Bonjour à tous,

Vous êtes tous et toutes cordialement invité.es à assister à la présentation de projet du prédoc III de Yihong Wu, le 3 septembre à 9h sur zoom.

Titre : Efficient LLM Agent Systems: From Learning to Inference

Date: Jeudi 3 septembre à 9h

Location: https://umontreal.zoom.us/j/81072799275?pwd=6jrCmbhDGSKPmeQTTiLEW1oNLWzg2A.1

Jury

Résumé

Recent large language model (LLM) agents have demonstrated remarkable capabilities in solving complex tasks, ranging from repository-level software development to mathematical reasoning and conjecture solving. Such tasks are often long-horizon, requiring hours of interaction and hundreds of thousands of tokens. However, supporting agents at this scale remains challenging because of the substantial computational and memory costs associated with long-context modeling, reinforcement learning, and autoregressive inference.

In this report, we investigate how to efficiently train and serve long-horizon LLM agents. On the training side, we discuss efficient long-context training and reinforcement learning techniques, including reducing the group size in GRPO and using asynchronous rollouts to improve training efficiency. On the inference side, we examine model quantization, KV-cache management, and agent-harness techniques such as subagent orchestration and context compression.

More broadly, we argue that efficient LLM agent systems require a three-way co-design of learning algorithms and model architectures, serving systems, and agent harnesses. Optimizing these components independently can overlook important interactions among them, whereas jointly designing the three layers provides new opportunities to improve the scalability and efficiency of long-horizon agents.