In-Context Learning for Data-Driven Censored Inventory Control

2026-05-14 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Optimization & Control · Depth: Expert, quick

Summary

A new method called in-context generative posterior sampling (ICGPS) is proposed for inventory control problems with decision-dependent censoring, specifically addressing the repeated newsvendor (R-NV) problem. This approach combines generative models, meta-trained offline, with online in-context autoregressive generation to make oracle actions on learned latent demand completions. ICGPS aims to overcome the brittleness of parametric Thompson sampling (TS) under prior mismatch and the transfer issues of offline imputation methods. Theoretically, ICGPS's Bayesian regret is bounded by a TS benchmark plus a deployment penalty related to completion mismatch, yielding sublinear Bayesian regret for R-NV. The method's online performance is linked to offline predictive quality. Practically, ICGPS is instantiated with ChronosFlow, a time-series transformer combined with a conditional normalizing-flow head, demonstrating performance matching correctly specified TS, outperforming myopic and UCB baselines, and showing robustness to prior mismatch and distribution shift on benchmark and real-world SuperStore datasets.

Key takeaway

For Machine Learning Engineers developing inventory control systems with censored demand, ICGPS offers a robust alternative to traditional Thompson sampling. Your teams should consider implementing ICGPS, particularly with ChronosFlow, to achieve performance comparable to well-specified TS while gaining resilience against prior mismatch and distribution shifts, especially in scenarios with heavy censoring.

Key insights

ICGPS uses generative models and in-context learning to improve inventory control under censored demand.

Principles

• Offline predictive quality transfers to online performance.
• Generative models can complete latent demand for decision-making.

Method

ICGPS meta-trains generative models offline, then deploys them online for in-context autoregressive generation to infer latent demand and guide inventory decisions.

In practice

• Instantiate ICGPS with ChronosFlow for time-series data.
• Apply ICGPS to inventory control with censored demand.

Topics

• In-Context Learning
• Censored Inventory Control
• Repeated Newsvendor Problem
• Generative Models
• Bayesian Regret

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Regime-Calibrated Demand Priors for Ride-Hailing Fleet Dispatch and Repositioning — Matching current demand to historical regimes significantly reduces ride-hailing wait times without complex forecasting models.
• Can In-Context Learning Support Intrinsic Curiosity? — ICL can support intrinsic curiosity for data selection in non-temporal settings, but not general MDPs.
• Beyond Importance Sampling: Rejection-Gated Policy Optimization — RGPO uses a differentiable acceptance gate to selectively trust samples, ensuring bounded gradient variance in policy optimization.
• UC-Search: Risk-Aware Test-Time Search for Delayed Constrained Time-Series Control — UC-Search provides a risk-aware test-time search framework for delayed, constrained time-series control decisions.
• MINTS: Minimalist Thompson Sampling — MINTS simplifies Bayesian decision-making by focusing priors and profiling nuisance parameters to handle structural constraints effectively.
• Learning from Failures: Correction-Oriented Policy Optimization with Verifiable Rewards — CIPO enhances LLM reasoning by converting failed trajectories into explicit, directional correction signals for policy optimization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.