verdi: retrieval is not transfer for continual world model optimization
2026-08-10 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors address the challenge of improving pretrained world models for specific goals, noting that past methods often start optimization from scratch without safely reusing prior strategies. They propose VERDI, a system that carefully tests and validates whether strategies from previous models can help new ones before applying them. VERDI uses probes to understand each model, ranks past optimization ideas, and only adopts those that pass strict checks. Their experiments show VERDI reduces computing costs and mistakes when transferring strategies between models.
foundation world modelsoptimizationtransfer learningcontinual learninginference probesexperimental validationnegative transferCtrl-WorldCosmos modelsGPU cost
Authors
Junyu Wu, Shiqin Nie, Youyi Kou, Baohua Yin, Guocai Yao, Qingyu Chen, Jingheng Ma, Shiji Zhou, Hongyong Song, Mingchen Zhuge, Sen Cui, Changshui Zhang
Abstract
Foundation world models have made remarkable progress in planning, simulation, and embodied intelligence. However, optimizing a pretrained world model toward a user-specified objective remains difficult: each campaign typically rediscovers optimization strategies from scratch, and the resulting knowledge rarely transfers to the next model. Existing research agents automate the optimization loop but treat successful strategies as directly reusable recipes, without principled safeguards for when transfer is appropriate. We argue instead that retrieval is not transfer: a strategy validated on one model is at best an optimization hypothesis for another, and becomes transferable knowledge only after target-side experimental valida- tion. Guided by this principle, we propose VERDI , a continual framework for evidence-licensed world model optimization. VERDI characterizes each world model through shared inference-time probes to construct an Optimization Fin- gerprint, retrieves relevant prior experience as ranked hypotheses, and validates every candidate under a frozen target-side verifier before admitting it as reusable evidence; contradictions among nearby fingerprints further trigger probe evolution, continually refining the diagnostic representation itself. Experiments on Ctrl-World, the Cosmos family, and RoboCoin show that VERDI reduces search cost by 68%, GPU cost by 69%, and negative transfer from 0.34 to 0.06, while predicting transfer outcomes with 83% sign accuracy.