Pith. sign in

REVIEW 3 major objections 4 minor 38 references

ZapGPT: Free-form Language Prompting for Simulated Cellular Control

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Trained on a single prompt—'form a cluster'—simulated cells learn to follow eight unseen instructions, including semantic opposites such as 'scatter apart'.

desk verdict ZapGPT is a neat proof-of-concept for free-form language control of a simple 2D cell simulation, but the headline generalization claim is weakened by using the same vision-language model as both the evolutionary fitness and the test evaluator; the 'drift apart' example shows the grader can be permissive. read the letter →

arxiv 2509.10660 v1 pith:X3Z2NDFG submitted 2025-09-12 cs.AI cs.MAq-bio.CB

classification cs.AIcs.MAq-bio.CB
keywords free-formlanguagepromptingprompt-to-interventionvision-languagemodelevaluationevolutionaryoptimizationsimulatedcellularcontrolgroundinggeneralizationtounseenpromptsdecentralizedbehavior
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 claims to be the first demonstration that free-form natural language alone—no engineered reward, no fixed command set—can guide the collective behavior of simulated cells. The pipeline maps a sentence to a spatial force field via a prompt-to-intervention model, then uses a vision-language model to score how well the resulting final image matches the prompt; an evolution strategy optimizes the mapping using only that language-based score. Trained on a single prompt, 'form a cluster', the system produces plausible, distinct behaviors for eight previously unseen instructions, including semantic opposites such as 'scatter apart'. If correct, this opens a route to controlling decentralized computational, robotic, or biological systems by speaking to them.

What carries the argument

The P2I (prompt-to-intervention) model maps a sentence embedding to a spatial vector field via a 5x5x64 latent and transposed convolutions; the field, applied as an external force over many timesteps, is the only channel through which language touches the cells. The second load-bearing component is the VLM-D2R evaluator (Mistral-Vision), which converts a rendered final image plus the original prompt into a free-form description and a scalar alignment score; that score is the fitness. The (mu+lambda) evolution strategy is the optimizer that closes the loop, since the simulator is non-differentiable.

What would settle it

Use an independent scoring source—human raters or explicit geometric criteria—on the same eight unseen prompts with the same trained controller, and check agreement with the VLM scores. The paper already contains a warning sign: 'Drift apart from one another' received a VLM score of 0.927 while showing no significant change in average pairwise distance, so if other prompts show similar divergence, the claimed language grounding collapses to grader compliance.

Watch

Extended reading notes

Core claim

ZapGPT's central claim is that a decentralized agent collective can be guided by free-form text without engineered rewards or prompt templates. The pipeline is closed loop: a SentenceTransformer encodes the prompt; a convolutional decoder expands that embedding into a low-resolution 2D vector field; the field is interpolated and applied to 50 simulated cells with local repulsion for 500 timesteps; a pretrained vision-language model (Mistral-Vision, called D2R) renders the final image, describes it, and returns a 0-to-1 alignment score; and a (mu+lambda) evolution strategy mutates the prompt-to-intervention weights to maximize that score. The authors trained the system on only the prompt 'for

Load-bearing premise

The argument depends on the vision-language model's alignment score being a genuine measure of whether a final image matches the prompt; since the same model both drives evolution and evaluates the trained system, a permissive grader would make the reported generalization an artifact.

Editorial extensions

If this is right

  • Natural language alone can serve as a control layer for decentralized systems, with no engineered fitness function or task-specific reward.
  • A controller evolved on one prompt can generalize to unseen instructions, including semantic opposites, without retraining—a form of zero-shot behavioral transfer.
  • The vision-language alignment score tracks physically measurable spatial change (average pairwise distance) for most tested prompts, so language-based evaluation can stand in for geometric ground truth.
  • The same evolution-plus-VLM loop can in principle be pointed at other non-differentiable simulators or embodied collectives, making language prompting a general interface for goal-directed emergence.
  • Language-guided control could extend beyond spatial morphology to transcriptional, metabolic, or bioelectric state spaces, where goals are naturally expressed in human terms.

