xiamoent/Agent-G2-alfworld-7b
xiamoent/Agent-G2-alfworld-7b is a 7.6 billion parameter language-agent checkpoint, initialized from Qwen2.5-7B-Instruct and post-trained with Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning. This model is specifically designed and optimized for research in the sandboxed ALFWorld text environment, excelling at long-horizon language agent tasks. It achieves a 98.4% success rate across all ALFWorld tasks, making it suitable for studying adaptive expert-prefix guidance and agentic reinforcement learning.
Loading preview...
Agent-G2 ALFWorld 7B Overview
xiamoent/Agent-G2-alfworld-7b is a specialized language-agent model, built upon the Qwen2.5-7B-Instruct architecture and enhanced through post-training with Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning. This 7.6 billion parameter model, with a configured context length of 32,768 tokens, is exclusively tailored for the ALFWorld / ALFRED text environment.
Key Capabilities and Features
- ALFWorld Specialization: Designed for research in the sandboxed ALFWorld text environment, not as a general-purpose chat model or physical robot controller.
- Gaussian Guidance: Utilizes Agent-G2's novel approach of sampling expert-prefix depth from an adaptive Gaussian distribution, updated from rollout statistics for policy optimization.
- High Success Rates: Achieves impressive success rates in ALFWorld tasks, including 100% for Pick, Look, Clean, Heat, and Cool tasks, and 98.4% across all tasks.
- Reinforcement Learning Focus: Employs expert-prefix guidance as a training mechanism, with the learned policy evaluated without requiring an expert trajectory at inference time.
- Structured Output: Expects English output in a specific
<think>...</think><action>...</action>format for reasoning and action selection.
Intended Use Cases
- Reproducing Agent-G2 Results: Ideal for replicating the project's findings within the ALFWorld text environment.
- Research on Language Agents: Suitable for studies on long-horizon language agents and agentic reinforcement learning.
- Adaptive Expert-Prefix Guidance: Useful for investigating the methodology of adaptive expert-prefix guidance.
- Action Selection Evaluation: Designed for evaluating action selection over environment-provided admissible action sets.
It's important to note that the model's performance is sensitive to prompt formatting and environment configuration, and it may produce malformed actions requiring environment-side validation. For faithful evaluation, users should utilize the ALFWorld environment, prompt template, and rollout loop provided by the Agent-G2 repository.