EnvLoop LabsLog inEnvLoop / Use cases / Reinforcement learning
Reinforcement learning from your own environments and verifiers, without building RL infrastructure. EnvLoop gives a company the building blocks of the loop: sandboxed environments where an agent attempts your tasks, verifiers that score each attempt, LoRA SFT Train Jobs that post-train on the attempts you keep, and evaluation against the base model. A Model Loop can run these steps round by round.
01 / The loop
Each step is an EnvLoop building block that you can run on its own through the CLI and API.
Package your tasks as an Environment Dataset. Each Task has its inputs and an evaluation definition, and a Job runs the agent’s attempts in sandboxes.
An independent verifier scores every attempt. The Job records each reward, the score coverage and aggregate pass metrics, so a failed run is not mistaken for a low score.
Keep the attempts that the verifiers accept, optionally as a Trajectory Dataset, and turn them into training records. This is verifier-filtered fine-tuning, also called rejection sampling.
Submit a LoRA SFT Train Job on Qwen3.8-27B, GLM-5.3-Flash or another model in the training catalog, and keep the checkpoint.
Run the same Dataset against the base model and the checkpoint, and keep the checkpoint only when it beats the base. A Model Loop reports this comparison for each round.
02 / What the reward does
Rewards decide what is evaluated and which data is kept; they do not update the model directly. The training step is supervised fine-tuning (LoRA SFT) on the selected examples. EnvLoop does not offer online RL training such as GRPO or PPO, and no single API runs the whole loop: you compose the steps through the CLI and API, or follow a Model Loop that runs rounds of data, training and evaluation.
03 / Evidence so far
The CUA-RSIBench pilot is the published example of these pieces working together. Researcher models built training data from verified GUI demonstrations of a fixed teacher, each candidate trained a Qwen3.5-4B LoRA student through EnvLoop training, and Harbor evaluated the students in E2B with an independent verifier. In the gpt-6-sol extension, the selected student scored 2/12 against 0/12 for the base student. The study describes itself as a small data-research pilot, not evidence of sustained recursive self-improvement.
04 / Questions
Yes, as a reward-driven loop built from EnvLoop’s building blocks. Your tasks run as environments in sandboxes, verifiers score each attempt, the verified attempts become training data, and a LoRA SFT Train Job post-trains the model before it is re-evaluated against its base. The reward decides evaluation and data selection; the training step itself is supervised fine-tuning, not online RL such as GRPO or PPO.
No. EnvLoop’s verifiers do not provide a reward inside training. They score evaluation attempts and decide which attempts become training examples, and the model is then post-trained with LoRA SFT on those examples.
LoRA SFT runs on Qwen3.8-27B and GLM-5.3-Flash, and DeepSeek-V4.1-Flash-1017 is a batch training preview for enabled projects without a serving profile. The training catalog in the documentation lists the models available to your project.