Main Page - Publication Page

Paper PDF

Code

Summary

This paper introduces Automated Search for Artificial Life (ASAL), a novel framework that leverages foundation models (FMs), particularly vision-language FMs like CLIP, to automate the discovery and analysis of Artificial Life (ALife) simulations. Historically, ALife research has been hampered by the reliance on manual design and trial-and-error to explore the vast combinatorial space of possible simulation configurations. ASAL addresses this by enabling three distinct search mechanisms:

• Supervised Target Search: This method allows researchers to find simulations that generate specific, predefined phenomena. By providing textual prompts (e.g., "a red oscillating blob") at different simulation timesteps, ASAL guides the search to discover configurations that produce the desired visual outcomes. This is crucial for identifying worlds similar to our own or testing hypothetical evolutionary scenarios.

• Open-Endedness Search: A significant challenge in ALife is discovering simulations that exhibit temporally open-ended novelty, mirroring the continuous innovation observed in natural evolution. ASAL tackles this by searching for simulations whose trajectories in the FM's representation space are historically novel, meaning they deviate significantly from their past states. This approach outsources the subjective assessment of "interestingness" to the rich representations learned by FMs.

• Illumination Search: To map out the diverse landscape of emergent behaviors within a substrate, ASAL employs an illumination strategy. It searches for a set of simulations that are visually distinct from each other in the FM's representation space, effectively creating a "simulation atlas" that showcases the breadth of possibilities within a given ALife system.

ASAL's key strengths lie in its generality and its ability to handle diverse ALife substrates, including Boids, Particle Life, Game of Life, Lenia, and Neural Cellular Automata, provided they can be rendered as images. The framework not only facilitates the discovery of novel ALife forms and behaviors—such as previously unseen Lenia organisms, exotic Boids flocking patterns, and open-ended cellular automata—but also enables the quantitative analysis of qualitative phenomena. By using FM embeddings, ASAL can provide human-aligned metrics for complexity, diversity, and novelty, moving beyond traditional, often inadequate, quantitative measures. The paper demonstrates that ASAL is agnostic to the specific FM used, with vision-language models like CLIP and vision-only models like DINOv2 proving effective, and highlights the superiority of deep FM representations over raw pixel-based comparisons for capturing human notions of diversity. This paradigm shift promises to significantly accelerate ALife research by automating the exploration of "life as it could be."

Akarsh Kumar MIT Sakana AI, Chris Lu OpenAI, Louis Kirsch The Swiss AI Lab IDSIA, Yujin Tang Sakana AI, Kenneth O. Stanley Independent, Phillip Isola MIT, David Ha Sakana AI

Abstract

With the recent Nobel Prize awarded for radical advances in protein discovery, foundation models (FMs) for exploring large combinatorial spaces promise to revolutionize many scientific fields. Artificial Life (ALife) has not yet integrated FMs, thus presenting a major opportunity for the field to alleviate the historical burden of relying chiefly on manual design and trial-and-error to discover the configurations of lifelike simulations. This paper presents, for the first time, a successful realization of this opportunity using vision-language FMs. The proposed approach, called Automated Search for Artificial Life (ASAL), (1) finds simulations that produce target phenomena, (2) discovers simulations that generate temporally open-ended novelty, and (3) illuminates an entire space of interestingly diverse simulations. Because of the generality of FMs, ASAL works effectively across a diverse range of ALife substrates including Boids, Particle Life, Game of Life, Lenia, and Neural Cellular Automata. A major result highlighting the potential of this technique is the discovery of previously unseen Lenia and Boids lifeforms, as well as cellular automata that are open-ended like Conway’s Game of Life. Additionally, the use of FMs allows for the quantification of previously qualitative phenomena in a human-aligned way. This new paradigm promises to accelerate ALife research beyond what is possible through human ingenuity alone.

Conclusion

Summary

This project launches a new paradigm in ALife by taking the first step towards using FMs to automate the search for interesting simulations. Our approach is effective in finding target, open-ended, and diverse simulations over a wide spectrum of substrates. Additionally, FMs enable the quantification of many qualitative phenomena in ALife, offering a path to replacing low-level complexity metrics with deep representations aligned with humans.

Discussion

Because this project is agnostic to the FM and substrate used, it raises the question of which ones to use. The choice of FMs seems to not matter much from our experiments, and FMs in general may also be converging to similar representations of reality (Huh et al., 2024). The proper substrate largely depends on the phenomena that is being studied (e.g. self-organization, open-ended evolution, etc.). The most expressive substrate would simply parameterize all the RGB pixels of an entire video, but is useless for studying emergence. The most insightful substrates bake in as little information as possible, while maintaining vast emergent capabilities. For example, the periodic table of elements can be defined with little information, yet gives rise to the entirety of the observable universe.

Eventually, with the proper substrate, more powerful FMs, and enough compute, this paradigm may allow researchers to automatically search for worlds which start off as “simple cells in primordial soup”, then undergo “a Cambrian explosion of complexity”, and eventually become “an artificial alien civilization.” Researchers could alternatively search for hypothetical worlds where life evolves without DNA. Finding open-ended worlds would solve one of ALife’s grand challenges (Bedau et al., 2000; Stanley et al., 2017). Illuminating such a substrate could help map the space of possible lifeforms and intelligences, giving a taxonomy of life as it could be in the computational universe.

This work can be generalized by replacing the image-language FM with video-language FMs that natively process the temporal nature of simulations (Tang et al., 2023; Xu et al., 2021) or with 3-D FMs to handle 3-D simulations. To leverage the recent advances of LLMs, images can be converted to text via image-to-text models, allowing all analyses to be done in text space.

Instead of searching for ALife simulations, a similar approach could be constructed for low-level physics research. For example, in Wolfram’s Physics Project (Wolfram, 2020), one could search for the hypergraph update rule which emerges structures that an FM considers natural. At a meta-level, LLMs could be useful for generating code that describes the substrates themselves, driven by higher-level research agendas, similar to Faldor et al. (2024); Lu et al. (2024b, c).