Source-linked AI summary
Inventory-Grounded Policy-Level Optimization for Training-Free AI Search
Wei Zhou, Tiandeng Wu, Jiandong Ding, Zhufeng Fan, Yi Cao
TL;DR
Early-stage AI search operates over changing inventories that cannot be encoded reliably in fixed prompts or strategies, motivating a training-free alternative. IGPO separates policy from environment facts, grounds decisions in runtime inventory evidence, and refines guidelines offline; its complete treatment produced a 3.17% relative CTR lift and 38.9% fewer audited bad cases in a 14-day A/B test.
Problem
Rapidly changing catalog items and metadata make fixed prompts and strategies unreliable, while fine-tuning, reinforcement learning, and static prompt patches fit poorly for early-stage AI search.
Method
IGPO learns Policy Guidelines for runtime inventory reasoning, injects them into retrieval and selection prompts, and uses grouped rollouts with inventory-guided exploration to refine them without retraining model weights.
Results
3.17% relative CTR lift and 38.9% fewer audited bad cases were observed in a 14-day online A/B test of the complete IGPO treatment.
Takeaways & Limitations
IGPO provides a practical way to adapt commercial AI search under dynamic inventory without retraining model weights.
Takeaways & Limitations
IGPO depends on the quality and coverage of runtime inventory evidence and cannot recover inventory or metadata absent from the underlying system.
Abstract
from arXiv · showhide
Early in deployment, an AI search system typically operates over a frequently updated product catalog, so the available items and their properties cannot be treated as stable knowledge that can be encoded in fixed prompts or strategies. Fine-tuning, reinforcement learning, and static prompt patches fit poorly: labels are scarce, rewards drift with inventory, model releases are costly, and prompt fixes quickly stale. We present Inventory-Grounded Policy-Level Optimization (IGPO), a training-free approach for fixed AI search pipelines. IGPO separates policy from environment facts: it learns Policy Guidelines for acting on runtime inventory evidence rather than memorizing available items. Online, IGPO grounds each query by probing the inventory and constructing an inventory portrait, then injects relevant Policy Guidelines into the retrieval and selection prompts. Offline, stochastic rollouts are grouped by query -- mixed outcome groups directly yield contrastive signal, and an inventory-guided exploration loop distinguishes missed retrieval routes from cases where no matching support is found under the observed inventory evidence. Since May 2026, IGPO has been deployed in a commercial smart-assistant AI search system. A 14-day online A/B test of the complete IGPO treatment shows a 3.17% relative CTR lift and a 38.9% reduction in audited bad cases.
1 Introduction
Early-stage AI search must adapt to rapidly changing inventories without treating catalog contents as fixed knowledge. IGPO addresses this by separating durable policy behavior from volatile inventory facts and refining Policy Guidelines through online evidence and offline exploration.
- Rapidly changing items, names, and metadata make catalogs unsuitable as static knowledge for fixed prompts or strategies.
- Inventory-aware failures may reflect absent items, missed retrieval routes, or retrieved items rejected by final selection.
- Policy Guidelines teach runtime reasoning from inventory evidence while avoiding claims that particular items remain permanently available.
- IGPO grounds queries with inventory portraits, injects relevant guidelines into retrieval and selection prompts, and refines them offline from grouped stochastic rollouts.
- IGPO is evaluated through offline replay, ablations, a 14-day online A/B test, and deployment in a commercial smart-assistant search system.
2 Related Work
Prior work addresses dynamic retrieval, prompt optimization, or inventory-aware search, but not post-deployment policy optimization under changing inventory uncertainty. IGPO instead makes inventory evidence central to policy optimization in a fixed search pipeline.
- Dynamic-collection and query-rewriting research addresses evolving corpora or lexical alignment rather than post-deployment policy optimization under inventory uncertainty.
- Weight-updating approaches require frequent model releases and stable rewards, which changing inventories undermine.
- Trace-only training-free optimization can encode transient inventory failures into brittle prompts instead of teaching runtime verification.
- IGPO treats inventory evidence as a first-class policy-optimization signal, using inventory portraits synthesized from retrieval probes.
- Figure 1 summarizes online conditioning with inventory evidence and Policy Guidelines alongside offline refinement through inventory-guided exploration.
