Paper-Grounded Lanes

This dossier now treats the major paper reproductions as first-class research lanes, not as ad hoc sweeps.

The lane model exists for one reason: each paper asks a different question, uses a different search space, and evaluates success differently. A single generic search surface cannot represent all of them without becoming vague or misleading.

Main Surfaces

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 lane
sensorimotor-2024 for the 2024 agency-discovery lane
qd-2024 for the 2024 quality-diversity lane
ecology-2025 for the 2025 Flow-Lenia intrinsic-evolution ecology lane
curiosity-2025 for the 2025 curiosity-driven universe exploration lane
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. The config directories keep the paper names because they are provenance surfaces, not the main CLI vocabulary.

Source Papers

Why These Lanes Exist

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.

Those are different research programs. They should share core simulation code when possible, but they should not be flattened into one search API.

Lane 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

Why it is useful:

It gives a reproducible task harness for motion and navigation.
It lets us debug optimization and morphology separately from open-ended discovery.

Main remaining 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do not include every original experiment wrapper, so parity is against the published paper/config regime rather than unpublished internal scripts.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 harnesses around the reaction-diffusion interpretation
asymptotic/original comparison outputs

Why it is useful:

It is the core-validation lane.
It checks whether our continuous Lenia-like dynamics remain mathematically defensible rather than only visually plausible.

Main remaining gap:

This lane is a validator, not a discovery engine. It should stay small and sharp.

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

Why it is useful:

This is the lane closest to our earlier “motion without convincing organisms” failure mode.
It gives a principled discovery loop instead of score-only sweeps.

Main remaining gap:

Real paper-scale runs still matter more than more harness code. This lane 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

Why it is useful:

This lane discovers repertoires, not one winner.
It is the right tool for broad species discovery when we care about stepping stones and diversity.

Main remaining 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

Why it is useful:

It moves us from isolated specimens to ecosystem behavior.
It is the correct lane for long-run multispecies questions.

Main remaining gap:

This lane should be judged by long-horizon ecology outputs, not by 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

Why it is useful:

It is the open-ended universe discovery lane.
It replaces narrow creature ranking with system-level novelty metrics.

Main remaining 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 lane
public-style polygon-library input
paper-style batch semantics and phase-map export

Why it is useful:

It maps the search space before we try to mine it for gliders or creatures.
It is the cleanest correction to blind classical Lenia sweeps.

Main remaining gap:

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

Shared Engineering Choices

These lanes 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 lanes are intended to run on the MLX/Metal path on Apple Silicon.

Relevant code:

What has already been done:

release-path runs for long paper lanes
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

The paper lanes are allowed to add runners, descriptors, and evaluation logic. They are not allowed to silently rewrite the core Flow-Lenia equations.

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

When a lane needs a paper-specific surface, the correct move is:

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

What To Read Next

For the current architecture:

For operating these lanes: