EMNLP 2026 / LLM Agents

Inventory-Grounded Policy-Level Optimization for Training-Free AI Search

An EMNLP 2026 Industry Track paper on training-free AI search, using inventory-grounded policy-level optimization to keep search behavior aligned with changing catalog and inventory conditions.

arXiv Publication record
Infographic for Inventory-Grounded Policy-Level Optimization for Training-Free AI Search

Research area

AI search, skill retrieval, NL2SQL benchmarks, LLM serving systems, and capability governance are treated as system surfaces that should be inspected before they are trusted.

Publication record

Authors
Wei Zhou, Tiandeng Wu, Jiandong Ding, Zhufeng Fan, Yi Cao
Venue
Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track (EMNLP '26)
Year
2026
Area
LLM Agents

Problem

AI search systems must reason over changing inventories, where a static prompt or fixed search policy can become stale as available items, attributes, and user intent patterns shift.

Approach

The work frames search behavior as a policy-level optimization problem grounded in inventory evidence, improving the search procedure without retraining the underlying foundation model.

Public evidence

The public paper reports deployment in a commercial smart-assistant search system since May 2026. In a 14-day online A/B test, the complete IGPO treatment produced a 3.17% relative CTR lift and reduced audited bad cases by 38.9%.

Takeaway

The paper extends the LLM agents line from skill retrieval and service evaluation into AI search systems where policy, evidence, and inventory state must stay aligned.