This repository contains the implementation of AgentSymbiotic, an advanced web agent framework. It introduces a symbiotic learning framework that enables large language models (LLMs) and small language models (SLLMs) to enhance each other iteratively, achieving superior performance in web-based automation tasks. Our code is built upon AgentOccam.
- Mutual Improvement Between LLMs and SLLMs
- Large LLMs generate high-quality trajectory data.
- Distilled small LLMs explore diverse strategies, enriching the dataset.
- Enhanced Knowledge Distillation
- Speculative Data Synthesis: Mitigates off-policy bias and improves SLLM adaptation.
- Multi-Task Learning: Enhances reasoning capabilities of small LLMs.
- Hybrid Mode for Privacy Preservation
- Automatically switches to local SLLMs when handling sensitive data.
- State-of-the-Art Performance
- Achieved 52% success rate with large LLMs and 49% with distilled small LLMs on the WEBARENA benchmark.