pith. sign in

arxiv: 2605.16477 · v1 · pith:KVQ34EJSnew · submitted 2026-05-15 · 💻 cs.LG · cs.CV

Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning

Mengye Ren This is my paper

Pith reviewed 2026-05-20 19:47 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords creativitymeta-learningdiffusion modelsgenerative modelsadaptationconcept formationreward signal
0
0 comments X

The pith

Creativity is generating stimuli that feel unfamiliar at first but become learnable after a few exposures.

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

The paper defines creativity as the production of concepts or images that an adaptive observer finds new on first encounter yet masters rapidly through limited examples. It implements this idea via a Creator-Appraiser setup: the Creator proposes a candidate, the Appraiser performs a few steps of inner-loop adaptation, and the measured improvement in the Appraiser supplies the reward that updates the Creator through meta-learning gradients. The diffusion model that acts as Creator stays completely frozen and receives no extra language prompts or parameter changes, yet the resulting outputs include stylistic shifts and concept blends absent from ordinary sampling. This framing turns the speed of adaptation itself into an objective signal for novelty.

Core claim

We propose that creativity is the production of stimuli that are unfamiliar to an adaptive observer at first sight, but quickly learnable from a few exposures. We formalize this as a Creator-Appraiser pair: a Creator generates a candidate, an Appraiser adapts to it for a few inner-loop learning steps, and the Appraiser's improvement becomes the reward the Creator optimizes through. We instantiate the framework with diffusion as the Creator, an autoencoder Appraiser on MNIST, and a CLIP Appraiser with a low-rank adapter for natural images. The diffusion model remains frozen with no additional language conditioning; the meta-learning gradient is enough to produce both stylistic variations and

What carries the argument

Creator-Appraiser pair in which the Appraiser's performance gain after a short inner-loop adaptation supplies the scalar reward that the Creator optimizes via meta-gradients.

If this is right

  • Stylistic variations and concept compositions appear that the base diffusion model never produces under standard sampling.
  • The diffusion generator requires neither parameter updates nor extra language conditioning to exhibit these new outputs.
  • The same Creator-Appraiser loop applies across simple data such as MNIST and complex natural images via lightweight adapters.
  • Creativity is operationalized as a meta-learning objective rather than as subjective novelty or diversity metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adaptation-speed reward could be applied to other frozen generators such as autoregressive or flow-based models.
  • Creativity scores derived from measured adaptation speed offer a quantitative alternative to human preference ratings.
  • The framework suggests experiments that test whether human observers also find these meta-learned stimuli faster to internalize than typical model outputs.

Load-bearing premise

That quick improvement in the Appraiser after a few adaptation steps supplies a stable, usable reward signal the Creator can optimize without any further updates or conditioning on the generator itself.

What would settle it

Generate new samples under the meta-learning objective, then measure whether a fresh Appraiser instance adapts to them in measurably fewer steps than to ordinary diffusion samples drawn from the same prompt; if adaptation speeds are statistically identical, the claimed reward signal is absent.

Figures

Figures reproduced from arXiv: 2605.16477 by Mengye Ren.

