Mimir

Neural circuit synthesis — learning to predict, generate, and repair computational circuits.

Neurosymbolic

Neural predictions are always verified symbolically via Asgard's simulator. No unchecked guesses.

Grammar-Constrained

Syntactically invalid circuits are impossible by construction. Every generated output is well-formed.

Curriculum Learning

Progressive difficulty training prevents mode collapse and builds robust generalization.

Behavioral Loss

Simulation-based loss compares predicted vs target circuit behavior during training.

Transformer Architecture

Sequence-to-sequence model over tokenized circuit ASTs with attention-based decoding.

Cloud Training

One-command training on GCP with GPU acceleration. Scale from laptop to cloud seamlessly.

Comprehensive Benchmarking

Validity, behavioral accuracy, diversity, and recovery metrics for thorough model evaluation.

Training Monitor

Real-time web dashboard for tracking training progress, loss curves, and curriculum stages.

Asgard Integration

Built on Gimle Asgard's circuit algebra and JAX runtime for symbolic verification.

How It Works

1

Generate

Create synthetic circuits with Asgard and simulate them to produce behavioral trajectories — input/output pairs that characterize each circuit's function.

2

Train

Train a transformer to predict circuit structure from behavioral trajectories, using curriculum learning and grammar-constrained decoding to ensure valid outputs.

3

Verify

Verify predictions symbolically — simulate the predicted circuit and compare its behavior against the target. Neural speed, symbolic correctness.

Getting Started

Install dependencies and launch a quick training run:

# Install with uv
uv sync --extra dev

# Train a small model with curriculum learning
uv run python scripts/train.py --model-size small --use-curriculum --curriculum-type fast