Pith. sign in

REVIEW 4 major objections 5 minor 39 references

E&E alternates local novelty search with VLM-generated linguistic goals to find more diverse Flow Lenia behaviors than standard exploration baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A hybrid exploration algorithm that alternates semantic novelty search with VLM-generated linguistic goals discovers more diverse Flow Lenia behaviors than novelty search alone.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful hybrid exploration idea, but the headline comparison is not compute-matched, and the semantic-goal mechanism isn't isolated from the extra compute. the 4 major comments →

arxiv 2509.03863 v1 pith:5NUQM372 submitted 2025-09-04 cs.AI

Expedition & Expansion: Leveraging Semantic Representations for Goal-Directed Exploration in Continuous Cellular Automata

Sina Khajehabdollahi , Gautier Hamon , Marko Cvjetko , Pierre-Yves Oudeyer , Cl\'ement Moulin-Frier , C\'edric Colas This is my paper
classification cs.AI
keywords continuous cellular automataFlow Lenianovelty searchgoal-directed explorationCLIP embeddingsDINO embeddingsvision-language modelsbehavioral diversity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper introduces Expedition & Expansion (E&E), an exploration algorithm for continuous cellular automata that alternates between local novelty-driven expansion and goal-directed expeditions. In expedition phases, a vision-language model proposes textual descriptions of hypothetical interesting patterns; these are embedded in CLIP space and treated as optimization targets. Tested on Flow Lenia, E&E reports consistently higher behavioral diversity than random search, a random genetic algorithm, and a tuned Novelty Search baseline, with the gain verified in an independent DINO embedding space. A genealogical analysis shows that the few solutions created during expeditions seed a disproportionate share of the final archive's descendants. The authors argue this hybrid breaks through the local novelty plateaus that pure Novelty Search hits, and that semantic spaces make both novelty and goals more human-aligned.

Core claim

Central claim: alternating undirected novelty search with periodic, language-guided expeditions in a shared embedding space escapes the plateau pure novelty search hits in high-dimensional behavioral spaces. On Flow Lenia, every K iterations a VLM writes a 5-15 word goal (e.g., 'a photo of a four-leaf clover with bright green heart-shaped leaves'); the goal is embedded with CLIP, and sep-CMA-ES optimizes parameters from the archive's nearest neighbor to minimize cosine distance between behavior and goal embeddings. With expeditions every 50 iterations over 10,000 iterations, E&E yields larger average pairwise embedding distances than all baselines in both CLIP and DINO. Expedition-created so

What carries the argument

The engine is the alternating loop between Novelty Search expansion and VLM-guided expeditions, both operating in CLIP's joint image-text embedding space. Novelty is defined as the average cosine distance from a behavior's embedding to its k=10 nearest neighbors; expedition goals are text embeddings of 5-15 word VLM-generated descriptions; sep-CMA-ES minimizes 1 minus cosine similarity between the behavior embedding and the goal embedding. The archive stores (parameter, embedding) pairs, and DINO provides an independently trained embedding space to check that gains are not an artifact of CLIP's geometry. The load-bearing idea is treating a linguistic goal as a hypothesis about an adjacent po

Load-bearing premise

The central claim assumes that average pairwise cosine distances between frozen CLIP and DINO embeddings of final 128x128 frames faithfully capture human-perceived visual and semantic diversity; if these embedding distances misalign with how people judge Flow Lenia patterns, 'more diverse' reduces to 'more spread in two embedding spaces.'

What would settle it

Collect human dissimilarity ratings for pairs of Flow Lenia final frames and compare them with CLIP and DINO cosine distances; a weak or zero correlation would falsify the claim that E&E's diversity gains are human-aligned. Alternatively, an adversarial search that maximizes CLIP and DINO spread while yielding frames humans rate as near-identical would directly refute the diversity measure.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Frequent expeditions (every 50-100 iterations) yield more diverse archives than pure Novelty Search with the best-tuned novelty bias, and the improvement transfers to DINO, indicating the diversity is not merely CLIP-specific overfitting.
  • Sparse expeditions (every 200-300 iterations) barely improve on Novelty Search in CLIP but reduce CLIP-specific novelty inflation, so expedition frequency trades raw diversity against computational cost.
  • Expedition-discovered behaviors are disproportionately selected as parents for later novelty search, meaning goal-directed jumps serve as stepping stones rather than isolated outliers in the archive.
  • Because the same loop only requires an embedding space, a parameterized simulator, and a goal-sampling model, the recipe can be transferred to other parameter-to-behavior search domains.
  • Even partially achieved goals can be useful: expeditions that fail to fully match a linguistic description still relocate search into new regions that later expansions populate.

