Workshop "Research in Action"

Abstract

This workshop brings together researchers and practitioners advancing the reliability, efficiency, and evaluability of LLM systems and agents in real-world applications. Speakers present procedural memory in LLM agents to enable reusable, adaptable skills across enterprise tasks; techniques for producing reliable answers; strategies for tabular data classification using LLMs; and a framework for automatically generating LLM-based judges from execution traces to streamline evaluation of agentic pipelines. Emphasis is placed on turning LLM capabilities into robust, measurable components through controlled skill transfer, uncertainty-aware decoding, data-efficient learning, and trace-driven evaluation, so that business can deploy agents with greater confidence, lower annotation cost, and clearer performance guarantees.