Paper 2601.16206
Computer Environments Elicit General Agentic Intelligence in LLMs
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 5
- GitHub
- 241 stars
01 In brief
Summary
The paper introduces LLM-in-Sandbox, a paradigm that virtualizes a computer as a minimal code sandbox (with bash, file_editor, and finish tools) to elicit general agentic intelligence in LLMs.
Without additional training, strong models (e.g., Claude-Sonnet-4.5-Think, GPT-5, DeepSeek-V3.2-Thinking) show consistent gains across mathematics, physics, chemistry, biomedicine, long-context understanding, and instruction following, with improvements up to +15.5% (Qwen3-Coder on math) and token reductions up to 8x in long-context tasks.
These gains arise from three meta-capabilities: external resource access, file management, and code execution.
Weaker models (e.g., Qwen3-4B-Instruct) fail to benefit without training.
The authors propose LLM-in-Sandbox-RL, a reinforcement learning method training models on general non-agentic context-based tasks within the sandbox, which enables weaker models to outperform their vanilla LLM mode and also improves text-only LLM mode via internalized reasoning patterns.
The sandbox uses a single shared Docker image (~1.1 GB) with negligible memory overhead, contrasting with task-specific environments in SWE agents.
The paradigm also enables capabilities beyond text generation, such as creating interactive maps, posters, videos, and music.
02 From the paper
Abstract
Agentic intelligence in large language models (LLMs) requires not only model intrinsic capabilities but also interactions with external environments. Equipping LLMs with computers now represents a prevailing trend. However, the computer environment's intrinsic value has not been systematically investigated, particularly its potential to elicit general capabilities. Here we introduce LLM-in-Sandbox, which virtualizes the computer as a code sandbox with only basic functionalities, and demonstrate that this minimal setting elicits computer-based meta-capabilities for general task solving: external resource access, file management, and code execution. Without additional training, strong models achieve substantial gains (up to 15.5%) across mathematics, physics, chemistry, biomedicine, long-context understanding, and instruction following, while reducing token consumption by up to 8 times. Furthermore, we develop LLM-in-Sandbox-RL to train models exclusively on non-agentic data within the sandbox, empowering weaker models to harness the environment and internalize these interactions. Our results demonstrate that computer environments elicit general intelligence, yield efficiency gains, and can be harnessed through training, serving as a promising foundation for generalist agents.