3 Method
IGPO optimizes a fixed AI search pipeline by separating policy guidelines from changing inventory facts. It grounds online decisions in an inventory portrait and improves guidelines offline through rollout analysis, exploration, patching, and validation.
- System objective: The optimized state is a Policy Guideline store, while the model weights and retriever remain unchanged.Each guideline encodes decision behavior from runtime inventory evidence rather than asserting that specific items exist.
- Online inventory grounding: Inventory grounding probes matching items and fuses probe results with scene profiles into a query-specific inventory portrait.The portrait includes probe confidence, category-level signals, field coverage, and scene-profile information.
- Online guideline injection: Stage-matched guidelines are retrieved from an inverted index and injected into the corresponding retrieval and selection prompts.Guidelines are selected using portrait categories, stage tags, and query-guideline embedding similarity.
- Offline optimization: Offline optimization alternates trajectory grouping, inventory-guided exploration, and validated guideline updates on a fixed inventory snapshot.The loop samples rollouts, explores portrait-covered categories for missed retrieval routes, proposes patches, and checks replay and regression gates before acceptance.
- Offline optimization: Exploration re-queries the Retrieval Planner using the inventory portrait and prior observations until the request is satisfied, covered categories are exhausted, or the turn limit is reached.New candidates are merged into the accumulated set, and a successful missed route supplies contrastive evidence.
4 Experiments
Experiments evaluate IGPO through offline comparisons, cross-snapshot analysis, ablations, and a 14-day online A/B test. IGPO improves retrieval and final-result metrics, reduces boundary errors, and produces positive online outcomes under the reported configurations.
- Evaluation setup: The evaluation uses 500 optimization queries, 500 disjoint validation queries, and 3,000 temporally later final-test queries, with expert annotations independent of exploration.
- Main offline results: 0.847 Candidate Recall@30 and 0.816 F1@8 improve over 0.691 and 0.643 production-baseline values, while false no-inventory falls from 30.6% to 4.8%.False match also decreases from 60.8% to 15.2% in the native single-run comparison.
- Controlled comparisons: At approximately 120M tokens, IGPO increases Candidate Recall@30 from 0.823 to 0.850 and F1@8 from 0.791 to 0.817 versus the GEPA-style optimizer.False no-inventory decreases from 12.7% to 4.9%, and false match from 37.3% to 15.6%.
- Cross-snapshot analysis: Across inventory-change categories, F1@8 ranges from 0.761 to 0.803, while affected queries reach 0.795 F1@8 with 8.2% false no-inventory and 23.1% false-match rates.The corresponding unaffected-query values are 0.822, 3.8%, and 13.3%.
- Ablation study: The largest ablation drop occurs when Retrieval and Selection share guidelines, reducing F1@8 to 0.597; removing scene-conditioned selection raises false match to 27.5%.Removing probe statistics most degrades recall and false no-inventory among inventory-grounded ablations, while removing exploration raises both boundary errors.
- Online experiment: 3.17% relative CTR lift and 38.9% reduction in audited bad cases are observed in the 14-day online A/B test of the complete IGPO treatment.The test compares 100,400 control users with 99,800 treatment users under 50/50 user-level randomization.
5 Conclusion
IGPO is a training-free approach for AI search over dynamic inventories that grounds online decisions in inventory portraits and stage-indexed Policy Guidelines. Replay-validated offline updates and online evaluation indicate improved search quality without retraining model weights.
- IGPO grounds online decisions in inventory portraits and stage-indexed Policy Guidelines for search over dynamic inventories.
- Offline rollouts and inventory-guided exploration support replay-validated updates to the search policy.
- Offline replay and ablations improve candidate recall and F1@8 without modifying model weights.
- 3.17% relative CTR lift and 38.9% fewer audited bad cases were observed in a 14-day online A/B test.
- IGPO has been deployed since May 2026 as a practical way to adapt commercial AI search without retraining model weights.
6 Limitations
IGPO depends on the quality and coverage of runtime inventory evidence, and it cannot recover inventory or metadata absent from the underlying system. Policy Guidelines also create maintenance overhead, with longer-term scaling and cross-system maintenance outside the evaluated scope.
- IGPO depends on retrieval probes, item metadata, and the production retrieval interface to construct reliable runtime inventory evidence.
- When inventory fields, scene coverage, or retrieval routes are incomplete, the optimizer may struggle to distinguish missed routes from absent matching support.
