sim2gym AccelGym

# loading

Loading tasks…

Pick a task and a model, then press Run episode.

The model designs a systolic array for a real neural-network layer. It may call the real SCALE-Sim simulator a few times before answering, and its answer is scored against the exhaustive optimum.

The model sees only the task prompt and the simulator's observations, never the optimum.

Insights

What sim2gym tests, and what we found, in plain language

Loading insights…

Leaderboard

Real SCALE-Sim scores from results/*.jsonl

Models vs baselines

Model	Type	n	Mean score	95% CI	Solved ≥0.95	Feasible	Sim calls	Latency s	Errors
Loading…

Score = exhaustive-optimum cycles ÷ answer cycles (real SCALE-Sim); 0 if infeasible or no answer. Baselines are not model inference.