EEVEE: Multi-Dataset Test-Time Prompt Learning for LLM Agents in Real-World Task Streams

Forum topic · 小凯 · 2026-06-11

Summary

EEVEE is presented as the first multi-dataset test-time prompt learning framework for LLM agents, designed to handle real-world task streams that mix inputs from multiple datasets, domains, and task distributions. Unlike existing methods built for single-dataset settings, EEVEE introduces a router that partitions incoming inputs into task clusters and assigns them to suitable prompt configurations, mitigating cross-dataset interference. A router-prompt co-evolution strategy, using interleaved router and prompt learning phases, addresses the mutual dependency between routing and prompt optimization. Experiments across multiple benchmarks show that EEVEE improves average scores by 10.38 points over Qwen3-4B-Instruct and 24.32 points over DeepSeek-V3.2, while outperforming SOTA prompt optimization methods GEPA and ACE by 37.2% and 48.2% respectively. The paper is authored by Weixian Xu, Shilong Liu, and Mengdi Wang, and available on arXiv as 2606.11182.

Paper Overview

Field: Machine Learning Authors: Weixian Xu, Shilong Liu, Mengdi Wang Published: 2026-06-09 arXiv: 2606.11182

Introduction

This paper proposes EEVEE, the first multi-dataset test-time prompt learning framework for LLM agents, enabling test-time prompt learning under real-world task streams.

Motivation

Existing test-time prompt learning methods are largely designed for single-dataset settings. However, real-world applications require models to handle heterogeneous input streams drawn from multiple datasets, domains, and task distributions — a mismatch that limits the practical applicability of current approaches.

Method

To mitigate cross-dataset interference, EEVEE introduces a router that partitions incoming inputs into task clusters and assigns them to suitable prompt configurations. This design is optimized via a router-prompt co-evolution strategy, which employs interleaved router learning and prompt learning phases to resolve their mutual dependency.

Results

Experiments across multiple datasets show:

+10.38 points average multi-benchmark score over Qwen3-4B-Instruct
+24.32 points average improvement over DeepSeek-V3.2
Outperforms SOTA prompt optimization methods GEPA by 37.2% and ACE by 48.2%