- IGPO cannot recover inventory or metadata that is not represented in the underlying system.
- Policy Guidelines introduce maintenance overhead as inventory items, providers, and catalog schemas change.
- Evaluation of maintenance covered a 31-day observation window at a 43–64-Guideline scale, while longer-term growth and cross-system evaluation remain outside scope.
7 Ethical Considerations
The paper describes privacy-conscious handling of production-style search traces, including anonymization, access controls, limited auditor exposure, and caution around releasing potentially identifying examples.
- Production-style search interactions were anonymized and stripped of personally identifiable information before analysis.
- Data handling followed internal retention and access-control policies.
- Human auditors saw only information needed to judge search quality.
- Released examples should be anonymized or synthetic when raw queries or inventory items could reveal users or providers.
A.1 Native-Configuration Results
Table 5 presents a breadth-oriented native-configuration comparison across systems using their originally evaluated backbones and optimization budgets. It is not a controlled comparison of optimization methods.
- Table 5 reports complete native-configuration results from a single run.
- The table provides comparison breadth rather than a controlled comparison because systems retain their original backbones and optimization budgets.
- The Production Baseline and IGPO rows correspond to the native-configuration block in Table 1, while controlled blocks report separate three-run comparisons.
- ReflectiveRAG-style, Reflexion-style, ReAct-style, and GEPA-style baselines use the same frozen DeepSeek-V3 backbone, production retrieval interface, and inventory portrait as IGPO's DeepSeek-V3 configuration.
- The baseline methods differ in retrieval orchestration, use of failure feedback, or optimization of the two-stage prompts.
- GEPA-style optimization uses reflective trajectory-based prompt revisions for both Retrieval Planner and Selection stages, with round-robin module updates.
A.3 Training Details for Weight-Updating Baselines
The weight-updating baselines use Qwen3-8B with supervised fine-tuning and, for SFT+GRPO, additional group-based reinforcement training under specified reward designs and rollout settings.
- Qwen3-8B serves as the student backbone for both Retrieval-Planner-only and full-pipeline SFT and SFT+GRPO variants.
- GRPO further trains the SFT checkpoint on a 5k-query high-variance subset with group size G=8, learning rate 1×10−6, and 8 NVIDIA H800 GPUs.
- The full-pipeline SFT+GRPO reward combines Candidate Recall@30 and final outcome reward for supported Queries, while inventory-absence Queries use outcome reward alone.
- IGPO’s rollout settings use N=5 stochastic rollouts per optimization Query, while query types distinguish exact, exploratory, composite, and inventory-absence cases.
B.1 Patch Replay Validation
Patch replay validation uses code checks, per-query and aggregate regression gates, and ranked combined replay before allowing guideline changes to update the deployed store.
- Candidate Patches first undergo schema, identifier, operation, scene/stage, and conflict checks before individual replay.
- Individual replay requires every Target Query to pass win/loss and uncertainty gates alongside applicable Background and fixed-set thresholds.
- Table 6 evaluates three paired trials per Target Query, pools Background rates across Query–seed judgments, and averages per-seed metric drops equally.
- Combined replay ranks surviving Patches by Target net win rate, F1 gain, Background loss, Background uncertainty, and proposal order before sequential combination checks.
B.2 Operational Update Statistics
Operational updates run through bounded, evidence-driven Patch generation and replay, with production batches showing selective Patch acceptance and ongoing guideline-store growth.
- Operational results: Nine overnight update batches over 31 days admitted 83 candidate Patches to replay, and 45 passed individual and combined validation for application.The applied share was 54.2%.
- Operational results: The guideline store grew from 43 to 62 entries and peaked at 64, while each update batch took a median 2.2 hours.Batches averaged 10.8K LLM calls and 30.8M tokens, with incremental maintenance averaging 0.7 person-hours weekly.
- Patch generation: Patch generation uses query summaries, inventory evidence, successful and failed patterns, deployed Guidelines, and rejected-Patch summaries to propose bounded operations.The workflow distinguishes retrieval and selection stages and supports add, revise, merge, and delete operations.
- Patch generation: Patch operations are selected according to failure coverage: add introduces missing methods, revise corrects existing rules, merge resolves overlap, and delete removes persistently harmful Guidelines.
- Patch generation: Guidelines must be reusable across similar Queries, specify applicability conditions and executable actions, and avoid encoding item identities or transient inventory facts.