Figure 1
Figure 1. Figure 1: (a) Creativity happens at the frontier of understanding formed by existing concepts on the represen￾tation manifold learned from past experience. (b) The Creator-Appraiser framework, with a Creator 𝐶 and an Appraiser 𝐴. 𝐴 learns the concept 𝑐 created by 𝐶 in the inner loop, while 𝐶’s goal in the outer loop is to make 𝑐 unfamiliar to 𝐴 at first but highly learnable or memorable. One path to building creativ… view at source ↗
Figure 2
Figure 2. Figure 2: Implementation of the Creator-Appraiser setup in the experiments. We illustrate both the autoen￾coder and CLIP appraisers. 𝜙 denotes the adaptation parameter in the Appraiser for concept learning. Algorithm 1 Creator-Appraiser Meta-Learning (one outer iteration) Require: Creator 𝐶 with learnable variable 𝜉; Appraiser 𝐴 with slow weights 𝜃 and fast-weight initializa￾tion 𝜙0 ; loss ℒ𝐴; inner steps 𝑇 , inner … view at source ↗
Figure 3
Figure 3. Figure 3: (a) Our guided images occupy the bottom-middle of the (𝐿0 , 𝐿𝑇 ) plane—moderately unfamiliar yet highly learnable; noise lies in the top-right corner, and our shaped reward significantly exceeds the best reference category (Korean). (b) As guidance scale 𝜂 increases, generated outputs become progressively more novel and complex while retaining stroke-like structure. 0.0 0.2 0.4 0.6 0.8 Unfamiliarity L0 0.0… view at source ↗
Figure 4
Figure 4. Figure 4: Decomposition of the shaped reward. (a) The complexity gate alone cannot distinguish structured from degraded stimuli at similar 𝐿0 . (b) Adding the improvement term separates structured scripts from noise blends. (c) The full shaped reward. (d) Relative improvement inverts the desired ranking, rewarding familiarity over structured novelty. ure 4 ablates each reward component. Panel (a), complexity gate al… view at source ↗
Figure 5
Figure 5. Figure 5: Baseline comparison. (a) Each point is one generated image in the (𝐿0 , 𝐿𝑇 ) plane; the dashed diag￾onal marks 𝐿𝑇 = 𝐿0 (no learning). Color gradients show increasing 𝜂 (or 𝜎𝑝) within each method. Random perturbation stays familiar (near origin). Repulsive guidance increases unfamiliarity but rides the diagonal where no learning occurs. Only our method achieves the bottom-right: unfamiliar and learnable. (b… view at source ↗
Figure 6
Figure 6. Figure 6: Curated outputs from the CLIP Appraiser at LoRA rank 𝑟 ∈ {8, 32}, 𝜂 = 10. Each row shows four samples from our method (left) and two vanilla SD samples (right) for the same prompt; rows 1–5 are ImageNet class-conditioned, rows 6–7 are open-ended text prompts. The meta-gradient produces visually distinct stylistic interpretations within each prompt. Page 8 of 25 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of LoRA rank on the natural-image setup. As 𝑟 grows, the inner-loop reward Δsim rises sharply while frozen-CLIP sim0 drifts down from the vanilla baseline; the Qwen2.5-VL relevance acceptance rate (fraction of samples judged on-prompt) drops from 65% at 𝑟=4 to 12% at 𝑟=32. The two latter curves trace the capacity-vs-fidelity trade-off: more inner-loop ca￾pacity buys a stronger reward signal at the c… view at source ↗
Figure 8
Figure 8. Figure 8: Representative images from each method. Unguided: standard DDPM digits. Random perturba￾tion (𝜎𝑝 = 2, 5): outputs are nearly identical to unguided—the diffusion process absorbs the noise. Repulsive guidance (𝜂 = 2, 5, 10): images become progressively corrupted; higher repulsion produces more unfamiliar but incoherent outputs. Ours (𝜂 = 2, 5; 𝜇 = 0.08): novel, glyph-like patterns with coherent stroke struct… view at source ↗
Figure 9
Figure 9. Figure 9: Effect of guidance scale 𝜂 on generated image statistics (𝜇 = 0.15). (a) Unfamiliarity 𝐿0 saturates quickly beyond 𝜂 = 2. (b) Residual unfamiliarity 𝐿𝑇 drops monotonically—higher 𝜂 produces more learnable outputs. 100 samples per (prompt, rank) cell at outer-loop guidance 𝜂 = 10, 𝐾 = 10 inner-loop steps with learning rate 10−3, and 50 denoising steps. Per-image scoring. Every generated image is scored alon… view at source ↗
Figure 10
Figure 10. Figure 10: Generated images at increasing guidance scales 𝜂, corresponding to the analysis in [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Effect of ideal initial loss 𝜇 on generated image statistics (𝜂 = 10). (a) Unfamiliarity 𝐿0 scales linearly with 𝜇, confirming the complexity gate controls the unfamiliarity target as designed. (b) Residual unfamiliarity 𝐿𝑇 remains flat (≈ 0.006) regardless of 𝜇—learnability is consistently high. Page 18 of 25 [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Generated images across ideal initial loss 𝜇 values (𝜂 = 10), corresponding to the analysis in [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Layer-specific learnability. The shaped reward when fine-tuning only individual decoder layers during the inner loop, against fine-tuning the full decoder (rightmost). Adapting only the first (high-level, semantic) layer recovers most of the full-decoder reward across categories, while adapting only the last (low￾level, textural) layer yields near-zero improvement on every category — learnability is drive… view at source ↗
Figure 14
Figure 14. Figure 14: Automatic top-K outputs from the Creator-Appraiser framework on 10 additional prompts not shown in the main text. Each row is one prompt; six columns are the random sample of top-8 by Δsim within the relevance gate (seed fixed). No manual curation. Page 21 of 25 [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Each picked Creator-Appraiser sample (top of each pair) shown alongside its DINOv2-nearest neighbor (bottom of each pair) selected from 1000 vanilla Stable Diffusion samples. Top 7 prompts are the manually curated picks from [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Effect of LoRA rank: Per prompt, the top Best-of-𝑁 survivor (Qwen2.5-VL relevance ≥ 5, ranked by Δsim) at LoRA rank 𝑟 ∈ {4, 8, 16, 32}. As rank grows the inner LoRA gains capacity to assimilate increasingly out-of-distribution outputs, so the meta-gradient is free to push the diffusion sample further from the prompt. Page 24 of 25 [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
read the original abstract

