COMMUNITY_COLLABORATION.md

Community Collaboration: Scaling Spectral World Models

Invitation

We are seeking research and engineering collaborators to help scale and independently test the Spectral World Models (SWM) project.

SWM investigates a world-model architecture in which observations are represented in multiresolution spectral coordinates and future latent states are evolved through action-conditioned structured operators. The project has completed a falsification-oriented experimental sequence through V15.1. Rather than presenting the current implementation as a finished or universally superior architecture, the results identify both useful properties and concrete limitations.

The current release is intentionally frozen so that the next stage can focus on independent scaling, replication, and new research hypotheses, rather than continued tuning against the existing benchmarks.

Why Collaborate?

The completed experiments suggest that structured spectral dynamics are scientifically interesting, but substantially larger-scale evaluation is needed.

Current evidence indicates that:

• structured operator constraints can provide controlled and stable latent rollouts;
• counterfactual trajectory training can improve directional action-effect alignment on controlled synthetic dynamics;
• symmetry audits reveal learned coordinate anisotropy that aggregate metrics can hide;
• the frozen V9 transition/training recipe does not transfer competitively to the tested visual Classic Control environments;
• dynamics-sensitive V15.1 diagnostics show that this external-transfer deficit is not explained solely by static-background bias in full-frame pixel MSE.

These results make SWM a useful starting point for researchers interested in spectral representations, operator learning, world models, causal dynamics, symmetry/equivariance, long-horizon prediction, and model-based planning.

The most valuable next experiments require more compute, broader environments, and independent implementations than the initial project can provide alone.

Collaborators We Are Looking For

We are particularly interested in individuals, laboratories, startups, and organizations with access to substantial GPU compute or relevant simulation infrastructure.

Potential collaboration areas include:

1. Large-scale training — increase model capacity, spectral-state dimension, operator rank, dataset size, rollout horizon, and environment diversity, with an emphasis on studying scaling behavior rather than simply tuning the existing benchmarks.
2. Independent replication — reproduce the V6–V15.1 results from the frozen artifact using independent seeds, hardware, and environments, reporting both positive and negative replications.
3. Symmetry-aware spectral dynamics — investigate whether explicitly equivariant spectral transitions can address the axis anisotropy exposed by V13.
4. Richer nonlinear operator models — explore nonlinear spectral transitions while retaining interpretable stability controls and compare against Koopman-inspired, neural-operator, state-space, transformer, and hybrid approaches.
5. Larger world-model benchmarks — evaluate on established visual-control, robotics, physics, video-prediction, or embodied-agent benchmarks.
6. Planning and control — extend evaluation beyond predictive rollouts to closed-loop planning and determine whether spectral structure provides planning advantages even when raw prediction error is not state of the art.
7. Distributed and accelerated implementations — multi-GPU/multi-node training, mixed precision, memory-efficient spectral operators, profiling, and optimized kernels.

Compute Partners

A particularly useful collaboration would involve a group able to provide sustained access to modern GPU clusters and engineering support.

Rather than merely rerunning the current configuration with more epochs, we would like to investigate genuine scaling questions:

• How does performance change as spectral-state dimension and operator rank increase?
• Do the stability properties survive at substantially larger parameter counts?
• Can symmetry-aware transitions eliminate the coordinate anisotropy observed in V13?
• Does broader multi-environment training improve external dynamics transfer?
• At what scale, if any, do structured spectral transitions become competitive with strong latent-dynamics and neural-operator baselines?
• Are there regimes in which spectral structure improves planning efficiency or long-horizon robustness even when one-step prediction is comparable?

Negative results are valuable. The objective is to determine where the approach works and where it does not.

Contracting and Sponsored Collaboration

We are also open to contract research, sponsored experimentation, and engineering collaborations with organizations interested in evaluating or extending this work.

Potential engagements include:

• funded benchmark and scaling studies;
• implementation and optimization of SWM variants on partner infrastructure;
• custom evaluations on proprietary simulation or control problems;
• development of symmetry-aware or domain-specific spectral world models;
• joint research prototypes;
• technical consulting around spectral/operator approaches to learned dynamics;
• jointly authored research where contributions warrant authorship.

Any commercial, sponsored, or contracting engagement would be established separately in writing, including scope, deliverables, funding, intellectual-property terms, publication expectations, confidentiality requirements, and ownership of newly developed work.

The open-source repository itself should not be interpreted as creating a commercial partnership or contractual obligation.

Reproducibility Baseline

Prospective collaborators should begin from the frozen reproducibility release rather than an undocumented development branch.

Project repository:
https://huggingface.co/kiruluta/Spectral-World-Models-Reproducibility

Frozen experimental release: v15.1-final

The release contains the implementation, test suite, experimental lineage, measured result tables, and diagnostics through V15.1.

A clean-environment validation of the final package completed with:

32 passed

V15.1 is the endpoint of the original experimental sequence. Proposed architectural changes should therefore be identified as new experiments or successor research, rather than retroactively modifying the frozen V15.1 result.

Suggested Collaboration Workflow

Stage 1 — Reproduce. Validate the frozen release on independent hardware and reproduce selected benchmark results.

Stage 2 — Profile. Identify compute, memory, data, optimization, and architectural bottlenecks.

Stage 3 — Scale or test a new hypothesis. Define the experiment before running it, including baselines, evaluation metrics, compute budget, stopping criteria, and falsification conditions.

Stage 4 — Report. Contribute reproducible configurations, logs, result tables, and analysis. Positive and negative results should both be preserved.

This structure is intended to reduce benchmark-specific post-hoc tuning and make collaborative results scientifically interpretable.

What to Include When Reaching Out

Please include:

• your name and affiliation or organization;
• the aspect of SWM that interests you;
• relevant research or engineering experience;
• available compute or simulation resources, if applicable;
• the experiment or extension you would like to pursue;
• whether you are proposing an open research collaboration, compute sponsorship, sponsored research, or a contracting engagement;
• links to relevant publications, repositories, or previous work when available.

For substantial experiments, a short proposed experiment specification is preferable to a general expression of interest.

Collaboration Principles

This project welcomes rigorous attempts to improve, reproduce, or falsify the approach. We especially value collaborations that predefine meaningful comparisons and stopping criteria, preserve strong baselines, report variance across seeds, distinguish mathematical stability from predictive or causal correctness, test symmetry and out-of-distribution behavior explicitly, publish negative findings alongside positive ones, and keep the frozen V15.1 evidence separate from successor architectures.

Citation

If you build directly on the frozen implementation or use its experimental results, please cite the V15.1 reproducibility artifact and accompanying manuscript.

Spectral World Models: V15.1 Final Reproducibility Artifact
Andrew J. Kiruluta, 2026
Hugging Face repository: https://huggingface.co/kiruluta/Spectral-World-Models-Reproducibility
Revision: v15.1-final

Contact / Discussion

Interested collaborators can open a discussion in the Hugging Face repository or contact the project author through the contact mechanism listed in the repository/profile.

A useful discussion title is:

Collaboration proposal: [short description of experiment]

Please indicate whether the proposal concerns independent replication, open research collaboration, compute sponsorship, funded/contract research, or engineering/infrastructure collaboration.

---

Status: Open to collaboration.

The objective of the next phase is not simply to make the existing benchmark numbers larger. It is to determine, with substantially greater compute and independent scrutiny, whether spectral/operator world models can scale into a useful foundation for learned dynamics—and to identify clearly the regimes in which they cannot.