reef
An open-source repository containing two implementations of self-improving agents: a harness loop for configuration search and a weight recipes loop for weight training. The project is designed to explore the boundary between AI self-improvement and reinforcement learning.
Timeline (last 90 days)
The poster proposed two directions for improvement: rejection memory and a conservative release policy, and implemented these policies in the tool reef that he maintains.
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 reef codebase contains two types of loops within the same codebase: the harness loop (framework loop) and the weight recipes loop.
The harness loop requires only a single model endpoint, needs no GPUs, and does not update model parameters.
All Facts (4)
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 reef codebase contains two types of loops within the same codebase: the harness loop (framework loop) and the weight recipes loop.
50%UnverifiedThe harness loop requires only a single model endpoint, needs no GPUs, and does not update model parameters.
50%UnverifiedThe poster proposed two directions for improvement: rejection memory and a conservative release policy, and implemented these policies in the tool reef that he maintains.
50%