GLOBE Framework: Enabling AI Agents to Self-Evolve in Generative Search

E-commerce product optimization faces new challenges under generative search: language agents must not only rewrite product listings, but also accumulate successful strategies across different categories to achieve continuous self-improvement. However, existing methods often treat structured listings as free text, causing effective strategies to remain hidden in prompts, search trajectories, or rewritten outputs, unable to be consolidated into reusable agent memory. To address this pain point, researchers propose the GLOBE framework, which stores optimization knowledge through explicit, field-level behavior manuals, driving continuous self-evolution of agents.

From Implicit Strategies to Explicit Memory: GLOBE's Core Design

The core idea of GLOBE is to extract optimization knowledge from implicit states and transform it into editable, traceable behavior entries. In each optimization round, an "Editor" role applies relevant behaviors and receives effectiveness signals based on ranking feedback; subsequently, a "Reflector" converts these feedback signals into behavior update proposals. This design frees the agent from black-box prompt tuning, instead building a structured optimization memory library around specific fields.

Compared to traditional monolithic prompt optimization, this field-level behavior representation offers greater interpretability. Each behavior explicitly corresponds to a rewriting strategy for a specific product field—for example, how to embed keywords in the title, or how to organize selling points in the description. When feedback arrives, the agent can pinpoint exactly which behavior needs adjustment, rather than blindly searching through a vague prompt space.

Field-level behavior representation refers to decomposing product listing optimization knowledge into executable strategy entries that correspond one-to-one with specific fields (such as title, description, selling points, etc.), rather than holistic prompt adjustments. In e-commerce scenarios, product information naturally possesses structured characteristics—titles, attributes, descriptions, and other fields each serve different search matching and conversion functions. The advantage of field-level representation is that when ranking feedback occurs, the agent can precisely identify which field's rewriting strategy caused the effectiveness change, enabling targeted corrections. In contrast, free-text prompt optimization mixes all fields together, making feedback signals difficult to attribute and strategies difficult to reuse. This representation also facilitates review and editing by human operators, moving the agent's learning process from a black box toward interpretability and intervenability.

Triple Review Mechanism: Ensuring Reliable and Controllable Evolution

In real e-commerce scenarios, agent self-evolution must balance effectiveness and risk. GLOBE designs three key roles to gatekeep the behavior update process. The "Arbitrator" removes redundant proposals to prevent behavior library bloat and conflicts; the "Curator" formally commits approved changes to the manual; and behaviors consistently proven harmful are deprecated.

This mechanism transforms each round of episodic feedback into persistent, field-level optimization memory. You may not have noticed, but automatic deprecation of persistently harmful behaviors is particularly important—it allows the agent to cut losses promptly while exploring new strategies, reducing the risk of deployment in actual business operations. Experiments also show that feedback-driven evolution combined with moderate review significantly reduces negative risks while improving effectiveness.

Episodic feedback (round feedback) is a concept commonly used in reinforcement learning and agent research, referring to the immediate effectiveness signals an agent receives within a single task episode or short-term interaction—for example, a ranking increase or decrease after a product rewrite. Unlike persistent memory, episodic feedback is transient and local; if not distilled, it disappears when the episode ends. GLOBE's key design lies in converting this short-term, per-episode feedback into persistent behavior update proposals through the Reflector, which are then reviewed and stored by the Arbitrator and Curator, achieving the transformation from "one-time experience" to "long-term memory." This process is analogous to humans summarizing reusable operation manuals through repeated practice, rather than starting from scratch with trial and error for each task.

Experimental Validation: Structured Behaviors Outperform Holistic Prompt Optimization

The research team's experiments compared two paradigms: structured behavior representation and holistic prompt optimization. Results show that after decomposing optimization knowledge into field-level behaviors, agent performance outperforms traditional monolithic prompt optimization. Further feedback-driven evolution and review mechanisms build on this to further improve effectiveness and reduce downside risk.

More notably, learned behaviors remain effective on unseen product categories. This indicates that the evolution of agent memory not only improves performance on current tasks but also produces transferable optimization capabilities. For e-commerce platforms, this means that when a new category launches, the agent can quickly adapt with its existing behavior manual, without having to explore from scratch.

Cross-category transfer has important practical significance in e-commerce scenarios. Different product categories (such as clothing, home appliances, beauty) differ significantly in search query patterns, consumer focus points, and keyword distributions; traditionally, each category requires separately accumulated optimization experience. GLOBE's experiments show that strategies learned in field-level behavior manuals can transfer to unseen categories, indicating that some optimization knowledge possesses generality beyond specific categories—for example, principles such as placing keywords early in titles and highlighting core selling points in descriptions apply across different categories. Of course, the effectiveness of transfer may be influenced by factors such as inter-category similarity and the abstraction level of behaviors, which is why the original article mentions that "boundary conditions for cross-category transfer" remain to be explored.

Implications for E-commerce AI Applications

GLOBE's value lies in providing a feasible path for language agents to continuously self-improve in vertical domains. Making optimization strategies explicit as field-level behaviors not only facilitates understanding and intervention by human operators, but also provides an auditable carrier for agent autonomous evolution. As generative search increasingly becomes the entry point for product discovery, this interpretable and transferable agent memory mechanism may become a new standard for e-commerce product optimization tools.

Of course, the framework is still at the research stage, and issues such as the scaling of behavior manuals and boundary conditions for cross-category transfer remain to be further explored. But the "structured memory + feedback evolution" approach it represents provides valuable reference for long-term operation of agents in real business environments.

One-line Summary

GLOBE framework makes e-commerce optimization knowledge explicit as field-level behavior manuals, driving continuous agent self-evolution.

Paragraph Summary

The GLOBE framework addresses new challenges in e-commerce product optimization under generative search by extracting optimization knowledge from implicit prompts and search trajectories and storing it as explicit, field-level behavior manuals. Through a triple review mechanism—where an Editor applies behaviors, a Reflector converts feedback, an Arbitrator removes redundancy, and a Curator commits changes—the framework transforms per-episode episodic feedback into persistent optimization memory and automatically deprecates harmful behaviors. Experiments show that structured behavior representation outperforms holistic prompt optimization, feedback-driven evolution further improves effectiveness and reduces risk, and learned behaviors transfer to unseen product categories, providing an interpretable and auditable path for continuous self-improvement of e-commerce AI.