Symbiotic Cooperation for Web Agents: Harnessing Complementary Strengths of Large and Small LLMs

Ruichen Zhang*1, Mufan Qiu*1, Zhen Tan*2, Mohan Zhang1, Vincent Lu3, Jie Peng1, Kaidi Xu4, Leandro Z. Agudelo5, Peter Qian3, Tianlong Chen1,
1University of North Carolina at Chapel Hill; 2Arizona State University; 3Ventus AI; 4Drexel University; 5Daice Labs
*Equal Contribution
PDF Code arXiv
Figure 2

Figure 1: Overview of the AgentSymbiotic framework. Step 1: The large LLM interacts with the environment to generate high-quality trajectories, which are then used to distill small LLMs. Step 2: Multi-task learning and Speculative Data Synthesis are applied during distillation to enhance the reasoning capabilities of the small LLM and mitigate off-policy bias between the two LLMs. Step 3: The small LLM further explores the environment to produce diverse and valuable trajectories. Step 4: Then the knowledge base containing high-quality trajectories and comprehensive trajectories is incorporated into the large LLM's RAG process, improving its performance. This iterative process establishes a symbiotic improvement cycle, enhancing both large and small LLMs over time.

Introduction

In this paper, we propose AgentSymbiotic, an iterative framework that couples data synthesis with task-performance, yielding a "symbiotic improvement" for both large and small LLMs.Our study uncovers a complementary dynamic between LLM types: while large LLMs excel at generating high-quality trajectories for distillation, the distilled small LLMs—owing to their distinct reasoning capabilities—often choose actions that diverge from those of their larger counterparts. This divergence drives the exploration of novel trajectories, thereby enriching the synthesized data.However, we also observe that the performance of small LLMs becomes a bottleneck in this iterative enhancement process. To address this, we propose two innovations in LLM distillation: a speculative data synthesis strategy that mitigates off-policy bias, and a multi-task learning approach designed to boost the reasoning capabilities of the student LLM.Furthermore, we introduce a Hybrid Mode for Privacy Preservation to address user privacy concerns. Evaluated on the WebArena benchmark, AgentSymbiotic achieves SOTA performance with both LLM types. Our best Large LLM agent reaches 52%, surpassing the previous best of 45%, while our 8B distilled model demonstrates a competitive 49%, exceeding the prior best of 28%.

AgentSymbiotic

Our contributions can be summarized as follows:
▶ Synergistic Framework. We present a novel framework that establishes an iterative, symbiotic cycle between large and small LLMs, enabling them to leverage their complementary strengths for mutual enhancement.

▶ Technical Innovations. We introduce two key advancements in distillation techniques: (a) a speculative data synthesis strategy to counteract off-policy bias and (b) a multi-task learning approach to maintain reasoning capabilities.

Figure 2

Figure 2: Overview of two key innovations in LLM distillation: (a) Speculative Data Synthesis, which mitigates off-policy bias by leveraging both large and small LLMs. At each step, the small LLM generates an action based on the observation, while the large LLM produces a set of top-K action candidates. If the small LLM's action is within the large LLM's top-K actions, it is accepted (✓); otherwise, the large LLM's action is chosen for subsequent interactions (✗). (b) Multi-task Learning, which enhances reasoning capabilities by training small LLM to predict both actions and rationales, enabling it to handle multiple tasks and address missing reasoning capabilities during distillation. CoT indicates Chain-of-Thought.


▶ State-of-the-art Performance. Experiments show that on the WebArena benchmark, our method achieves state-of-the-art performance with both LLM types: the large LLM achieves 52\%, surpassing the previous best open-source 45\%, while our 8B distilled LLaMA-3 model achieves 49\%, approaching the performance of agents based on Claude-3.5.


Method Model SR (%) ↑
Learn-by-InteractClaude-3.539.2*
API-Based AgentGPT-4o43.9*
AgentOccamGPT-4-Turbo45.7*
AgentOccamClaude-3.548.5
AgentSymbioticClaude-3.552.1
Vanilla promptingLLaMA-1B2.4
Vanilla promptingLLaMA-8B5.6
Vanilla promptingDeepSeek-R1-8B8.5
Learn-by-InteractCodegemma-7B17.9*
Learn-by-InteractCodestral-22B28.0*
AgentSymbioticLLaMA-1B24.1
AgentSymbioticDeepSeek-R1-8B43.6
AgentSymbioticLLaMA-8B48.5
AgentSymbiotic-HybridClaude-3.5 + LLaMA-8B50.5

Comparison of final success rates (SR) among various large LLM and small LLM base agents on WebArena. Scores marked with * indicate cited scores from the corresponding papers' experiment scores.



▶ Hybrid Mode for Privacy Preservation. We integrate a hybrid mode for privacy preservation that directs sensitive tasks to a local small LLM, ensuring that private user data remains secure.


Figure 2

Figure 3: The Privacy Detector analyzes each step's observation and action for private data. If detected, a local small LLM ensures confidentiality by predicting the next action and reason. Otherwise, a cloud-based large LLM handles predictions, leveraging its superior reasoning capabilities for non-sensitive tasks.



While the large LLM offers superior reasoning and broader knowledge, many web tasks involve confidential or high-stakes information (e.g., passwords and payment details) that must be handled with care. To safeguard user privacy, we propose a hybrid mode that allows the system to automatically switch between a local small LLM and a cloud-based large LLM. This hybrid mode operates as follows:

Privacy Detection. Before processing any environment observation or action, the content is scanned by a local model, which flags potential private information—such as personally identifiable data or security tokens.

Local Processing. If the observation or action is deemed private, the decision-making step is delegated to the small LLM deployed locally.

Cloud Processing. If no private information is detected, the agent leverages the large, cloud-based LLM to benefit from its advanced capabilities.


Figure 1 Figure 2

Figure 4: The Privacy Detector detects and categorizes tasks containing privacy information. We analyze their distribution to understand privacy interactions better.





BibTeX

@article{zhang2025symbiotic,
            title={Symbiotic Cooperation for Web Agents: Harnessing Complementary Strengths of Large and Small LLMs},
            author={Zhang, Ruichen and Qiu, Mufan and Tan, Zhen and Zhang, Mohan and Lu, Vincent and Peng, Jie and Xu, Kaidi and Agudelo, Leandro Z and Qian, Peter and Chen, Tianlong},
            journal={arXiv preprint arXiv:2502.07942},
            year={2025}
          }