Where Pith is reading between the lines

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

  • Editorial extension: If the diversity gains generalize, the essential ingredient may be the model's ability to sample goals from the distribution of plausible-but-absent patterns; a purely visual generative model given archive statistics might reproduce the jump without language.
  • Editorial extension: The genealogical overrepresentation suggests a testable mechanism—expedition solutions sit in low-density regions of embedding space, making them preferred novelty parents; measuring local density before and after each expedition would confirm or reject this.
  • Editorial extension: A human similarity study on pairs of Flow Lenia final frames would directly test the 'human-aligned' framing; the paper's evidence is currently embedding distances, and the authors themselves flag novelty hacking as an open risk.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces Expedition & Expansion (E&E), a hybrid exploration algorithm for Flow Lenia that alternates between local novelty search in a CLIP embedding space and periodic 'expeditions' guided by VLM-generated linguistic goals. The authors claim that E&E consistently discovers more diverse behaviors than random parameter search, random GA, and novelty search, and that expedition-discovered solutions act as stepping stones that disproportionately influence later exploration. Diversity is measured by average pairwise CLIP and DINO embedding distances over final 128x128 behavior frames. The paper includes a parameter analysis, qualitative examples of generated goals and optimization trajectories, a genealogical analysis, and a public code/visualization link.

Significance. The core idea of alternating local novelty expansion with semantically grounded, VLM-generated goal-directed expeditions is novel and potentially valuable for open-ended exploration in artificial life and related domains. The authors deserve credit for including an independent DINO-space evaluation, which partially addresses the concern that results are merely an artifact of optimizing CLIP-specific features, and for openly acknowledging limitations such as computational cost and novelty hacking. The genealogical analysis is a useful attempt to study the long-term impact of expedition-discovered solutions. However, the current experimental validation has several load-bearing weaknesses: the main comparison is not compute-matched, the headline hyperparameters are selected using the same metrics as the final evaluation, and the genealogical null model is not appropriate under novelty-proportional selection. If the claimed effects survive controlled comparisons, this would be a solid contribution; at present the evidence is suggestive but not conclusive.