Reading between the lines

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

  • Because the same VLM is both fitness function and grader, the claimed generalization is only as strong as that grader; an independent test with human raters would confirm whether the alignment scores correspond to perceived behavior.
  • The 'drift apart from one another' case—high VLM score but no significant change in pairwise distance—suggests the mapping may sometimes exploit image-level cues (e.g., spread to corners) rather than the literal instruction, and is worth probing with more prompts of that form.
  • If the P2I embedding is doing the work, then interpolating between prompt embeddings should produce intermediate spatial behaviors; that is a testable prediction the paper does not run.
  • The framework suggests a reciprocal design: if cells could generate language as well as consume it, language models could become a shared interface between humans and living collectives—an idea the authors raise but leave open.
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

3 major / 4 minor

Summary. The paper introduces ZapGPT, a language-driven control pipeline for simulated cellular collectives. A Prompt-to-Intervention (P2I) model maps a sentence embedding produced by SBERT into a spatial vector field via a convolutional decoder; the field is applied as an external force to 50 simulated agents over 500 timesteps. The rendered final state is then scored by a Mistral-Vision vision-language model (D2R), which returns a scalar alignment score measuring how well the image matches the prompt. P2I weights are evolved using a (µ+λ) evolution strategy with the D2R score as the fitness signal, training only on the single prompt “form a cluster.” The authors report that the evolved model generalizes to eight unseen prompts, including scattering prompts with meanings opposite to the training prompt, based on high D2R alignment scores and, for most prompts, statistically significant changes in average pairwise distance. The paper claims this demonstrates an abstract, compositional mapping between language and spatial dynamics, enabling conceptual inversion beyond the training distribution.

Significance. If the claims hold, the work is a useful step toward using free-form natural language as a control interface for decentralized, agent-based systems without engineered reward functions. The paper has concrete strengths: the approach is simple and reproducible, code and data are provided, multiple random seeds and vector-field resolutions are tested, and the use of an independent behavioral metric (average pairwise distance) partially corroborates the VLM-based training signal. The independent PWD evidence supports clustering and three of four scattering prompts, which is nontrivial and suggests the learned mapping is not purely arbitrary. However, the central generalization claim is weakened by a training/evaluation overlap: the same VLM that provided the fitness signal during evolution is also the sole semantic grader for the test prompts. The paper's own data contain a dissociation between VLM score and objective behavior. These issues are fixable with additional controls and independent evaluation, but they are load-bearing for the paper's main conclusion.

major comments (3)
  1. [§3.1.6, Algorithm 1, §4.5] The D2R alignment score is used as the fitness function during evolution (Section 3.1.6, Algorithm 1), and the same D2R model is then used to evaluate the unseen test prompts (Section 4.5: “we evaluated each with the same D2R model used during training”). Consequently, high D2R scores on unseen prompts may partly reflect grader permissiveness or a prompt-agnostic bias rather than genuine semantic alignment. The independent PWD metric validates clustering and 3/4 scattering prompts, but it does not cover all prompts and cannot by itself establish the abstract compositional mapping claimed in Section 4.5. I recommend replacing or supplementing the D2R-based test evaluation with a held-out VLM that was never used as a fitness signal, or with objective behavioral metrics defined for every prompt, plus a demonstration that D2R scores on unseen prompts correlate with those independent measures
  2. [Table 1 vs. Figure 11] The dissociation for “Drift apart from one another” is direct evidence that the VLM grader can award a high alignment score when the objective behavior is absent. Table 1 reports a mean VLM score of 0.927 with p=1.30e−7, yet Figure 11 shows no significant change in average pairwise distance for this prompt. Since the paper's own exception demonstrates that the VLM can be permissive in exactly the scattering case central to the semantic-inversion claim, the high scores for other scattering prompts cannot be interpreted without additional evidence that the VLM scores track dispersal. The manuscript should either explain this dissociation, or qualify the semantic-inversion conclusion and provide an independent measure of dispersion for all scattering prompts.
  3. [§4.5, Table 1] The statistical analysis in Table 1 tests only whether the mean VLM score exceeds 0.5 (one-sided Wilcoxon test). This does not establish that the score reflects alignment with the specific prompt, because the D2R might assign generally high scores to most rendered multi-agent configurations regardless of wording. To support the generalization claim, the paper needs baselines such as: scores for the same rendered images paired with an unrelated prompt; scores from an untrained/random P2I model; or scores for random vector-field interventions. Without such controls, the reported p-values only show that the scores are above an arbitrary threshold, not that the model responds differentially to prompt semantics.
minor comments (4)
  1. [§4.1, Figures 4–6] Figures 4, 5, and 6 contain only an asterisk and appear to be placeholder captions; they should either be removed or replaced with the actual training-curve figures.
  2. [§4.4] The text describes panels (a–h) of Figure 9 but does not explicitly map prompt names to panels; adding this mapping would improve reproducibility.
  3. [§4.5] The phrase “some quantitative evaluation” in the section title undersells the content; consider renaming to “Quantitative evaluation of unseen prompts” or similar.
  4. [§5.2] The limitations section notes the black-box nature of the VLM but does not mention that the same VLM was used for both fitness and evaluation. This training/evaluation overlap should be acknowledged as a limitation and discussed.

