ASAL++

Guiding Evolution of Artificial Life Using Vision-Language Models

Nikhil Baid · Hannah Erlebach · Paul Hellegouarch · Frederico Wieser

Paper · arXiv · ASAL paper · ASAL project · Citation

Published in ALIFE 2025: Ciphers of Life — Proceedings of the Artificial Life Conference 2025 (MIT Press).

Overview

ASAL++ extends the supervised-target mechanism in Automated Search for Artificial Life (ASAL) from fixed, human-written objectives to a sequence of targets proposed by a multimodal foundation model. An inner ASAL loop uses CLIP and Sep-CMA-ES to optimise an artificial-life simulation towards a natural-language target. An outer Evolver Model observes the resulting video, proposes the next target, and warm-starts the next search from the previous best configuration.

The paper studies two variants in Lenia using Gemma-3-4b-it:

• Evolved Supervised Targets (EST): optimise only the latest generated target. EST produces larger morphological changes and higher visual novelty, but weaker continuity.
• Evolved Temporal Targets (ETT): optimise the complete target history at sampled rollout times. ETT produces more coherent and interpretable trajectories, but can stagnate in local minima.

The repository contains the paper experiments, ablations, branching “trees of life”, preserved outputs, exploratory notebooks, a pinned JAX ASAL fork, and an experimental PyTorch Lenia implementation.

Contributions

Component	ASAL baseline	ASAL++
Target source	Fixed prompts supplied before search	A multimodal model proposes targets from rollout history
Search structure	One supervised, open-endedness, or illumination search	Repeated warm-started ASAL searches in an autonomous outer loop
Temporal control	Human-specified target sequence	EST and ETT generated target trajectories
Exploration	Supervised-target, open-endedness, and illumination searches	Iterated target evolution, branching trees, and environmental prompting
Implementation	JAX reference implementation	JAX experiments plus experimental PyTorch Lenia/model adapters

Method

At outer iteration i:

1. Optimise substrate parameters theta_i against either the latest target (EST) or the target history (ETT).
2. Render the best simulation as a video rollout.
3. Sample up to --max_images rollout frames and pass them, with the relevant prompt history, to the Evolver Model.
4. Generate the next target p_(i+1).
5. Warm-start the next Sep-CMA-ES search from theta_i and repeat.

The method is agnostic to the underlying artificial-life substrate provided that it exposes a parameterised simulation and a 2D rendering. The reported experiments focus on Lenia. The pinned JAX fork also contains inherited ASAL implementations for Boids, Particle Life, Particle Life++, Particle Lenia, discrete and continuous neural cellular automata, and Life-like cellular automata.

See docs/architecture.md for the formal objects, interfaces, and code boundaries.

Results

Paper-scale experiments use nine initial prompts, eight outer-loop iterations, 2,000 inner Sep-CMA-ES steps per iteration, CLIP for visual/text alignment, and Gemma-3-4b-it as the Evolver Model.

Method	Mean visual OE score	Final − initial OE	Qualitative outcome
EST	0.052 ± 0.002	0.009 ± 0.004	Greater visual novelty; limited long-range continuity
ETT	0.049 ± 0.002	0.002 ± 0.004	More coherent trajectories; occasional repetition and convergence

Aggregate ± values are standard errors across the nine prompt-conditioned runs; each run’s OE summary uses the standard deviation over its eight ASAL++ iterations.

The open-endedness score is 1 − max cosine similarity between a sampled frame and its earlier frames in CLIP image-embedding space. Higher is more visually novel. The results support open-ended characteristics, not indefinite open-ended evolution; the paper discusses model bias, Lenia’s radial symmetry, compute cost, vanishing activations, and limited quantitative benchmarks.