What does it mean to create a new concept, rather than retrieve a familiar one? Repeatedly sampling a generative model at the same prompt produces variations with similar styles and typical content. We propose that creativity is the production of stimuli that are unfamiliar to an adaptive observer at first sight, but quickly learnable from a few exposures. We formalize this as a Creator-Appraiser pair: a Creator generates a candidate, an Appraiser adapts to it for a few inner-loop learning steps, and the Appraiser's improvement becomes the reward the Creator optimizes through. We instantiate the framework with diffusion as the Creator, an autoencoder Appraiser on MNIST, and a CLIP Appraiser with a low-rank adapter for natural images. The diffusion model remains frozen with no additional language conditioning; the meta-learning gradient is enough to produce both stylistic variations and concept compositions that the base model does not generate on its own.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes that creativity consists in producing stimuli unfamiliar to an adaptive observer yet quickly learnable from few exposures. It formalizes this idea as a Creator-Appraiser meta-learning pair: a frozen diffusion Creator generates candidates while an Appraiser (autoencoder on MNIST or CLIP with low-rank adapter on natural images) performs a few inner-loop adaptation steps; the scalar improvement in the Appraiser supplies the reward that the Creator optimizes via meta-gradients. The authors claim that this procedure yields stylistic variations and concept compositions outside the base diffusion model's typical outputs without any additional language conditioning or parameter updates to the Creator.

Significance. If the empirical claims are borne out, the framework would supply a principled, parameter-light route to conceptual novelty in frozen generative models by treating adaptation speed itself as the objective. The approach could influence research on meta-learning for creativity, efficient exploration of latent spaces, and the design of reward signals that do not require external human labels or fine-tuning.

major comments (3)
  1. [Abstract and §3] Abstract and §3 (formalization of the reward): the scalar improvement in Appraiser loss/accuracy after a small number of inner-loop steps is defined directly as the Creator's reward. Because this quantity is constructed from the same adaptation process the Creator is trying to exploit, and because the diffusion model receives neither language-conditioning updates nor parameter changes, it is unclear whether the resulting meta-gradient reliably selects for conceptual unfamiliarity rather than low-level statistics (pixel histograms, CLIP feature norms, etc.). A variance analysis or comparison against a non-adaptive baseline reward is needed to establish that the signal is load-bearing for the central claim.
  2. [§4.1] §4.1 (MNIST instantiation): the manuscript asserts that the meta-learning gradient produces outputs the base model does not generate on its own. Without reported quantitative metrics (e.g., novelty scores, human preference rates, or distance from the base distribution), ablations against random or fixed Appraisers, or error analysis, it is impossible to determine whether the observed variations are attributable to the proposed mechanism or to incidental effects of the optimization procedure.
  3. [§4.2] §4.2 (CLIP + low-rank adapter Appraiser): the rapid adaptability of the low-rank adapter risks eroding selectivity; an adapter that improves on almost any input will yield a reward signal that is only weakly correlated with conceptual novelty. The paper should demonstrate, via controlled experiments, that the adapter's improvement remains informative for the specific notion of “unfamiliar but memorable” rather than collapsing to generic feature statistics.
minor comments (2)
  1. [Notation and figures] A diagram or pseudocode block illustrating the inner-loop Appraiser updates and the outer-loop meta-gradient flow would clarify the Creator-Appraiser interaction for readers.
  2. [Introduction] The abstract and early sections would benefit from explicit comparison to prior meta-learning-for-generation and creativity-in-AI literature to better situate the novelty of the reward construction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on our work. We address each of the major comments below and have revised the manuscript accordingly to clarify the role of the meta-learning signal and to provide additional empirical support.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (formalization of the reward): the scalar improvement in Appraiser loss/accuracy after a small number of inner-loop steps is defined directly as the Creator's reward. Because this quantity is constructed from the same adaptation process the Creator is trying to exploit, and because the diffusion model receives neither language-conditioning updates nor parameter changes, it is unclear whether the resulting meta-gradient reliably selects for conceptual unfamiliarity rather than low-level statistics (pixel histograms, CLIP feature norms, etc.). A variance analysis or comparison against a non-adaptive baseline reward is needed to establish that the signal is load-bearing for the central claim.

