Get Started
Home
Topics
Search
Library
Research questionHow can production systems tune inference-time configurations online from noisy live feedback without representative validation data?Each configuration can be assessed only after serving live requests, while the resulting feedback is noisy. Without representative validation data, the system must identify effective settings from the evidence generated during operation.
AI
AI Agents
Inference Optimization
Machine Learning
Statistical Machine Learning
Technology
Latest papersRecent research connected to this question, newest first.Bandits in Prod: Hyperparameter Optimization at Inference TimeThe problem applies to production systems choosing settings such as model selection, retrieval depth, prompting strategy, or decoding temperature, including classical machine-learning models and LLM-based agents. The source formalizes the setting as an infinitely many-armed bandit and provides theoretical regret analysis alongside experiments across these configuration tasks.research paper · Sep 2, 2026
Related questions
How can functional bilevel optimization adapt online as learning objectives change over time?How can self-driving laboratories adapt long-running Bayesian optimization campaigns as scientific goals and execution conditions change?How can online time-series forecasters adapt to evolving, unseen patterns without costly continual parameter updates?How can production recommender systems continually optimize retrieval, ranking, and serving as users and content change?