Preprint: MetaExp enables cross-task experience reuse in LLMs, boosting automation benchmarks by up to 14.8 points
New framework proves experience learning can become a transferable capability for LLM agents, enabling post-deployment improvement without retraining.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Whether language model agents can convert past interactions into reusable experience is a key bottleneck for continuous evolution after deployment. MetaExp introduces a scalable framework that trains experience learning itself as a general model capability through verifiable task families with controlled cross-task structures. The design employs two complementary mechanisms: Shared-Tool fixes black-box tools to elicit operational experience, while Shared-Workflow preserves latent procedures to elicit procedural experience.
The research team collected and filtered 4,956 sequential interaction trajectories from 1,000 task families to fine-tune Qwen3.5-122B-A10B. Results show that with parameters completely frozen during downstream evolution, MetaExp improves by 8.5 points on SpreadsheetBench, peaks at 2.3 on APEX-Agents, and gains 14.8 points on AutomationBench. These improvements persist even under domain-excluded training settings, supporting experience learning as an effective non-parametric continuous improvement capability.
This result is currently an author-run benchmark (E2 grade) and has not been independently reproduced. The paper does not disclose specific hardware configurations or absolute baseline scores, so the practical engineering significance of the +14.8 point gain must be evaluated against original task difficulty. Additionally, the method relies on pre-constructed verifiable task families; its generalization boundary in open-world unsupervised scenarios remains to be validated.