Selected ETT trajectories

The labels below are the human-written seed prompts, not claims that the final morphology literally depicts the phrase.

a flower	a monkey	an extraterrestrial life

Installation

Requirements

• Python 3.11 and uv;
• Git LFS and Git submodule support;
• a CUDA-capable GPU for practical paper-scale runs;
• Hugging Face access to google/gemma-3-4b-it;
• FFmpeg for video generation; Graphviz/Pydot for graph rendering.

git clone --recurse-submodules https://github.com/fredericowieser/ASALPlusPlus.git
cd ASALPlusPlus
bash scripts/setup.sh
source .venv/bin/activate
huggingface-cli login

The setup script installs the research environment from configs/requirements.txt, installs the src/ packages in editable mode, and selects a CUDA-enabled JAX build when nvcc is available.

A dependency-free structural check is available before downloading model weights:

python scripts/verify_repository.py

Quick start

The commands below reproduce the structure of one paper run. They download large foundation-model checkpoints and are not lightweight smoke tests.

# EST: optimise only the newest generated target
python scripts/main_evolutionary_prompts_last_only.py \
  --save_dir analysis/data/runs/est-microbe \
  --substrate lenia \
  --prompts "a microbe" \
  --coef_prompt 1.0 \
  --N 8 \
  --n_iters 2000 \
  --temp 0.1

# ETT: optimise the accumulated temporal target sequence
python scripts/main_evolutionary_prompts_temporal.py \
  --save_dir analysis/data/runs/ett-microbe \
  --substrate lenia \
  --prompts "a microbe" \
  --coef_prompt 1.0 \
  --coef_softmax 0.5 \
  --N 8 \
  --n_iters 2000 \
  --temp 0.1

For paper-scale sweeps and branching experiments, see docs/experiments.md.

Repository layout

.
├── src/                    # JAX submodule, PyTorch implementation, shared utilities
├── scripts/                # search, ablation, analysis, rendering, and setup entry points
├── configs/                # W&B sweeps, dependency list, and prompt sets
├── analysis/
│   ├── data/               # preserved runs, scores, summaries, and checkpoints
│   └── graphs/             # branching trajectories and graph visualisations
├── notebooks/              # tutorials and exploratory notebooks (Git LFS)
├── docs/                   # architecture, experiment map, and reference material
├── assets/                 # method figures and curated README animations
├── pyproject.toml
├── CITATION.cff
└── README.md

The committed analysis artefacts are large. Historical configuration snapshots retain the paths recorded during the original runs; current entry points use analysis/data/ and analysis/graphs/. See analysis/README.md for provenance notes.

Implementation status

• Canonical paper path: JAX ASAL inner loop in src/asal/ with ASAL++ drivers in scripts/.
• PyTorch path: experimental Lenia-only implementation in src/asal_pytorch/. CLIP and Gemma-3 are registered; BLIP-2, CLIP4Clip, and VideoCLIP remain source-level adapters requiring further integration.
• Notebooks: exploratory research records; some retain Colab-specific cells and cached outputs.
• FlowLenia sweep: retained for provenance, but the currently pinned submodule does not expose the flenia identifier required by that historical configuration.

Citation

@inproceedings{baid2025guiding,
  title     = {Guiding Evolution of Artificial Life Using Vision-Language Models},
  author    = {Baid, Nikhil and Erlebach, Hannah and Hellegouarch, Paul and Wieser, Frederico},
  booktitle = {ALIFE 2025: Ciphers of Life: Proceedings of the Artificial Life Conference 2025},
  publisher = {MIT Press},
  year      = {2025},
  doi       = {10.1162/isal.a.850}
}

This work builds on ASAL:

@article{kumar2024automating,
  title  = {Automating the Search for Artificial Life with Foundation Models},
  author = {Kumar, Akarsh and Lu, Chris and Kirsch, Louis and Tang, Yujin and Stanley, Kenneth O. and Isola, Phillip and Ha, David},
  year   = {2024},
  url    = {https://arxiv.org/abs/2412.17799}
}

Acknowledgements and licence status

We thank Laura Ruis for supervision and guidance, Akarsh Kumar for feedback, and Catherine Bacon for editing. The JAX implementation under src/asal/ is a pinned fork of the ASAL repository and retains its upstream licence.

The surrounding ASAL++ repository does not currently declare a top-level software licence. The paper is available from arXiv under CC BY 4.0; that paper licence does not automatically license the code.