    Authors: We agree that establishing the specificity of the adaptive reward signal is crucial for supporting our central claim. In the revised manuscript, we have added a non-adaptive baseline where the reward is based on the initial Appraiser loss without any inner-loop adaptation steps. This baseline produces outputs closer to the base diffusion model's typical generations, whereas the adaptive reward leads to more novel compositions. We also include an analysis of the variance in the reward signal across sampled candidates, showing that it varies meaningfully with conceptual features rather than solely low-level statistics. These additions are in the updated §3 and experimental sections. revision: yes

  2. Referee: [§4.1] §4.1 (MNIST instantiation): the manuscript asserts that the meta-learning gradient produces outputs the base model does not generate on its own. Without reported quantitative metrics (e.g., novelty scores, human preference rates, or distance from the base distribution), ablations against random or fixed Appraisers, or error analysis, it is impossible to determine whether the observed variations are attributable to the proposed mechanism or to incidental effects of the optimization procedure.

    Authors: We acknowledge the need for quantitative validation in the MNIST experiments. The revised version includes novelty scores computed as the increase in reconstruction accuracy after adaptation, along with distances in latent space from the base model's typical outputs. We also report results from ablations using a fixed (non-adapting) Appraiser and a random Appraiser, both of which fail to yield the same level of conceptual variation. A small-scale human study on preference for novelty is added as well. These changes strengthen the evidence that the meta-learning mechanism is responsible. revision: yes

  3. Referee: [§4.2] §4.2 (CLIP + low-rank adapter Appraiser): the rapid adaptability of the low-rank adapter risks eroding selectivity; an adapter that improves on almost any input will yield a reward signal that is only weakly correlated with conceptual novelty. The paper should demonstrate, via controlled experiments, that the adapter's improvement remains informative for the specific notion of “unfamiliar but memorable” rather than collapsing to generic feature statistics.

    Authors: This is a valid concern regarding the selectivity of the low-rank adapter. To address it, we have performed additional controlled experiments in the revision: we compare the reward signal on held-out novel concepts versus on-distribution images and versus images with manipulated low-level features (e.g., altered histograms). The results show that the improvement is significantly higher for conceptually novel but learnable stimuli, and does not correlate strongly with generic statistics like feature norms. We have included these experiments and corresponding figures in §4.2. revision: yes

Circularity Check

1 steps flagged

Reward defined directly as Appraiser adaptation improvement implements the proposed definition of creativity by construction

specific steps
  1. self definitional [Abstract]
    "We propose that creativity is the production of stimuli that are unfamiliar to an adaptive observer at first sight, but quickly learnable from a few exposures. We formalize this as a Creator-Appraiser pair: a Creator generates a candidate, an Appraiser adapts to it for a few inner-loop learning steps, and the Appraiser's improvement becomes the reward the Creator optimizes through."

