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

Source

• Source name: OpenReview
• Source kind: paper
• Authors / organization: Ruichen Zhang, Mufan Qiu, Zhen Tan, Mohan Zhang, Xiaopeng Lu, Jie Peng, Kaidi Xu, Leandro Z. Agudelo, Peter Zhenghao Qian, Tianlong Chen
• Venue: ICLR.cc/2026/Conference
• Published: UNKNOWN
• Collected: 2026-06-29
• URL: https://openreview.net/forum?id=J7sbjTbOll
• PDF: https://openreview.net/pdf/879770078e33420e3f89d8fdee987b7831364d89.pdf
• DOI: UNKNOWN

Plain-Language Brief

Researchers propose AgentSymbiotic, an iterative framework that enhances both large and small Language Model (LLM) performance in web browsing tasks by combining data synthesis with task-performance. Large LLMs generate high-quality trajectories for distillation while the distilled small LLMs offer distinct reasoning capabilities that lead to novel trajectories. The study also addresses user privacy concerns through a Hybrid Mode. AgentSymbiotic outperformed the State of the Art (SOTA) on the WEBARENA benchmark.

What the Document Says

• Researchers propose an iterative framework called AgentSymbiotic.
• The framework combines data synthesis with task-performance for both large and small LLMs.
• Large LLMs generate high-quality trajectories for distillation.
• Distilled small LLMs offer distinct reasoning capabilities that lead to novel trajectories.
• The study observes a complementary dynamic between large and small LLMs.
• Off-policy bias in data synthesis is mitigated through a speculative strategy.
• A multi-task learning approach boosts the reasoning capabilities of the student LLM.
• A Hybrid Mode for Privacy Preservation is introduced.
• AgentSymbiotic outperformed SOTA on the WEBARENA benchmark.
• The best Large LLM agent reached 52%, surpassing the previous best of 45%.
• The 8B distilled model demonstrated a competitive 49%, exceeding the prior best of 28%.

Key Concepts Mentioned

• Language Models (LLMs)
• Web browsing agents
• Iterative framework
• Data synthesis
• Task-performance
• Speculative data synthesis strategy
• Multi-task learning
• Hybrid Mode
• Privacy Preservation
• SOTA performance

Claims Made by the Document

• AgentSymbiotic outperforms SOTA on the WEBARENA benchmark.
• The best Large LLM agent reaches 52% and surpasses the previous best of 45%.
• The 8B distilled model demonstrates a competitive 49% and exceeds the prior best of 28%.

Evidence Used

• 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%.
• Our 8B distilled model demonstrates a competitive 49%, exceeding the prior best of 28%.

Methods / Approach

• Iterative framework combining data synthesis with task-performance
• Speculative data synthesis strategy
• Multi-task learning

Limitations Stated by the Authors

The study does not specify when or where AgentSymbiotic will be released.

Limitations Noted by the Agent

Agent interpretation — not from the source.

Not stated in source

Open Questions

• When and where will AgentSymbiotic be released?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The researchers mention that the code will be released upon acceptance, but no specific date or venue is provided.

Raw Abstract / Excerpt

Web browsing agents powered by large language models (LLMs) have shown tremendous potential in automating complex web-based tasks. Existing approaches typically rely on large LLMs (e.g., GPT-4o) to explore web environments and generate trajectory data, which is then used either for demonstration retrieval (for large LLMs) or to distill small LLMs (e.g., Llama3) in a process that remains decoupled from the exploration. 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%. Code will be released upon acceptance.