Circularity Check

1 steps flagged · score 6.0 of 10

Generalization is scored by the same VLM that was the fitness signal; 'Drift apart' shows the grader can be permissive.

  1. fitted input called prediction [Section 3.1.6 / Algorithm 1; Section 4.5 / Table 1]
    "This score serves as the fitness signal during training... (Sec 3.1.6). Algorithm 1: 'Query VLM with I and prompt p to get fitness score f... P←top µ models in C by score f.' Section 4.5: 'we ran 30 simulations per test prompt and evaluated each with the same D2R model used during training.' Table 1 reports 'alignment scores significantly above 0.5' as evidence of generalization."

    The P2I model is evolved to maximize the D2R alignment score, and the same D2R model is then used as the evaluation metric for unseen prompts. High 'alignment scores' are therefore re-measurements of the exact objective that the evolutionary search optimized, not an independent test of semantic grounding. The paper does include an independent PWD metric, but for scattering it supports only 3/4 prompts; in the 'Drift apart from one another' case the VLM awards 0.927 while PWD shows no significant dispersal, demonstrating that the VLM can be permissive. Thus the central claim of abstract compositional generalization rests partly on the same grader that served as fitness, making the result partially circular by construction.

full rationale

The paper's principal evidence for generalization is the D2R alignment score, yet D2R is explicitly the fitness function used during evolution (Algorithm 1; Section 3.1.6). Testing with the same model (Section 4.5) means high scores are not independent confirmation of semantic grounding. The independent pairwise-distance analysis provides some external support, but it is partial and contradicts the VLM in the 'Drift apart' case, showing the grader can be permissive. This is a clear instance of the evaluation metric overlapping with the optimized objective, though not a fully deterministic reduction because the test prompts were unseen during training. Score 6 reflects the partial circularity: the central generalization claim is heavily dependent on the same VLM that defined the fitness, with only secondary independent validation.

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

The system uses existing pretrained tools (SBERT, Mistral-Vision) and simple physics; no new physical entities or forces are proposed. The central fitted object is the P2I decoder, plus environment and optimization hyperparameters. The main unstated axiom is that the VLM score is a trustworthy semantic yardstick, which the paper's own Drift apart data point calls into question.

free parameters (4)
  • P2I model weights = evolved (not stored in paper)
    The projection and convolutional decoder weights are the fitted quantities, optimized by evolution against VLM scores. This is the core learned object.
  • Vector field grid resolution = 2x2, 5x5, 10x10
    Three resolutions were tested; the choice affects fitness curves and generalization behavior.
  • Evolution hyperparameters = mu=5, lambda=15, G=50, sigma=0.1, 30 seeds
    Chosen by hand without sensitivity analysis; results are conditional on this schedule.
  • Simulation constants = 50 agents, 500x500 canvas, T=500, repulsion force
    Fixed environment parameters; the claimed behaviors and generalization are dependent on these choices.
assumptions (4)
  • domain assumption Mistral-Vision alignment scores are valid measures of prompt-image semantic alignment.
    The entire training and evaluation loop treats the VLM score as a faithful fitness and as ground truth for test prompts. Invoked in Sections 3.1.5 and 4.5.
  • domain assumption SBERT embeddings capture the semantic contrasts (e.g., cluster vs scatter) needed for behavioral differentiation.
    Figure 2 supports this for two sample prompts; the decoder operates on these embeddings. Section 3.1.2.
  • domain assumption The 2D physics simulation with repulsion and an external field is a sufficient substrate for emergent language-guided behavior.
    All experiments use this environment; no transfer to real cells is claimed. Section 3.1.4.
  • standard math (mu+lambda) evolution strategy can effectively search the P2I weight space given VLM evaluations.
    Background optimization method; its effectiveness for this specific noisy fitness landscape is not proven. Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ZapGPT: Free-form Language Prompting for Simulated Cellular Control." pith.science (2026). https://pith.science/paper/X3Z2NDFG