    The definition of creativity is explicitly tied to 'quickly learnable from a few exposures,' which is then operationalized as the scalar improvement after inner-loop steps; therefore the Creator's meta-optimization of that improvement is optimizing the defined creativity quantity by construction.

full rationale

The paper proposes a formalization of creativity and then defines the optimization objective to be exactly the quantity named in that formalization. While the overall framework is a novel modeling choice and the experiments may demonstrate interesting outputs, the central reward signal reduces to the input definition without an independent external benchmark. This matches the moderate circularity noted in the reader's take. No self-citations, uniqueness theorems, or other enumerated patterns appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the premise that rapid adaptation improvement constitutes a valid creativity signal, with no free parameters or invented entities explicitly quantified in the abstract.

axioms (1)
  • domain assumption Rapid inner-loop adaptation improvement reliably indicates conceptual novelty and learnability.
    This premise underpins the reward definition and is invoked when stating that the Appraiser's improvement becomes the Creator's optimization target.
invented entities (1)
  • Creator-Appraiser pair no independent evidence
    purpose: To operationalize creativity as meta-learning optimization.
    New modeling construct introduced to link generation with rapid adaptation feedback.

pith-pipeline@v0.9.0 · 5677 in / 1303 out tokens · 29522 ms · 2026-05-20T19:47:33.782534+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 12 internal anchors

  1. [1]

    2004 , publisher=

    The Creative Mind: Myths and Mechanisms , author=. 2004 , publisher=

    work page 2004

  2. [2]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Image Style Transfer Using Convolutional Neural Networks , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    work page

  3. [3]

    Generative Adversarial Networks

    Generative Adversarial Nets , author=. Advances in Neural Information Processing Systems (NIPS) , year=. 1406.2661 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    A Style-Based Generator Architecture for Generative Adversarial Networks

    A Style-Based Generator Architecture for Generative Adversarial Networks , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=. 1812.04948 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    Denoising Diffusion Probabilistic Models

    Denoising Diffusion Probabilistic Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=. 2006.11239 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  6. [6]

    Computer Vision -

    Jack Lu and Ryan Teehan and Mengye Ren , title =. Computer Vision -. 2024 , url =. doi:10.1007/978-3-031-73027-6\_23 , timestamp =

    work page doi:10.1007/978-3-031-73027-6 2024

  7. [7]

    2018 , publisher=

    Reinforcement Learning: An Introduction , author=. 2018 , publisher=

    work page 2018

  8. [8]

    ACM Transactions on Graphics , volume=

    Conceptlab: Creative concept generation using vlm-guided diffusion prior constraints , author=. ACM Transactions on Graphics , volume=. 2024 , publisher=

    work page 2024

  9. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Enhancing creative generation on stable diffusion-based models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=. 2503.23538 , archivePrefix=

    work page arXiv

  10. [10]

    arXiv preprint arXiv:2505.03667 , year=

    Distribution-Conditional Generation: From Class Distribution to Creative Generation , author=. arXiv preprint arXiv:2505.03667 , year=

    work page arXiv

  11. [11]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    One-shot learning of object categories , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2006 , publisher =

    work page 2006

  12. [12]

    Proceedings of the Annual Meeting of the Cognitive Science Society (CogSci) , year =

    One shot learning of simple visual concepts , author =. Proceedings of the Annual Meeting of the Cognitive Science Society (CogSci) , year =

    work page

  13. [13]

    Prototypical Networks for Few-shot Learning

    Prototypical networks for few-shot learning , author =. Advances in Neural Information Processing Systems (NIPS) , year =. 1703.05175 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Language models are few-shot learners , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2020 , eprint=

    work page 2020

  15. [15]

    Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

    Model-agnostic meta-learning for fast adaptation of deep networks , author=. International Conference on Machine Learning (ICML) , year=. 1703.03400 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Assran et al

    Self-supervised learning from images with a joint-embedding predictive architecture , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=. 2301.08243 , archivePrefix=

    work page arXiv

  17. [17]

    Proceedings of the 25th International Conference on Machine Learning (ICML) , pages=

    Extracting and composing robust features with denoising autoencoders , author=. Proceedings of the 25th International Conference on Machine Learning (ICML) , pages=. 2008 , doi=

    work page 2008

  18. [18]

    Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volume 1: Foundations , editor=

    Learning internal representations by error propagation , author=. Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volume 1: Foundations , editor=. 1986 , doi=

    work page 1986

  19. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Masked autoencoders are scalable vision learners , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2022 , eprint=

    work page 2022

  20. [20]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=. 2019 , eprint=

    work page 2019

  21. [21]

