Fluidworld Reaction Diffusion Models

FluidWorld: Reaction-Diffusion Dynamics for World Modeling

Prior Belief Challenged

Assumption: Self-attention (Transformers) is necessary for effective predictive world modeling. Attention mechanisms capture long-range dependencies and enable flexible computation.

Falsifying Experiment: PDE-Based World Models

Train three architecturally-comparable models (all ~800K parameters) on video prediction (UCF-101):

• FluidWorld: Multi-scale reaction-diffusion PDEs
• Transformer Baseline: Standard self-attention world model
• ConvLSTM Baseline: Recurrent architecture

Parameter Budget Matching: Crucial—all three models have identical parameter counts to isolate architectural differences.

Key Results:

Single-Step Metrics (Converge):

• MSE reconstruction: FluidWorld 2× lower than Transformer (0.001 vs 0.002)
• Spatial activation statistics: Comparable

Multi-Step Rollouts (Diverge Dramatically):

• At horizon h=2: Transformer and ConvLSTM outputs collapse into gray frames
• At horizon h=3: FluidWorld maintains recognizable structure; competitors unrecoverable
• SSIM trajectory: FluidWorld shows measurable recovery cycles (66.8% of rollouts show recovery, p<10⁻⁴⁹)

Critical Finding: Single-step optimization masks architectural differences; only extended autoregressive rollouts reveal advantages.

Revised Principle: Aperiodic Spatial Processing > Dense Attention

Core Mechanism: Reaction-diffusion PDEs provide three properties absent from attention:

1. O(N) Locality: Multi-scale Laplacian diffusion operates on spatially-local neighborhoods, avoiding O(N²) pairwise interactions.

2. Implicit Error Correction: Laplacian smoothing (∇²) acts as low-pass filter, attenuating prediction errors during rollouts. When accumulated error reaches high frequencies, diffusion progressively removes it.

3. Graceful Degradation: System operates at edge-of-chaos criticality—50% state corruption recovers autonomously without explicit recovery mechanisms.

Mathematical Basis:

Reaction-diffusion equation combines:

• Multi-scale Laplacian: ∇² applied with dilations {1, 4, 16} for long-range diffusion
• Position-wise reaction: MLP processes local dynamics
• Global memory terms: Accumulate context across time steps
• BeliefField persistent state: Tracks temporal context via gated PDE evolution

# Reaction-diffusion forward pass: iterative PDE integration
# For each timestep, apply:
# u_new = u + dt * (
#     diffusion_1*Laplacian(u) +
#     diffusion_4*Laplacian_dilated4(u) +
#     diffusion_16*Laplacian_dilated16(u) +
#     MLP(u, global_context) +  # reaction term
#     gating * memory_update
# )
# Repeat for adaptive steps (3-12) until convergence or timeout

Implications for Architecture Design

1. Spatial Inductive Bias Works: Locality constraints improve long-horizon stability; dense attention's flexibility isn't necessary for video prediction.

2. Energy-Based Thinking: PDEs naturally formulate as energy minimization; reward functions implicitly build in diffusion-based regularization.

3. Bio-Inspired Mechanisms: Lateral inhibition, synaptic fatigue, Hebbian diffusion not just biologically plausible—functionally critical for stability.

4. Computational Efficiency Trade-off: FluidWorld slower per-step (~1 it/s vs 5.2 it/s for Transformer) but converges quickly; long-horizon planning breaks Transformer's advantage.

Opened Research Directions

• JEPA-Style Latent Prediction: Early results (0.827 cosine similarity at step 19) suggest PDE strength extends to latent-space world models
• Action-Conditioned Control: Current work unconditional; architecture ready for control-signal integration
• Higher Resolution: Current 64×64; PDE advantages should amplify at 128×128+ due to O(N) vs O(N²) scaling
• Multi-Modal Fusion: Memory terms can integrate proprioception, action history; preliminary results promising
• Quantitative Rollout Metrics: FVD, LPIPS needed for standardized evaluation beyond SSIM
• Bio-Mechanism Ablation: Individual contributions of lateral inhibition, fatigue, Hebbian dynamics unclear; systematic studies planned

Deployment Considerations

• Training Cost: 5-8× slower than Transformer baseline; suitable for offline training, not interactive fine-tuning
• Inference Speedup Potential: Adaptive early stopping and reduced diffusion timesteps under investigation
• Dataset Scale: Evaluated on UCF-101 (small); scaling to larger datasets unknown
• Generalization: Single-dataset validation; cross-dataset rollouts need quantification before claiming general advantage