GRADUS: Discovering and Recovering Agent Lessons from the Environment's Latent Syllabus

Key points

• Training LLM agents for policy-constrained tool environments, such as an airline backend with booking, refund, and cancellation policies, requires tasks that exercise the environment's rules individually and in combination.1
• Existing synthesis pipelines let a model propose tasks and retain only those it immediately solves, biasing training data toward easier behaviors.1
• Without seed tasks, Gradus treats a fixed environment as a latent syllabus, extracting guarded behavioral operators that capture how actions depend on current conditions and how their effects reshape subsequent decisions.1
• Crucially, Gradus jointly calibrates tasks and solutions against the environment: execution feedback identifies which side to refine while preserving the intended behavioral requirements.1

Sources (1)

• [1]GRADUS: Discovering and Recovering Agent Lessons from the Environment's Latent Syllabus
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 07:47 AM

  We introduce Gradus (Graded Rungs of Ascending Difficulty from an Underlying System), which decouples what a task should teach from the concrete task and solution used to teach it.

  Training LLM agents for policy-constrained tool environments, such as an airline backend with booking, refund, and cancellation policies, requires tasks that exercise the environment's rules individually and in combination.

Extractive summary: sentences quoted from the sources.