MetaExp: Learning to Self-Improve through Experience via Verifiable Cross-Task Environment Scaling

Key points

• Existing methods either optimize external experience scaffolds or train experience-related behavior on specific task distributions, leaving open whether experience learning itself can become a transferable model capability.1
• In this work, we introduce MetaExp, a scalable framework that trains this capability through verifiable task families with controlled cross-task structure.1
• We propose two complementary designs: Shared-Tool families keep black-box tools fixed across tasks to elicit operational experience, while Shared-Workflow families preserve latent procedures across changing realizations to elicit procedural experience.1
• We scale environments under both designs, collect and filter sequential interaction trajectories, and fine-tune Qwen3.5-122B-A10B on 4,956 trajectories from 1,000 task families.1

Sources (1)

• [1]MetaExp: Learning to Self-Improve through Experience via Verifiable Cross-Task Environment Scaling
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 08:29 AM

  Language-model agents can accumulate interaction feedback, but reliably turning that feedback into reusable experience remains challenging.

  Existing methods either optimize external experience scaffolds or train experience-related behavior on specific task distributions, leaving open whether experience learning itself can become a transferable model capability.

Extractive summary: sentences quoted from the sources.