EEVEE

Test-time prompt learning in the real world for self-improving agents

EEVEE: Towards Test-time Prompt Learning in the Real World for Self-Improving Agents

Princeton/AI2 Collaboration Paper: arXiv:2606.11182 Project Page Code

EEVEE is a multi-dataset test-time prompt learning framework for LLM agents. It targets mixed real-world task streams with different datasets, domains, formats, and evaluation rules, where agents need to keep learning and adapting as contexts change.

Key Ideas

  • Learned routing assigns each input to a specialized prompt slot.
  • Router-prompt co-evolution jointly improves routing behavior and prompt behavior.
  • Cross-dataset interference reduction makes prompt learning more robust under mixed task streams.
  • Practical self-improvement helps agents adapt across heterogeneous task streams while preserving useful learned behavior.

Research Direction

EEVEE connects self-improving agents with a broader definition of intelligence. Strong agents should analyze incoming contexts, learn how to specialize, adapt across heterogeneous tasks, and preserve useful behavior over time.

References