EmbeddingGemma-300M Tetris System-1 Decision Engine
This model is a real-time System 1 decision engine built on top of google/embeddinggemma-300m, designed to evaluate and select optimal 2D Tetris placements without generating autoregressive text tokens.
Inspired by non-autoregressive decision architectures, this model combines a bidirectional embedding backbone with a 2-layer MLP readout head to output calibrated probabilities over 40 discrete piece placements in under 10ms.
Architecture & Benchmark
- Base Backbone:
google/embeddinggemma-300m(Frozen, bfloat16) - Decision Head: 2-Layer MLP (Linear 768 -> 256 -> GELU -> Linear 40)
- Inference Latency: ~8ms (GPU forward pass) / ~70ms (with multi-objective candidate reranking)
- Empirical Performance: Achieved 21+ continuous line clears across 60 blocks in real-time arcade control.
Technical Article & Post-Mortem
- Full research walkthrough: Beyond Next-Token Prediction: Building Sub-10ms "System 1" Decision Engines with EmbeddingGemma