RLOO (REINFORCE Leave-One-Out)

RLOO is a simple policy gradient baseline that avoids a critic. Advantage for each sample is computed against the average of its siblings (leave-one-out), which acts as a per-prompt variance-reduction baseline.

Reference: Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs.

Canonical Scripts

Script	Infer	Train	Platform
run_qwen3_8b_fsdp.sh	vLLM	FSDP	NVIDIA

Override any argument via env vars at the top of the script (e.g. MODEL_PATH=Qwen/Qwen3-14B bash run_qwen3_8b_fsdp.sh).

Key Flags

• algorithm.adv_estimator=rloo
• actor_rollout_ref.rollout.n=5 — RLOO needs ≥2 samples per prompt; 5 is a common default.
• actor_rollout_ref.actor.use_kl_loss=False and algorithm.use_kl_in_reward=True — RLOO typically uses reward-side KL, not loss-side KL.