DeepRule turns messy contracts and negotiation notes into structured inputs, reconciles distributor and manufacturer incentives through constrained game-theoretic optimization, and distills interpretable pricing rules — pilot retail deployments report profit gains over existing B2C baselines, though evidence rests on proprietary, non-public evaluations.

Main Finding

DeepRule proposes an end-to-end, tri-level framework that integrates LLM-based semantic parsing, game-theoretic constrained optimization, and LLM-guided symbolic regression to automatically generate auditable assortment and pricing business rules for traditional (partially digitalized) retail. Empirically, the pipeline yields higher profits than systematic B2C baselines while producing operationally feasible, interpretable rules—narrowing the gap between academic optimality and industrial operationality.

Key Points

• Problem addressed: systematic misalignment between theoretical pricing/assortment models and real-world retail complexities (unstructured textual inputs, time-varying entangled features, hierarchical business constraints, and lack of interpretability).
• Tri-level architecture:
• Hybrid Knowledge Fusion Engine — uses LLMs to semantically parse unstructured text (negotiations, approvals), convert them into structured priors/reward functions, and fuse these with managerial rules and multi-source data.
• Game-Theoretic Constrained Optimization — models manufacturer–distributor interactions with bilateral utility functions and hierarchical constraints so profit allocation becomes endogenous to optimization.
• Interpretable Decision Distillation — uses LLM-guided symbolic regression (SR) / rule search to distill black-box strategies into auditable formulaic assortment + price rules, embedding economic priors (e.g., sign/shape constraints) as hard constraints.
• Feature engineering innovations:
  • LLM-extracted priors (pprior), reward/penalty functions R(s,a), and decision sets D.
  • Customer–store affiliation aggregation using spatial (Haversine) distance, scale-weighted aggregation, and LLM similarity thresholds.
  • Dual-tower encoding to avoid dominance of high-dimensional features and style embeddings for SKU attributes (PCA(BERT) or multimodal decomposition).
• Demand and revenue modeling:
  • DNN predicts shipment units (decoupling pricing from quantity); revenue computed as units × price.
  • Regularized MSE loss; confidence-based sample cleaning with LLM-assisted posterior correction.
• Data augmentation and robustness:
  • Rule-prior-guided pseudo-labeling (RLAIF-style) with theoretical bound-style arguments on pseudo-label noise when labeling rule informativeness exceeds thresholds.
  • Multi-source weighted fusion of DNN output and rule outputs (strict vs soft rules) with dynamic weights produced by LLM.
• Assortment-pricing rule search:
  • Symbolic regression over expressive operator set; methods evaluated include evolutionary GP-style search, RL-based search, and LLM-guided reasoning with Monte Carlo search refinements.
  • Fitness objective balances profit from stocked items, penalization for deviation from market prices, and expression parsimony/complexity.
• Empirical claim: validated in real retail deployments showing profit improvements and operational feasibility (details deferred to appendix).

