EcoAI explores geometry-driven world models that prioritize structure and latent dynamics over brute-force scaling, with a deliberate focus on compact and interpretable architectures.
Geometry first. Parameters second. Scale last.
EcoAI investigates world models under strict capacity constraints, emphasizing latent structure, geometric organization, and temporal dynamics as primary drivers of intelligence. Rather than relying on scale, the research explores how architectural inductive bias can support prediction, reasoning, and planning in compact systems.
We study sub-million-parameter world models trained entirely from scratch on TinyImageNet, using benchmark performance as a probe for latent structure quality rather than an optimization objective. Our results show that architectural design alone can substantially improve the efficiency–performance tradeoff.
These findings suggest geometry-driven organization as a promising direction for interpretable, resource-efficient world modeling beyond brute-force parameter scaling.
EcoAI architectures decompose intelligence into perception, structured memory, latent world dynamics, and decision-oriented planning. The objective is not to generate pixels, but to model how the world evolves in latent space.
Latent spaces are treated as structured manifolds rather than flat embeddings. Distances, directions, and trajectories carry semantic meaning. Understanding emerges from geometry, not supervision.
Lightweight persistence mechanisms enable stable internal state without exposing internal mechanics.
A compact output layer translates structured representations into predictions with minimal overhead.
Two compact world models trained from-scratch demonstrate the potential of geometry-first design.
~430K parameters — baseline compact design focused on core structure.
~580K parameters — enhanced stability and representation consistency.
We scale structure, not parameter count. The goal: accessible, reproducible progress toward compact world models.
The diagram below presents the core architectural pipeline of EcoAI_VN–WM–V3, a geometry-driven world model designed for structured understanding, imagination, and planning under strict capacity constraints.
Figure. EcoAI_VN–WM–V3 decomposes intelligence into perception, latent world dynamics, structured representation, and imagination-driven reasoning. Geometry acts as the organizing principle across all stages.
All benchmark results are reported on the TinyImageNet dataset (64×64), with models trained entirely from scratch and without external pretraining or teacher supervision.
Benchmarks are treated as probes rather than objectives. Performance is interpreted as evidence of latent structure quality, transferability, and stability — not as an end goal in itself.
Top‑1: 42% · Top‑5: 72%
Top‑1: 45% · Top‑5: 82%
These results indicate that strong inductive biases and careful architectural design can close the gap to much larger models on visual benchmarks.
This benchmark evaluates open-loop imagination stability for EcoAI_VN–WM–V3. The model is rolled out for K=100 steps without observation feedback, testing whether coherent world dynamics can be maintained over long horizons.
Model details.
• 578,633 parameters (sub-1M)
• Geometry-driven latent world model
• No decoder supervision during rollout
Despite its compact size, the model maintains structural coherence throughout the imagination horizon. This highlights the role of representation geometry over parameter scale.
The result suggests that long-horizon stability can emerge from carefully structured latent dynamics, not merely from increasing model size.
This benchmark evaluates extreme long-horizon open-loop imagination for EcoAI_VN–WM–V3.1. The model is rolled out for K = 1000 consecutive steps without observation feedback, directly probing the stability of its latent dynamics.
Model details.
• 578,633 parameters (sub-1M)
• Geometry-driven latent world model
• Fully trained from scratch
• No decoder supervision during imagination
Despite the increasing rollout difficulty imposed by curriculum imagination, the model maintains coherent latent trajectories across the full 1000-step horizon, indicating controlled error accumulation rather than open-loop divergence.
Loss increases smoothly with rollout horizon and follows an approximately linear trend in log–log space, suggesting multiplicative but controlled error propagation during long-horizon imagination.
Training remains stable across increasing rollout horizons, with no signs of divergence even as K is extended to 1000.
Overall, these results demonstrate that 1000-step open-loop stability can emerge from carefully structured latent dynamics and geometry-aware representations, without reliance on large-scale pretraining or external supervision.
Next. Building on this stability regime, we transition toward EcoAI_VN–WM–V4, focusing on explicit dynamical metrics such as Lyapunov-style stability analysis.
High-level essays and conceptual notes explain the motivations and philosophical underpinnings of EcoAI. Detailed algorithms remain private to protect intellectual property.
Conceptual overview of how data structure shapes learning and representation.
High-level discussion on stability, structure, and emergent behavior in compact models.
EcoAI is not focused on building larger models, but on understanding how intelligence can emerge from structure, dynamics, and geometry under severe capacity constraints.
For partnerships or research discussions, reach out.
Contact EcoAI