major comments (4)
  1. [Methods, Expedition phase; Figure 2] The main comparative experiment is not compute-matched. Figure 2 reports diversity as a function of 'iterations', where one iteration is one archive-insertion step. But each E&E expedition runs sep-CMA-ES for 350 steps with population N=16, which is roughly 5,600 Flow Lenia simulations, while an expansion step is a single simulation. Over 10,000 iterations with K=50, E&E performs about 179 expeditions, corresponding to roughly 1.0 million Flow Lenia simulations, versus about 10,000 for the baselines. Thus the reported diversity advantage may be a budget effect rather than evidence for the semantic-expedition mechanism. The authors should add compute-matched baselines and, crucially, a control where expeditions optimize toward randomly chosen CLIP embeddings instead of VLM-generated goals, to isolate the contribution of the semantic goals.
  2. [Results, Calibrating the expansion phase and Effect of expedition frequency] The headline parameters alpha=4 and K=50 are selected by sweeping the same CLIP/DINO diversity metrics that are later used to report the final comparative results. This is a selection-on-the-test-metric procedure, and no held-out evaluation or multiple-comparison correction is reported. With only five seeds and no significance tests, the claim that E&E 'consistently outperforms' baselines is not yet established. The authors should either report a nested evaluation (e.g., tuning on one set of seeds and evaluating on a held-out set) or provide statistical tests that account for the parameter search, and should show the sensitivity of the main comparison to nearby values of alpha and K.
  3. [Genealogical analysis] The claim that expedition-discovered solutions are 'disproportionately influential' rests on comparing their descendant fraction (42.2% +/- 13.7%) with a 4.64% expectation computed under a uniform random-GA model. But the expansion phase is a Novelty Search with selection probability p proportional to NOV^alpha, not uniform sampling. Expedition-discovered solutions may be preferentially selected simply because they are new and hence have high novelty at insertion, not because they unlock new behavioral niches. The correct control is a model or experiment that accounts for novelty-proportional selection—for example, comparing expedition solutions with expansion solutions matched for novelty at the time of insertion, or running a version of E&E where expeditions insert random high-novelty solutions rather than goal-optimized ones. The current analysis does not isolate the causal role of th
  4. [Representing behaviors with semantic embeddings; Limitations] The paper repeatedly frames the discovered diversity as 'human-aligned' and 'meaningful from a human perspective', but no human evaluation is provided. The diversity metrics are cosine distances between CLIP and DINO embeddings of final 128x128 frames, and the paper itself acknowledges 'novelty hacking' in the Limitations. Without human ratings or at least a qualitative user study, the claim that larger embedding distances correspond to meaningful perceptual or semantic differences is unsupported. The authors should either add a small human evaluation (e.g., pairwise preference or perceived difference ratings on sampled behaviors) or substantially soften the human-alignment framing and restrict the claims to diversity in the two embedding spaces.
minor comments (5)
  1. [Figure 2] The legend uses inconsistent capitalization ('Random Param Search' vs. 'Random GA'), and the top/bottom row labels 'CLIP Diversity (Training)' and 'DINO Diversity (Evaluation)' may confuse readers because both spaces are also used for hyperparameter selection. Clarify which results are used for tuning and which for final evaluation.
  2. [Algorithm 1] The pseudo-code abstracts the expedition as a single 'SEP-CMAES' call but does not indicate that this call consumes thousands of Flow Lenia simulations. A comment noting the computational cost would make the iteration-vs-evaluation distinction transparent.
  3. [Results, Table 1] The first goal in Table 1 is marked '(pre-defined)', but the text says the goals are VLM-generated. Clarify the role of the six hand-crafted seed goals and how often they are used.
  4. [Results, Calibrating the expansion phase] The text states that alpha > 4 'collapses in DINO' based on Figure 2, but the figure panel is small and does not show clear error bars. Add annotations or a separate plot to make this important point visible.
  5. [Methods, Expedition phase] The VLM is identified only as 'OpenAI's o4-mini'. Provide the exact model version/date and prompt template in an appendix for reproducibility.

Circularity Check

1 steps flagged

Minor CLIP training/evaluation overlap; DINO evaluation provides independent support, so no load-bearing circularity.

specific steps
  1. self definitional [Methods ('Representing behaviors with semantic embeddings'); 'Baselines and diversity measure'; Figure 2 axis labels]
    "During both the expansions and expeditions, novelty and goal alignment are computed as cosine distances in this structured space. ... To evaluate the diversity of an archive of solutions, we compute the average pairwise distance between embeddings in the archive. This is done separately in two distinct representation spaces: CLIP and DINO."

    E&E's expansion novelty is defined as average CLIP cosine distance to k-nearest neighbors and its expedition objective minimizes 1 − sim(ϕ(behavior(θ)), egoal) in CLIP space. The reported 'CLIP Diversity' is the average pairwise CLIP-embedding distance across the archive, i.e., success is measured as spread in the same embedding space that the algorithm explicitly optimizes. The paper even labels the CLIP curve 'CLIP Diversity (Training)' and the DINO curve 'DINO Diversity (Evaluation)'. Thus the CLIP-based part of the diversity claim is partly self-confirming by construction. However, the DINO metric is not used in the E&E objective, so the central claim that gains transfer to an independent representation space is not circular.

