verdi: retrieval is not transfer for continual world model optimization

2026-08-10

AI Summary

This work addresses the inefficiency of existing pre-trained world models in goal-oriented optimization, which often suffer from inadequate knowledge transfer mechanisms, leading to redundant exploration and wasted experience. To overcome this, the authors propose the VERDI framework, which introduces the principle that “retrieval does not imply transfer.” VERDI constructs optimization fingerprints to retrieve historical policies, generates hypotheses from them, and employs a frozen validator to filter only empirically valid knowledge for transfer, enabling evidence-driven continual improvement. The approach integrates shared inference-time probes, validation-based hypothesis selection, and probe-contrastive representation evolution, substantially enhancing transfer efficiency and reliability. Experiments on Ctrl-World, Cosmos, and RoboCoin demonstrate a 68% reduction in search cost and a 69% reduction in GPU usage, while negative transfer drops from 0.34 to 0.06 and symbolic prediction accuracy reaches 83%.

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.

Problem

• world model optimization
• knowledge transfer
• continual learning
• retrieval
• negative transfer

Innovation

• continual optimization
• world model
• evidence-licensed transfer
• optimization fingerprint
• negative transfer