This page focuses on example task setups only: task type, seeds, and evolution parameters for each run configuration. Best-node artifacts are shown on the dedicated Best Node page.
These tasks are examples, not hard-coded limits of CliffSearch. The evolutionary loop is task-agnostic: to use CliffSearch on a custom problem, you supply a benchmark that is interfaced with the search runtime and exposes the artifact contract plus the primary metric to optimize.
Reported losses are validation-set losses, averaged over the benchmark seeds and hyperparameter settings defined for each task. For the evolved transformer hyper-connections and optimizer tasks, these benchmark hyperparameters are fixed across nodes rather than tuned per node, so the absolute losses are typically higher than what per-node retuning would achieve. When some benchmark runs fail, the aggregate uses worst-successful-loss imputation for the missing runs. The benchmark is meant to support fair relative comparison across nodes, not individually tuned final training runs.
Evolve custom manifold hyper-connection modules inside a fixed transformer NanoGPT training stack with 4 hyper-connections on the Shakespeare dataset, evaluated with 3 benchmark random seeds. The website includes analyzed runs for both theory+code and matched code+design-intent artifact modes. In config names this mode appears as code_only; formal theory_content is empty but summary_md still carries the node's design and ideation principles.
Evolve EvoOptimizer(torch.optim.Optimizer) while keeping a regular NanoGPT stack on the Shakespeare dataset fixed, again evaluated with 3 benchmark random seeds.
Example ablation task: evolve EvoOptimizer(torch.optim.Optimizer) on small native classification benchmarks with linear and MLP models; the reported validation loss is averaged over 32 benchmark runs per node across 4 datasets, 2 seeds, and a 2x2 learning-rate/weight-decay grid.
| Task | Bundle | Prompt | Mode | Seeds | Status |
|---|