@misc{pith2026250910660,
  author       = {Pith},
  title        = {Pith review of: ZapGPT: Free-form Language Prompting for Simulated Cellular Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X3Z2NDFG}},
  note         = {Machine review of arXiv:2509.10660}
}
read the original abstract

Human language is one of the most expressive tools for conveying intent, yet most artificial or biological systems lack mechanisms to interpret or respond meaningfully to it. Bridging this gap could enable more natural forms of control over complex, decentralized systems. In AI and artificial life, recent work explores how language can specify high-level goals, but most systems still depend on engineered rewards, task-specific supervision, or rigid command sets, limiting generalization to novel instructions. Similar constraints apply in synthetic biology and bioengineering, where the locus of control is often genomic rather than environmental perturbation. A key open question is whether artificial or biological collectives can be guided by free-form natural language alone, without task-specific tuning or carefully designed evaluation metrics. We provide one possible answer here by showing, for the first time, that simple agents' collective behavior can be guided by free-form language prompts: one AI model transforms an imperative prompt into an intervention that is applied to simulated cells; a second AI model scores how well the prompt describes the resulting cellular dynamics; and the former AI model is evolved to improve the scores generated by the latter. Unlike previous work, our method does not require engineered fitness functions or domain-specific prompt design. We show that the evolved system generalizes to unseen prompts without retraining. By treating natural language as a control layer, the system suggests a future in which spoken or written prompts could direct computational, robotic, or biological systems to desired behaviors. This work provides a concrete step toward this vision of AI-biology partnerships, in which language replaces mathematical objective functions, fixed rules, and domain-specific programming.

Figures

Figures reproduced from arXiv: 2509.10660 by the authors.

Figure 1
Figure 1. Overview of the ZapGPT pipeline. A natural language prompt (a) is converted into a spatial vector field [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Cosine similarity between prompt embeddings using BERT (left) vs. SentenceTransformer (right). While [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. P2I architecture. A 384-dimensional prompt embedding is projected into a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: * (a) Grid 2 × 2 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: * (b) Grid 5 × 5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: * (c) Grid 10 × 10 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Evaluation of clustering behavior using average pairwise distance, comparing best and worst performing seeds for each grid size. Best seeds consistently reduce inter-agent distance after simulation, while worst seeds often fail to induce meaningful change. 4.5 Some Qua…
Figure 9
Figure 9. Figure 9: Behavioral responses of the trained ZapGPT model to various [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Distance change across clustering prompts (30 trials each). Significant reduction confirms successful agent [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Distance change across scattering prompts (30 trials each). Increases confirm dispersal responses. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 9 linked inside Pith

  1. [1]

    Self- organization in biological systems

    Scott Camazine, Jean-Louis Deneubourg, Nigel Franks, J Sneyd, Eric Bonabeau, and Guy Theraulaz. Self- organization in biological systems. 2001

  2. [2]

    Open problems in artificial life.Artificial life, 6(4):363–376, 2000

    Mark A Bedau, John S McCaskill, Norman H Packard, Steen Rasmussen, Chris Adami, David G Green, Takashi Ikegami, Kunihiko Kaneko, and Thomas S Ray. Open problems in artificial life.Artificial life, 6(4):363–376, 2000

  3. [3]

