Optimizing Continuous Prompts for Generation. arXiv preprint arXiv:2101.00190. Available at: https://doi.org/10.48550/arXiv.2101.00190

  29. [37]

    Qi, T., Zong, C., et al. (2021). HieRec: Hierarchical User Interest Modeling for Personalized News Recommendation. In C. Zong et al. (Eds.), Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on N...

  30. [38]

    Y., et al

    Wei, J., Bosma, M., Zhao, V. Y., et al. (2023). Chain‑of‑Thought Prompting Elicits Reasoning in Large Language Models. arXiv preprint arXiv:2201.11903. Available at: https://doi.org/10.48550/arXiv.2201.11903

  31. [39]

    Reimers, N., & Gurevych, I. (2019). Sentence‑BERT: Sentence Embeddings using Siamese BERT‑Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP‑IJCNLP...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.