full rationale

The derivation chain is largely self-contained. E&E is a new combination of VLM goal generation, CLIP-based novelty, sep-CMA-ES goal optimization, and archive expansion; its reported diversity advantage is measured on archives it produced, and the main non-CLIP metric (DINO) is external to the search objective. The genealogical analysis compares empirical descendant proportions to a random-GA expectation, which is an independent statistical null model, not a fitted parameter. Hyperparameter sweeps for α and K are standard tuning, not fitted prediction. Self-citations (Etcheverry et al. 2020; Colas et al. 2022; etc.) anchor standard methods but are not load-bearing premises or uniqueness theorems. The remaining concern is a partial overlap: one of the two headline diversity metrics (CLIP) is the same space where E&E computes novelty and goal alignment, so the CLIP result is not an impartial test; the paper itself labels it 'Training'. This lowers confidence in the CLIP-specific comparison but does not affect the DINO-based evidence. The equal-iteration versus unequal-simulation-budget issue is a possible confound, but it is an experimental-design concern (compute budget), not a circularity of definitions or self-citation. Overall score 2.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central empirical claims rest on two pillars: semantic embeddings (CLIP and DINO) provide valid measures of behavioral diversity, and the chosen algorithmic constants are representative rather than cherry-picked. Several constants (alpha, K, k, CMA-ES budget, archive size) were hand-set or swept on the same evaluation metrics, which is the main source of fragility. No new physical or mathematical entities are introduced; the linguistic goals are algorithmic constructs, not entities requiring independent evidence.

free parameters (7)
  • novelty bias alpha = 4 (selected from sweep {0,1,4,7,9})
    Controls parent sampling probability p ~ NOV^alpha in expansion; selected using CLIP and DINO diversity on the same task.
  • expedition frequency K = 50 (selected from {50,100,200,300})
    Number of expansion iterations between expeditions; chosen as a balance between diversity and compute, tuned on the evaluation metrics.
  • nearest neighbors k = 10
    Used in the novelty score as the size of the neighbor set.
  • CMA-ES budget = 350 steps, population 16, sigma_init=0.1
    Fixed optimization budget and hyperparameters for every expedition.
  • archive seed size S = 1000
    Number of random parameter configurations used to initialize the archive.
  • VLM context sample size n = 25
    Number of archive images shown to the VLM to ground goal generation.
  • hand-crafted seed goals = 6 predefined goals
    Initial goals in the goal archive; their content influences the style of generated goals.
axioms (5)
  • domain assumption The final frame of a 500-step Flow Lenia simulation (128x128 image) is a sufficient behavioral descriptor.
    Behaviors are represented as final states; temporal dynamics and stability are not captured (acknowledged in Future work).
  • domain assumption CLIP cosine distance between a text-goal embedding and a behavior embedding is a valid objective for goal-directed search.
    Expedition objective is minimization of 1-sim(phi(behavior), e_goal); if CLIP text-image alignment is weak for these artificial patterns, optimization targets may be ill-posed.
  • domain assumption Average pairwise distance in CLIP and DINO embeddings measures the diversity of discovered behaviors.
    This is the reported diversity metric; no human validation that larger average distance implies more meaningfully distinct patterns.
  • domain assumption DINO embeddings serve as an independent check that novelty is not a CLIP artifact.
    DINO is trained without language; the paper assumes diversity gains that transfer to DINO reflect general visual diversity.
  • domain assumption VLM-generated linguistic goals are useful hypotheses about unseen behaviors.
    The VLM is prompted from a sample of archive images and previous goals; there is no direct measure of goal quality other than downstream diversity.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Expedition & Expansion: Leveraging Semantic Representations for Goal-Directed Exploration in Continuous Cellular Automata." pith.science (2026). https://pith.science/paper/5NUQM372