    Bioelectric signaling: Reprogrammable circuits underlying embryogenesis, regeneration, and cancer.Cell, 184(8):1971–1989, 2021

    Michael Levin. Bioelectric signaling: Reprogrammable circuits underlying embryogenesis, regeneration, and cancer.Cell, 184(8):1971–1989, 2021. 11 APREPRINT- SEPTEMBER16, 2025

  4. [4]

    Programmable self-assembly in a thousand-robot swarm.Science, 345(6198):795–799, 2014

    Michael Rubenstein, Alejandro Cornejo, and Radhika Nagpal. Programmable self-assembly in a thousand-robot swarm.Science, 345(6198):795–799, 2014

  5. [5]

    Synthetic collective intelligence.Biosystems, 148:47–61, 2016

    Ricard Solé, Daniel R Amor, Salva Duran-Nebreda, Núria Conde-Pueyo, Max Carbonell-Ballestero, and Raul Montañez. Synthetic collective intelligence.Biosystems, 148:47–61, 2016

  6. [6]

    Word2vec to behavior: morphology facilitates the grounding of language in machines

    David Matthews, Sam Kriegman, Collin Cappelle, and Josh Bongard. Word2vec to behavior: morphology facilitates the grounding of language in machines. In2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4153–4160. IEEE, 2019

  7. [7]

    Large language models for robotics: A survey.arXiv preprint arXiv:2311.07226, 2023

    Fanlong Zeng, Wensheng Gan, Yongheng Wang, Ning Liu, and Philip S Yu. Large language models for robotics: A survey.arXiv preprint arXiv:2311.07226, 2023

  8. [8]

    Top-down models in biology: explanation and control of complex living systems above the molecular level.Journal of The Royal Society Interface, 13(124):20160555, 2016

    Giovanni Pezzulo and Michael Levin. Top-down models in biology: explanation and control of complex living systems above the molecular level.Journal of The Royal Society Interface, 13(124):20160555, 2016

Show all 38 references
  1. [9]

    A text-guided protein design framework.Nature Machine Intelligence, 7(4):580–591, March 2025

    Shengchao Liu, Yanjing Li, Zhuoxinran Li, Anthony Gitter, Yutao Zhu, Jiarui Lu, Zhao Xu, Weili Nie, Arvind Ramanathan, Chaowei Xiao, Jian Tang, Hongyu Guo, and Anima Anandkumar. A text-guided protein design framework.Nature Machine Intelligence, 7(4):580–591, March 2025

  2. [10]

    Molformer: Motif-based transformer on 3d heterogeneous molecular graphs

    Fang Wu, Dragomir Radev, and Stan Z Li. Molformer: Motif-based transformer on 3d heterogeneous molecular graphs. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 5312–5320, 2023

  3. [11]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

  4. [12]

    Automating the search for artificial life with foundation models.arXiv preprint arXiv:2412.17799, 2024

    Akarsh Kumar, Chris Lu, Louis Kirsch, Yujin Tang, Kenneth O Stanley, Phillip Isola, and David Ha. Automating the search for artificial life with foundation models.arXiv preprint arXiv:2412.17799, 2024

  5. [13]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

  6. [14]

    Oxford university press, 2009

    Melanie Mitchell.Complexity: A guided tour. Oxford university press, 2009

  7. [15]

    Artificial life: An overview

    Christopher G Langton. Artificial life: An overview. 1997

  8. [16]

    WW Norton & Company, 2011

    Terrence W Deacon.Incomplete nature: How mind emerged from matter. WW Norton & Company, 2011

  9. [17]

    Being emergence vs

    Jason Winning and William Bechtel. Being emergence vs. pattern emergence: Complexity, control and goal- directedness in biological systems. InThe Routledge handbook of emergence, pages 134–144. Routledge, 2019

  10. [18]

    Morphogenetic fields in embryogenesis, regeneration, and cancer: Non-local control of complex patterning.Biosystems, 109(3):243–261, September 2012

    Michael Levin. Morphogenetic fields in embryogenesis, regeneration, and cancer: Non-local control of complex patterning.Biosystems, 109(3):243–261, September 2012

  11. [19]