Data & Methods

• Data modalities:
  • Structured: transactional histories, SKU attributes, geolocation, inventory, promotional tags.
  • Unstructured: distributor agreements, approval docs, negotiation logs, sales assessments — parsed by locally deployed LLMs into structured vectors (pprior), reward functions R(s,a), and datasets D.
• Feature pipeline:
  • LLM semantic parsing → structured priors; spatial aggregation with Haversine distances and scale weights; temporal features via window stats and Fourier coefficients; dual-tower feature encoding (high-dim / low-dim fusion).
  • SKU style embeddings via PCA(BERT) or multimodal projection; final SKU vector eSKU concatenates style, sparse, dense parts.
• Prediction model:
  • DNN fθ on input tensor X = [Xcust ⊕ Xsku ⊕ Xprice ⊕ Xpromo] predicting units; revenue computed post-prediction.
  • Training uses regularized MSE and cross-validation for fusion weights (β).
• Data augmentation & cleaning:
  • LLM pseudo-labeling (RLAIF-inspired) with controlled noise analysis (pseudo-label bias bounds tied to rule informativeness).
  • Confidence metric identifies low-confidence samples; LLM produces (judgment, reason, rule) to correct/flag samples; ambiguous cases sent for human verification.
• Rule-guided decision module:
  • Symbolic regression search over functions f: R^n → {0,1}×R+ with h(x) (assortment) and g(x) (price).
  • Search algorithms: evolutionary operators (structural expansion/pruning, param perturbation), RL-based policy for expression construction, and LLM/Monte-Carlo-guided tree search.
  • Fitness Φ(f) includes realized profit term, deviation penalty, and complexity regularizer; economic priors enforced as hard constraints during search (e.g., non-positive/ non-negative elasticity).
• Constrained optimization:
  • Game-theoretic layer models manufacturer and distributor utilities, encodes multi-tier business rules as hierarchical constraints; uses distributed optimization architecture to scale to many decision variables.
• Validation:
  • Real retail environment experiments (claimed profit improvements vs B2C baselines and operational constraint satisfaction). Paper refers to appendix for formal mathematical details and additional experimental setups.

Implications for AI Economics

• Integrating unstructured textual data into demand/price models:
  • DeepRule demonstrates a practical pathway to harvest economic priors and negotiation/contractual information from text via LLMs, improving personalization and heterogeneity capture in low-digitalization settings.
  • Quantifies the value of non-transactional information for pricing strategies and suggests new data sources for empirical IO analyses.
• Multi-agent endogenous objectives:
  • Modeling manufacturer–distributor interactions endogenously within optimization enables the study of profit allocation mechanisms, contract design, and incentive compatibility in data-driven pricing—bridging mechanism design and ML-based pricing.
• Interpretability and governance:
  • Symbolic-rule distillation with economic priors produces auditable strategies, making algorithmic pricing more amenable to managerial validation, regulatory review, and cross-organization adoption.
  • Embedding hard economic constraints into SR reduces risk of economically implausible rules (e.g., wrong elasticity signs).
• Industrial adoption and operational feasibility:
  • The closed-loop pipeline (LLM ingestion → predictive models → constrained optimization → rule distillation → human-in-the-loop validation) addresses real-world operational constraints and may improve uptake in legacy retail systems.
• Research directions and policy questions:
  • Measuring welfare/competitive impacts of automated rule generation in multi-tier supply chains (does automated profit redistribution create new market power concerns?).
  • Empirical evaluation of how much unstructured text improves welfare, pricing efficiency, or consumer surplus.
  • Need for transparency standards around LLM-derived priors in economic decision-making.
• Limitations and cautionary notes:
  • Reliance on LLM quality and local deployment: biases, hallucinations, and label noise from LLMs can propagate into economic decisions; the paper proposes controls but further robustness analysis is needed.
  • Lack of formal theoretical guarantees (e.g., regret bounds) for the combined pipeline; many components (symbolic search, DNN prediction with pseudo-labeling, game-theoretic solver) are heuristic and engineering-heavy.
  • Computational and scaling costs (LLM parsing, SR search over large operator spaces) may be substantial for very large SKU universes; distributed optimization helps but costs were not fully quantified.
  • Generalizability: approach targets partially digitized traditional retail; performance in fully online marketplaces or different institutional contexts requires evaluation.
• Suggested follow-ups for AI economics research:
  • Formal quantification of information value from textual sources (value-of-information experiments).
  • Robustness and adversarial analysis of LLM-extracted priors in market settings.
  • Theoretical analyses linking SR-constrained solutions to welfare/regret benchmarks.
  • Study of incentive compatibility and strategic responses when multiple downstream agents adopt similar automated rule-generation systems.

Overall, DeepRule is a pragmatic, systems-level contribution that demonstrates how LLMs, predictive ML, and symbolic-rule search can be composed to produce economically sensible, auditable pricing and assortment rules for operational retail settings—opening multiple empirical and theoretical avenues for AI economics.