@misc{pith2026250903863,
  author       = {Pith},
  title        = {Pith review of: Expedition & Expansion: Leveraging Semantic Representations for Goal-Directed Exploration in Continuous Cellular Automata},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5NUQM372}},
  note         = {Machine review of arXiv:2509.03863}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Discovering diverse visual patterns in continuous cellular automata (CA) is challenging due to the vastness and redundancy of high-dimensional behavioral spaces. Traditional exploration methods like Novelty Search (NS) expand locally by mutating known novel solutions but often plateau when local novelty is exhausted, failing to reach distant, unexplored regions. We introduce Expedition and Expansion (E&E), a hybrid strategy where exploration alternates between local novelty-driven expansions and goal-directed expeditions. During expeditions, E&E leverages a Vision-Language Model (VLM) to generate linguistic goals--descriptions of interesting but hypothetical patterns that drive exploration toward uncharted regions. By operating in semantic spaces that align with human perception, E&E both evaluates novelty and generates goals in conceptually meaningful ways, enhancing the interpretability and relevance of discovered behaviors. Tested on Flow Lenia, a continuous CA known for its rich, emergent behaviors, E&E consistently uncovers more diverse solutions than existing exploration methods. A genealogical analysis further reveals that solutions originating from expeditions disproportionately influence long-term exploration, unlocking new behavioral niches that serve as stepping stones for subsequent search. These findings highlight E&E's capacity to break through local novelty boundaries and explore behavioral landscapes in human-aligned, interpretable ways, offering a promising template for open-ended exploration in artificial life and beyond.

Figures

Figures reproduced from arXiv: 2509.03863 by C\'edric Colas, Cl\'ement Moulin-Frier, Gautier Hamon, Marko Cvjetko, Pierre-Yves Oudeyer, Sina Khajehabdollahi.

Figure 1
Figure 1. Figure 1: Overview of Expedition & Expansion. (a) E&E alternates between expansion steps guided by Novelty Search and expedition steps driven by VLM-generated linguistic goals. (b) Example E&E exploration tree in Flow Lenia with expansions and expeditions steps. (c) Example expedition: E&E minimizes the distance between the embedding of the solution’s behavior (initialized to known solution #1362) and the VLM-genera… view at source ↗
Figure 2
Figure 2. Figure 2: Diversity analysis. Lines show the diversity of discovered solutions as measured by the average pairwise distance of CLIP (top) or DINO (bottom) solution embeddings (N=5 seeds, mean ± standard error). Left: Diversity discovered by Novelty Search (NS) under different novelty biases (p = NOVα). Center: Diversity discovered by E&E’s under different expedition frequencies (1/K). Right: Diversity discovered by … view at source ↗
Figure 3
Figure 3. Figure 3: Goal-directed expeditions. Starting from the nearest neighbor in the archive, sep-CMA-ES optimizes be￾haviors towards VLM-generated linguistic goals. Shown are key steps (0, 100, 200, 350) illustrating gradual alignment and novel pattern discovery. Genealogical analysis To evaluate the long-term impact of expedition-discovered solutions, we perform a genealogical analysis that traces the descendants of sol… view at source ↗
Figure 4
Figure 4. Figure 4: Expeditions pioneers new novelty niches. We embed the final archives in 2D with UMAP. Orange points come from a expansion-only (NS) run; blue points from E&E, with red stars marking behaviors created during expe￾dition phases. Top: example of an expedition seeding a new cluster beyond previously explored regions. Bottom: Exam￾ples behaviors discovered by expansions (blue) and expedi￾tions (red). ioral spac… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages · 5 internal anchors

  1. [1]

    Allen, J. L. (1972). An analysis of the exploratory process: The lewis and clark expedition of 1804-1806. Geographical Review , pages 13--39

    work page 1972

  2. [2]

    Askvideos-videoclip: Language-grounded video embeddings

    AskVideos (2024). Askvideos-videoclip: Language-grounded video embeddings. GitHub

    work page 2024

  3. [3]

    and Doudna, J

    Barrangou, R. and Doudna, J. A. (2016). Applications of crispr technologies in research and beyond. Nature biotechnology , 34(9):933--941

    work page 2016

  4. [4]

    Caron, M., Touvron, H., Misra, I., J \'e gou, H., Mairal, J., Bojanowski, P., and Joulin, A. (2021). Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650--9660

    work page 2021

  5. [5]

    Chan, B. W.-C. (2018). Lenia-biology of artificial life. arXiv preprint arXiv:1812.05433

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    and Lipson, H

    Clune, J. and Lipson, H. (2011). Evolving 3d objects with a generative encoding inspired by developmental biology. ACM SIGEVOlution , 5(4):2--12

    work page 2011

  7. [7]

