TLDR: A new benchmark called Amazon-Bench has been introduced to better evaluate AI web agents on e-commerce platforms like Amazon. Unlike previous benchmarks that focused mainly on product search, Amazon-Bench includes a wider range of tasks such as account management and gift card operations. Crucially, it also assesses agent safety, identifying “harmful failures” where agents might negatively impact a user’s account (e.g., buying the wrong item). The research found that current AI agents struggle with complex tasks and pose significant safety risks, highlighting the need for more robust and reliable development.
Web agents, powered by advanced large language models (LLMs), hold immense potential for automating various tasks on e-commerce websites. From searching for products to managing user accounts, these agents promise to streamline online interactions. However, a recent research paper introduces a new benchmark, Amazon-Bench, that reveals current evaluation methods fall short and highlights significant challenges in agent performance and safety.
The Gaps in Current Evaluation
Existing benchmarks for e-commerce web agents primarily suffer from two major limitations. Firstly, they tend to focus almost exclusively on product search tasks, such as “Find an Apple Watch.” This narrow scope fails to capture the vast array of functionalities offered by real-world e-commerce platforms like Amazon, which include complex operations like account management, address updates, wishlist creation, and gift card operations.
Secondly, current evaluations often overlook the critical aspect of safety. Agents might successfully complete a user query but inadvertently cause negative impacts, such as purchasing the wrong item, deleting a saved address, or incorrectly configuring auto-reload settings for a gift card. These unintended changes pose real risks to users and are not adequately assessed by existing benchmarks.
Introducing Amazon-Bench: A Comprehensive Approach
To address these crucial gaps, researchers from The Pennsylvania State University and Amazon have proposed Amazon-Bench. This new benchmark is designed to provide a more holistic and realistic evaluation of web agents in e-commerce environments.
At its core, Amazon-Bench utilizes a functionality-grounded user query generation pipeline. This innovative approach involves feeding real webpage content and interactive elements (like buttons and checkboxes) to LLMs. This allows the system to generate diverse and realistic user queries that span a broad range of tasks, including adding delivery addresses, managing wishlists, and interacting with brand stores.
Evaluating Performance and Safety
Beyond generating diverse tasks, Amazon-Bench introduces an automated evaluation framework that assesses both the performance and, critically, the safety of web agents. The framework categorizes outcomes into three types:
-
Success: The agent successfully completes the task without any negative impact.
-
Benign Failure: The agent fails to complete the task, but no changes are made to the user’s account or status.
-
Harmful Failure: The agent performs actions that result in a negative impact on the user, such as making an incorrect purchase, modifying account settings without explicit instruction, or adding unwanted items to the cart. Even if the main task is eventually completed, harmful side effects still classify it as a harmful failure.
This distinction is vital for understanding the true reliability of web agents in real-world scenarios where user data and financial transactions are involved.
Key Findings and Agent Challenges
The research systematically evaluated various agents, including Deepseek-R1, GPT-4o, GPT-o4-mini, Claude-3.7, GPT-4.1, WebVoyager, and Nova-Act, on the Amazon-Bench. The findings highlight significant challenges:
-
Struggling with Complexity: Current agents often struggle with complex queries that go beyond simple product searches. Tasks involving store interactions, for instance, proved particularly challenging.
-
Safety Risks: A notable finding was the presence of safety risks. Product interaction and account management tasks showed higher rates of harmful failures, where agents made unintended changes to user accounts or carts. For example, an agent might add two items to a cart when the user only requested one, or get stuck in a loop searching for a non-existent button.
-
Performance Variation: While GPT-4.1 achieved the highest overall success rate, Nova-Act demonstrated the lowest harmful failure rate. This indicates a trade-off between task completion and safety across different models.
-
Efficiency: Agents also varied in efficiency, measured by the number of steps taken and tokens used per query. GPT-4.1 was found to be the most efficient in terms of steps.
Also Read:
- Benchmarking AI’s Tool-Using Abilities in the Real World
- Beyond Jailbreaks: Unpacking the True Criminal Potential of Large Language Models
The Path Forward
The introduction of Amazon-Bench underscores the need for developing more robust and reliable web agents. While LLMs show promise, their current capabilities in handling diverse e-commerce functionalities and ensuring user safety require substantial improvement. The benchmark’s focus on functionality-grounded queries and a nuanced safety evaluation provides a crucial tool for guiding future research and development in this rapidly evolving field.
The paper also acknowledges limitations, such as its focus on single-turn tasks and the absence of user-context awareness, pointing to promising directions for future work in creating even more sophisticated and personalized AI agents for the web.
