Paper-Grounded Runs

Paper reproductions are separate run families, not ad hoc sweeps.

Each paper asks a different question, searches a different space, and measures success differently. One generic command would hide those differences.

Entrypoints

The current paper-grounded entrypoints are:

evolve with paper configs for the 2022 Flow-Lenia task suite
reaction-diffusion-2023 for the 2023 reaction-diffusion validation work
sensorimotor-2024 for the 2024 agency-discovery work
qd-2024 for the 2024 quality-diversity work
ecology-2025 for the 2025 Flow-Lenia intrinsic-evolution ecology work
curiosity-2025 for the 2025 curiosity-driven universe exploration work
atlas-2026 for the 2026 classical Lenia parameter-space atlas

The old user-facing names leniabreeder-2024, flowlenia-ecology-2025, and ai-scientist-2025 are retired. Config directories keep paper names for provenance, not as CLI vocabulary.

Source Papers

Why These Are Separate

The 2022 paper is task-conditioned optimization, not open-ended discovery.
The 2024 agency paper is goal-conditioned diversity search with a curriculum and robustness battery.
The 2024 QD paper is repertoire discovery with MAP-Elites and AURORA.
The 2025 papers move from isolated organisms to ecosystem- and universe-scale discovery metrics.
The 2026 atlas paper maps parameter space before glider or creature hunting.

These are different experiments. Share core simulation code where possible; do not flatten them into one search API.

Map

Flow-Lenia 2022

Primary code:

What we re-implemented:

directed motion
angular motion
obstacle navigation
chemotaxis
paper-style OpenES loop with antithetic Gaussian noise
paper-locked task configs and output capture

Role:

Reproducible task runner for motion and navigation.
Separates optimization and morphology debugging from open-ended discovery.

Known gap:

Public sources do not include every original experiment wrapper, so parity is against the published paper/config regime rather than unpublished internal scripts.

Reaction-Diffusion 2023

Primary code:

What we re-implemented:

validation and emulation tools around the reaction-diffusion interpretation
asymptotic/original comparison outputs

Role:

Core validation path.
Checks whether the continuous dynamics remain mathematically defensible, rather than only visually plausible.

Known gap:

This is a validator, not a discovery engine. Keep it small.

Sensorimotor 2024

Primary code:

What we re-implemented:

history/archive bootstrap
goal sampling curriculum
source selection and mutation
local optimization toward sampled goals
paper-style evaluation battery for movement, agency, and obstacle robustness

Role:

Closest to the earlier “motion without convincing organisms” failure mode.
Principled discovery loop instead of score-only sweeps.

Known gap:

Real paper-scale runs matter more than more support code. This work now needs broad experiments and result comparison.

QD 2024

Primary code:

What we re-implemented:

MAP-Elites repertoire search around the Aquarium-centered structured genotype space
time-series metrics after a developmental phase
hard viability filtering for empty/full/spread failures
AURORA with a VAE-trained latent descriptor
COM-centered 32x32x3 phenotype crops
8-dimensional latent descriptors
unsupervised descriptor as mean latent trajectory
unsupervised fitness as negative average latent dispersion

Role:

Repertoire discovery, not one-winner search.
Broad species discovery when stepping stones and diversity matter.

Known gap:

The open issue is paper-scale validation, not missing AURORA machinery.

Ecology 2025

Primary code:

What we re-implemented:

paper-locked intrinsic-evolution ecology runs
evolutionary activity and diversity-oriented summaries
strict validation of the paper regime before execution

Role:

Moves the dossier from isolated specimens to ecosystem behavior.
Long-run multispecies work.

Known gap:

Judge this work by long-horizon ecology outputs, not small smoke tests.

Curiosity 2025

Primary code:

What we re-implemented:

curiosity-driven Flow-Lenia universe exploration
paper-specific goal spaces for ecosystem and movement modes
archive/coverage accounting
MP4-size and non-neutral activity style metrics

Role:

Open-ended universe discovery.
Replaces narrow creature ranking with system-level novelty metrics.

Known gap:

As with ecology, the real test is paper-scale exploration, not tiny smoke coverage.

Atlas 2026

Primary code:

What we re-implemented:

classical Lenia mu/sigma atlas
public-style polygon-library input
paper-style batch semantics and phase-map export

Role:

Maps the search space before glider or creature mining.
Corrects blind classical Lenia sweeps.

Known gap:

The remaining work is full atlas scale and result inspection, not another sweep heuristic.

Shared Engineering Choices

These run families share one principle: paper-specific logic lives above the core simulator.

That means:

shared core simulation stays in FlowLenia.swift and related runtime code
paper-specific runners live in Sources/LeniaCore/Layer/
user-facing entrypoints live in Sources/LeniaCLI/
paper configs live under configs/papers

This separation matters because it lets us reproduce papers without silently changing the core equations.

Metal and MLX

The heavy runs use the MLX/Metal path on Apple Silicon.

Relevant code:

What has already been done:

release-path runs for long paper jobs
Instruments/signpost support for hot-path profiling
rollout sync cleanup so metrics do not force unnecessary host/device barriers
external artifact/log routing for long runs on remote volumes

The current hotspot conclusion is still the same:

reintegrationBatched remains the strongest custom-kernel candidate if we need another serious speed step

Invariants

Paper-specific commands can add runners, descriptors, and evaluation logic. They must not silently rewrite the core Flow-Lenia equations.

The guardrail is the invariant test in LeniaCoreTests.swift, especially testPopulationKernelsMatchSingleKernelPath.

When paper work needs a new command, the rule is:

add it in a paper runner
prove it with focused tests
keep the core path numerically consistent

Reading Order

For the current architecture:

For operating these runs: