LLM-in-Sandbox Elicits General Agentic Intelligence

This model, daixuancheng/Qwen3-4B-Instruct-2507-LLM-in-Sandbox-RL, is a 4 billion parameter instruction-tuned checkpoint derived from the Qwen/Qwen3-4B-Instruct-2507 base model. It has been specifically trained using Reinforcement Learning (RL) within the novel LLM-in-Sandbox framework, as detailed in the paper "LLM-in-Sandbox Elicits General Agentic Intelligence" (arXiv:2601.16206).

Key Capabilities

• Agentic Intelligence: Fine-tuned to demonstrate general agentic intelligence, enabling autonomous decision-making and interaction.
• RL-Trained: Benefits from Reinforcement Learning on a specialized dataset (llm-in-sandbox-rl) to enhance its agentic behaviors.
• Tool Use: Supports features like enable-auto-tool-choice and tool-call-parser hermes for integrating external tools.
• Scalable Inference: Optimized for serving with vllm, including prefix caching and tensor parallelism.

Good For

• Research in Agentic AI: Ideal for researchers exploring agentic capabilities and autonomous systems.
• Simulated Environments: Excels in tasks requiring interaction and decision-making within sandbox-like settings.
• Reproducing Paper Results: Provides the exact checkpoint used in the "LLM-in-Sandbox Elicits General Agentic Intelligence" paper for verification and further study.

For detailed usage and reproduction of results, refer to the RL training code and inference code.