GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search

What it claims

GLOBE, from LG Electronics' AI Lab (arXiv v1, October 2026), rewrites structured product listings so that an LLM-based generative ranker places them higher. Its strategies live in a handbook of field-scoped "behaviors," not in one rewrite prompt. Each behavior is tied to the title, features or description field. The authors report that this beats monolithic prompt optimization, reduces harmful rank drops, and transfers to unseen product categories and to unseen ranking models. The summary leaves out the claim about unseen ranking models.

How it works

The objective is rank gain: the target product's rank before the rewrite minus its rank after, with all other candidates held fixed.

Each behavior record holds:

• an ID and a name
• an editing instruction
• the permitted target fields
• a list of conflicting behaviors
• helpful, harmful and neutral counts

The conflict list is a soft constraint. The editor is told to avoid combining those behaviors, but it is not blocked from doing so.

Each iteration runs over a batch of K query-product pairs. The LLM Editor sees the query, the listing and the full active handbook. It picks behaviors, edits only their permitted fields, and reports which ones it used. The ranker then re-ranks the whole candidate set.

Statistics update only after the full batch. Credit assignment is coarse: every behavior applied in an episode gets that episode's direction (helpful, harmful or neutral), because individual contributions cannot be isolated. A behavior's score is helpful count minus harmful count. A behavior stays active while it has fewer than N_min uses (probation) or while its score is non-negative. Deactivated behaviors are kept for traceability but leave the active pool.

The Reflector proposes only ADD or MODIFY deltas. It never deletes. The Moderator filters proposals in two passes:

1. Embedding deduplication rejects a proposal whose cosine similarity to any active behavior exceeds τ. The overview figure uses 0.9.
2. An LLM then checks the remaining proposals for functional overlap in application condition and editing operation.

The Curator commits the survivors to form the next handbook.

flowchart TD
  H[Active handbook] --> E[Editor rewrites K listings]
  E --> R[Ranker: rank gain]
  R --> S[Update counts per applied behavior]
  S --> D{Score < 0 after probation?}
  D -- yes --> X[Deactivate, keep stored]
  S --> F[Reflector: ADD / MODIFY]
  F --> M{Moderator: cosine > τ or functional overlap?}
  M -- yes --> Drop[Reject proposal]
  M -- no --> C[Curator commits]
  C --> H

Limits

The fetched text stops partway through the Moderator description. It contains no experimental tables, datasets, ranker or editor model names, values for K or N_min, costs, or code release, so none of the claimed gains can be checked from this page. Because credit goes to every behavior in an episode, a weak behavior that is often used alongside strong ones can stay active. The editor also self-reports which behaviors it applied, and the page describes no verification of those reports.