    Colas, C., Karch, T., Sigaud, O., and Oudeyer, P.-Y. (2022). Autotelic agents with intrinsically motivated goal-conditioned reinforcement learning: a short survey. Journal of Artificial Intelligence Research , 74:1159--1199

    work page 2022

  8. [8]

    Colas, C., Madhavan, V., Huizinga, J., and Clune, J. (2020). Scaling map-elites to deep neuroevolution. In Proceedings of the 2020 Genetic and Evolutionary Computation Conference , pages 67--75

    work page 2020

  9. [9]

    Colas, C., Teodorescu, L., Oudeyer, P.-Y., Yuan, X., and C \^o t \'e , M.-A. (2023). Augmenting autotelic agents with large language models. In Conference on Lifelong Learning Agents , pages 205--226. PMLR

    work page 2023

  10. [10]

    Cully, A. (2019). Autonomous skill discovery with quality-diversity and unsupervised descriptors. In Proceedings of the Genetic and Evolutionary Computation Conference , pages 81--89

    work page 2019

  11. [11]

    and Demiris, Y

    Cully, A. and Demiris, Y. (2017). Quality and diversity optimization: A unifying modular framework. IEEE Transactions on Evolutionary Computation , 22(2):245--259

    work page 2017

  12. [12]

    Du, Y., Watkins, O., Wang, Z., Colas, C., Darrell, T., Abbeel, P., Gupta, A., and Andreas, J. (2023). Guiding pretraining in reinforcement learning with large language models. In International Conference on Machine Learning , pages 8657--8677. PMLR

    work page 2023

  13. [13]

    Etcheverry, M., Moulin-Frier, C., and Oudeyer, P.-Y. (2020). Hierarchically organized latent modules for exploratory search in morphogenetic systems. Advances in Neural Information Processing Systems , 33:4846--4859

    work page 2020

  14. [14]

    Toward Artificial Open-Ended Evolution within Lenia using Quality-Diversity

    Faldor, M. and Cully, A. (2024). Toward artificial open-ended evolution within lenia using quality-diversity. ArXiv , abs/2406.04235

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    Forestier, S., Portelas, R., Mollard, Y., and Oudeyer, P.-Y. (2022). Intrinsically motivated goal exploration processes with automatic curriculum learning. Journal of Machine Learning Research , 23(152):1--41

    work page 2022

  16. [16]

    Gravina, D., Khalifa, A., Liapis, A., Togelius, J., and Yannakakis, G. N. (2019). Procedural content generation through quality diversity. In 2019 IEEE Conference on Games (CoG) , pages 1--8. IEEE

    work page 2019

  17. [17]

    Discovering Sensorimotor Agency in Cellular Automata using Diversity Search

    Hamon, G., Etcheverry, M., Chan, B. W.-C., Moulin-Frier, C., and Oudeyer, P.-Y. (2024). Discovering sensorimotor agency in cellular automata using diversity search. arXiv preprint arXiv:2402.10236

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    Janmohamed, H., Wolinska, M., Surana, S., Pierrot, T., Walsh, A., and Cully, A. (2024). Multi-objective quality-diversity for crystal structure prediction. In Proceedings of the Genetic and Evolutionary Computation Conference , pages 1273--1281

    work page 2024

  19. [19]

    and Oudeyer, P.-Y

    Kaplan, F. and Oudeyer, P.-Y. (2004). Maximizing Learning Progress : An Internal Reward System for Development . In Embodied artificial intelligence , pages 259--270. Springer

    work page 2004

  20. [20]

    O., Isola, P., and Ha, D

    Kumar, A., Lu, C., Kirsch, L., Tang, Y., Stanley, K. O., Isola, P., and Ha, D. (2024). Automating the search for artificial life with foundation models

    work page 2024

  21. [21]

    and Stanley, K

    Lehman, J. and Stanley, K. O. (2011). Abandoning objectives: Evolution through the search for novelty alone. Evolutionary Computation , 19(2):189--223

    work page 2011

  22. [22]

