OpenClaw-RL
An online reinforcement learning scheme within the reef codebase, designed for live production traffic. It enables a model to update its weights during real conversations while serving users, using binary reward signals extracted from users' subsequent replies.
Timeline (last 90 days)
reef's OpenClaw-RL recipe uses 72 GSM8K problems as the task stream, with Qwen3-4B serving simultaneously as the policy model and the process reward model, Qwen3-32B as the student model, using seven GPUs, and setting a threshold that must be reached before an update is accepted.
The weight recipes loop includes SAO, OpenClaw-RL, TTT-Discover, and Guidance-TTT, and requires GPUs and a full training stack.
OpenClaw-RL is designed specifically for live traffic, with the policy updated directly from the conversation itself, and the reward signal is a binary value read from the user's next-turn response.
All Facts (3)
reef's OpenClaw-RL recipe uses 72 GSM8K problems as the task stream, with Qwen3-4B serving simultaneously as the policy model and the process reward model, Qwen3-32B as the student model, using seven GPUs, and setting a threshold that must be reached before an update is accepted.
50%UnverifiedThe weight recipes loop includes SAO, OpenClaw-RL, TTT-Discover, and Guidance-TTT, and requires GPUs and a full training stack.
50%UnverifiedOpenClaw-RL is designed specifically for live traffic, with the policy updated directly from the conversation itself, and the reward signal is a binary value read from the user's next-turn response.
50%