First, set the local path in hydra_configs/env/default.yaml and an OpenAI API key in hydra_configs/qwen_0.5B_cg.yaml

Then, to run the letter avoidance experiment with 1000 generations per algorithm and horizon H=32, we run the following:
$ bash run_cg.sh 1000 32
$ bash run_cg_winrate.sh 1000 32

The first script prints the average log-probabilities for each algorithm (TW = ActionLevelRS, ResetTW = ActionLevel-Reset, WJS0.1 = VGB with alpha=0.1, etc). The second script prints the winrate of VGB (for each choice of alpha) against each baseline.