    Lopes, M., Lang, T., Toussaint, M., and Oudeyer, P.-Y. (2012). Exploration in model-based reinforcement learning by empirically estimating learning progress. Advances in neural information processing systems , 25

    work page 2012

  23. [23]

    Mordvintsev, A., Randazzo, E., Niklasson, E., and Levin, M. (2020). Growing neural cellular automata. Distill , 5(2):e23

    work page 2020

  24. [24]

    and Clune, J

    Mouret, J.-B. and Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv preprint arXiv:1504.04909

    Pith/arXiv arXiv 2015

  25. [25]

    Niklasson, E., Mordvintsev, A., Randazzo, E., and Levin, M. (2021). Self-organising textures. Distill . https://distill.pub/selforg/2021/textures

    work page 2021

  26. [26]

    Papadopoulos, V., Doat, G., Renard, A., and Hongler, C. (2024). Looking for complexity at phase boundaries in continuous cellular automata. In Proceedings of the Genetic and Evolutionary Computation Conference Companion , pages 179--182

    work page 2024

  27. [27]

    P \'e r \'e , A., Forestier, S., Sigaud, O., and Oudeyer, P.-Y. (2018). Unsupervised learning of goal spaces for intrinsically motivated goal exploration. arXiv preprint arXiv:1803.00781

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  28. [28]

    Plantec, E., Hamon, G., Etcheverry, M., Oudeyer, P.-Y., Moulin-Frier, C., and Chan, B. W.-C. (2022). Flow-lenia: Towards open-ended evolution in cellular automata through mass conservation and parameter localization. The 2023 Conference on Artificial Life

    work page 2022

  29. [29]

    H., Dalal, M., Lin, S., Nair, A., Bahl, S., and Levine, S

    Pong, V. H., Dalal, M., Lin, S., Nair, A., Bahl, S., and Levine, S. (2019). Skew-fit: State-covering self-supervised reinforcement learning. arXiv preprint arXiv:1903.03698

    Pith/arXiv arXiv 2019

  30. [30]

    Pourcel, J., Colas, C., Molinaro, G., Oudeyer, P.-Y., and Teodorescu, L. (2024). Aces: Generating a diversity of challenging programming puzzles with autotelic generative models. Advances in Neural Information Processing Systems , 37:67627--67662

    work page 2024

  31. [31]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. (2021). Learning transferable visual models from natural language supervision. In International conference on machine learning , pages 8748--8763. PmLR

    2021

  32. [32]

    Reinke, C., Etcheverry, M., and Oudeyer, P.-Y. (2019). Intrinsically motivated discovery of diverse patterns in self-organizing systems. arXiv preprint arXiv:1908.06663

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  33. [33]

    and Hansen, N

    Ros, R. and Hansen, N. (2008). A simple modification in cma-es achieving linear time and space complexity. In International conference on parallel problem solving from nature , pages 296--305. Springer

    work page 2008

  34. [34]

    Schaul, T., Horgan, D., Gregor, K., and Silver, D. (2015). Universal value function approximators. In International conference on machine learning , pages 1312--1320. PMLR

    work page 2015

  35. [35]

    Takagi, H. (2002). Interactive evolutionary computation: Fusion of the capabilities of ec optimization and human evaluation. Proceedings of the IEEE , 89(9):1275--1296

    work page 2002

  36. [36]

    A., Chan, S., Strouse, D., Wang, J., Banino, A., and Hill, F

    Tam, A., Rabinowitz, N., Lampinen, A., Roy, N. A., Chan, S., Strouse, D., Wang, J., Banino, A., and Hill, F. (2022). Semantic exploration from language abstractions and pretrained representations. Advances in neural information processing systems , 35:25377--25389

    work page 2022

  37. [37]

    Tong, Z., Song, Y., Wang, J., and Wang, L. (2022). Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems , 35:10078--10093

    work page 2022

  38. [38]

    V., Zhou, D., et al

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems , 35:24824--24837

    work page 2022

  39. [39]

    Zhang, J., Lehman, J., Stanley, K., and Clune, J. (2023). Omni: Open-endedness via models of human notions of interestingness. arXiv preprint arXiv:2306.01711

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.