    Adams and Michael Levin

    Dany S. Adams and Michael Levin. Endogenous voltage gradients as mediators of cell-cell communication: strategies for investigating bioelectrical signals during pattern formation.Cell and Tissue Research, 352(1):95–122, February 2012

  12. [20]

    Elowitz and Stanislas Leibler

    Michael B. Elowitz and Stanislas Leibler. A synthetic oscillatory network of transcriptional regulators.Nature, 403(6767):335–338, January 2000

  13. [21]

    Ewen Cameron, Caleb J

    D. Ewen Cameron, Caleb J. Bashor, and James J. Collins. A brief history of synthetic biology.Nature Reviews Microbiology, 12(5):381–390, April 2014

  14. [22]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018

  15. [23]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langua...

  16. [24]

    Language to rewards for robotic skill synthesis.Arxiv preprint arXiv:2306.08647, 2023

    Wenhao Yu and Gileadi et al. Language to rewards for robotic skill synthesis.Arxiv preprint arXiv:2306.08647, 2023

  17. [25]

    Do as i can and not as i say: Grounding language in robotic affordances

    Michael Ahn, Anthony Brohan, and Noah Brown et al. Do as i can and not as i say: Grounding language in robotic affordances. InarXiv preprint arXiv:2204.01691, 2022

  18. [26]

    Lamda: Language models for dialog applications.arXiv preprint arXiv:2201.08239, 2022

    Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications.arXiv preprint arXiv:2201.08239, 2022. 12 APREPRINT- SEPTEMBER16, 2025

  19. [27]

    Evaluating large language models trained on code

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021

  20. [28]

    Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36:8634–8652, 2023

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36:8634–8652, 2023

  21. [29]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023

  22. [30]

    From text to life: On the reciprocal relationship between artificial life and large language models

    Eleni Nisioti, Claire Glanois, Elias Najarro, Andrew Dai, Elliot Meyerson, Joachim Winther Pedersen, Laetitia Teodorescu, Conor F Hayes, Shyam Sudhakaran, and Sebastian Risi. From text to life: On the reciprocal relationship between artificial life and large language models. I...

  23. [31]

    Lifegpt: Topology-agnostic generative pretrained transformer model for cellular automata.arXiv preprint arXiv:2409.12182, 2024

    Jaime A Berkovich and Markus J Buehler. Lifegpt: Topology-agnostic generative pretrained transformer model for cellular automata.arXiv preprint arXiv:2409.12182, 2024

  24. [32]

    Le, Patrick Erikson, Yanbo Zhang, Michael Levin, and Josh Bongard

    Nam H. Le, Patrick Erikson, Yanbo Zhang, Michael Levin, and Josh Bongard. Giving simulated cells a voice: Evolving prompt-to-intervention models for cellular control, 2025

  25. [33]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019

  26. [34]

    Mistral-small-3.1-24b-instruct-2503

    Mistral AI. Mistral-small-3.1-24b-instruct-2503. https://huggingface.co/mistralai/Mistral-Small-3. 1-24B-Instruct-2503, 2024. Accessed: 2025-05-06

  27. [35]

    Evolution strategies–a comprehensive introduction.Natural computing, 1:3–52, 2002

    Hans-Georg Beyer and Hans-Paul Schwefel. Evolution strategies–a comprehensive introduction.Natural computing, 1:3–52, 2002

  28. [36]

    Difftaichi: Differentiable programming for physical simulation.arXiv preprint arXiv:1910.00935, 2019

    Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation.arXiv preprint arXiv:1910.00935, 2019

  29. [37]

    The symbol grounding problem.Physica D: Nonlinear Phenomena, 42(1-3):335–346, 1990

    Stevan Harnad. The symbol grounding problem.Physica D: Nonlinear Phenomena, 42(1-3):335–346, 1990

  30. [38]

    Biological autonomy.A philo, 2015

    Alvaro Moreno and Matteo Mossio. Biological autonomy.A philo, 2015. 13 APREPRINT- SEPTEMBER16, 2025 Figure 10: Distance change across clustering prompts (30 trials each). Significant reduction confirms successful agent convergence. 14 APREPRINT- SEPTEMBER16, 2025 Figure 11: Di...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.