Inference-time steering experiments for multi-agent TextCraft and Number Search rollouts using SMC.
The experiments use platoon for the agent implementation and environments/task.
uv syncexamples/textcraft_multi_agent.py: prompt-intersection TextCraft rollouts.examples/number_search_cfg.py: Number Search rollouts with a structural CFG constraint.examples/number_search_cot.py: Number Search rollouts with the same CFG plus CoT sentence constraints.examples/run_vllm_rollouts.py: baseline rollouts against a plain OpenAI-compatible vLLM server.examples/run_genlm_shard_jobs.sh: shard launcher for the genlm-based experiments above.
steerbot/utils/llm.py: provides core wrapper interface to register and expose genlm models using the OpenAI/LiteLLM API.
uv run python examples/number_search_cfg.py --task-start 0 --task-end 20./examples/run_genlm_shard_jobs.sh cfg --gpus 0,1 --task-start 0 --task-end 20