    Neural Computation , volume=

    A fast learning algorithm for deep belief nets , author=. Neural Computation , volume=. 2006 , publisher=

    work page 2006

  22. [22]

    Auto-Encoding Variational Bayes

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Journal of Machine Learning Research (JMLR) , volume=

    A neural probabilistic language model , author=. Journal of Machine Learning Research (JMLR) , volume=. 2003 , doi=

    work page 2003

  24. [24]

    CAN: Creative Adversarial Networks, Generating "Art" by Learning About Styles and Deviating from Style Norms

    Can: Creative adversarial networks, generating" art" by learning about styles and deviating from style norms , author=. arXiv preprint arXiv:1706.07068 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    ACM Computing Surveys , volume=

    Creativity and machine learning: A survey , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    work page 2024

  26. [26]

    arXiv preprint arXiv:2510.10715 , year=

    VLM-Guided Adaptive Negative Prompting for Creative Generation , author=. arXiv preprint arXiv:2510.10715 , year=

    work page arXiv

  27. [27]

    Notes upon

    Lovelace, Ada , journal=. Notes upon. 1843 , annote=

    work page

  28. [28]

    Problems of Information Transmission , volume=

    Three approaches to the quantitative definition of information , author=. Problems of Information Transmission , volume=. 1965 , annote=

    work page 1965

  29. [29]

    The Universal Turing Machine: A Half-Century Survey , editor=

    Logical depth and physical complexity , author=. The Universal Turing Machine: A Half-Century Survey , editor=. 1988 , doi=

    work page 1988

  30. [30]

    An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion

    An image is worth one word: Personalizing text-to-image generation using textual inversion , author=. International Conference on Learning Representations (ICLR) , year=. 2208.01618 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research , year=. 2304.07193 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    IEEE Transactions on Autonomous Mental Development , volume=

    Formal theory of creativity, fun, and intrinsic motivation (1990--2010) , author=. IEEE Transactions on Autonomous Mental Development , volume=. 2010 , publisher=

    work page 1990

  33. [33]

    and Schmidhuber, J

    Srivastava, Rupesh Kumar and Steunebrink, Bas R. and Schmidhuber, J. First experiments with. Neural Networks , volume=. 2013 , publisher=. doi:10.1016/j.neunet.2013.01.022 , eprint=

    work page doi:10.1016/j.neunet.2013.01.022 2013

  34. [34]

    1960 , publisher=

    Conflict, Arousal, and Curiosity , author=. 1960 , publisher=

    work page 1960

  35. [35]

    Learning Transferable Visual Models From Natural Language Supervision

    Learning Transferable Visual Models From Natural Language Supervision , author=. International Conference on Machine Learning (ICML) , year=. 2103.00020 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and Zhong, Humen and Zhu, Yuanzhi and Yang, Mingkun and Li, Zhaohai and Wan, Jianqiang and Wang, Pengfei and Ding, Wei and Fu, Zheren and Xu, Yiheng and Ye, Jiabo and others , journal=

    work page

  37. [37]

    High-Resolution Image Synthesis with Latent Diffusion Models

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=. 2112.10752 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , eprint=

    work page 2022

  39. [39]

    Curiosity-driven Exploration by Self-supervised Prediction

    Curiosity-Driven Exploration by Self-Supervised Prediction , author=. International Conference on Machine Learning (ICML) , year=. 1705.05363 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2601.03220 , year=

    From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence , author=. arXiv preprint arXiv:2601.03220 , year=

    work page arXiv

  41. [41]

    Automatica , volume=

    Modeling by shortest data description , author=. Automatica , volume=. 1978 , publisher=

    work page 1978

  42. [42]

    arXiv preprint arXiv:2206.08549 , year=

    Rarity score: A new metric to evaluate the uncommonness of synthesized images , author=. arXiv preprint arXiv:2206.08549 , year=

    work page arXiv

  43. [43]

    The Philosophy of Creativity: New Essays , editor=

    An experiential account of creativity , author=. The Philosophy of Creativity: New Essays , editor=. 2014 , doi=

    work page 2014