Vibrant Labs
The bottleneck for RL-trained LLMs isn't algorithms — it's environments.
Core trick
Runs an LLM multi-agent simulation, then mines its interactions into reset/step/reward environments for RL training.
Description
The bottleneck for RL-trained LLMs isn't algorithms — it's environments. We have GRPO and compute, but no scalable way to create diverse, realistic training environments.
EnterpriseSimulator explores one approach: growing RL environments from simulated worlds instead of hand-crafting them.
1. Simulate a rich world — A Smallville-style multi-agent sim where LLM-powered customers, staff, and managers interact, producing organic scenarios.
2. Mine tasks — A task miner extracts RL-ready scenarios automatically from simulation data.
3. Train via OpenEnv — Each task becomes a gym-like env with reset/step/reward. The agent interacts with simulated customers. Reward = resolution + satisfaction + efficiency.
The key insight: world simulation is environment generation. You don't write scenarios — you grow them. This pattern (simulate → mine → train) could generalize to any domain.
README excerpt
# CoreCraftSim  ## Why? The bottleneck for RL-trained LLMs isn't algorithms — it's environments. We have GRPO, we have compute, but we don't have a scalable way to create the diverse, realistic environments needed to train agents. Research areas like **Unsupervised Environment Design (UED)** and **Open-endedness** point toward solutions, but there's a gap between theory and practice. CoreCraftSim is an experiment exploring one concrete approach: **what if you could grow RL environments from simulated worlds instead of hand-crafting them?** ## Enter Simulators This work is inspired by [Generative Agents: Interactive Simulacra of Human Behavior](https://arxiv.org/abs/2304.03442) — the idea that LLM-powered agents can simulate believable human behavior in rich, persistent worlds. We apply this to enterprise environments: instead of a small town, we simulate a company with customers, support agents, and managers. ## Enter CoreCraftSim A Smallville-style multi-agent simulation where LLM-powered customers, support agents, and managers interact through a shared database. This produces organic, realistic scenarios — not scripted ones. ``` Simulate World → Mine Tasks → Train Agents ```  The simulation runs on a tick-based engine (48 ticks x 5 min) with 4 phases per tick: 1. **Customers** generate messages based on their persona and situation 2. **Routing** assig