VERDI:Verified Contradictions for Continual World-Model Optimization

VERDI: Retrieval Is Not Transfer for Continual World Model Optimization

Abstract Summary:

Foundation world models have made progress in planning, simulation, and embodied intelligence, but 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. The paper argues 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 validation. The authors 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 Fingerprint, 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.

Introduction Summary:

World model optimization remains fundamentally non-cumulative: each optimization campaign typically starts from a pretrained world model, explores numerous optimization strategies, validates those that improve the target objective, and terminates once the objective is achieved. While this process continuously produces valuable optimization knowledge, including successful optimization strategies, failed attempts, empirical evidence, and optimization trajectories, such knowledge is rarely reused beyond the current model. A key challenge is that optimization knowledge is inherently model-dependent, yet existing systems lack a principled way to relate heterogeneous world models from the perspective of optimization. Similar architectures do not necessarily exhibit similar optimization behaviors, while architecturally different world models may benefit from remarkably similar optimization strategies. The authors hypothesize that these characteristics can be actively measured through a shared set of inference-time probes, constructing an Optimization Fingerprint for each world model and defining an Interventional Repair Geometry (IRG) over these fingerprints.

Method Summary:

VERDI formulates continual world model optimization as a four-stage process: Characterize → Retrieve → Validate → Accumulate. Given a target world model and a user-specified optimization objective, VERDI first characterizes the model through a shared library of inference-time probes, producing an Optimization Fingerprint that captures optimization-relevant behavior. The fingerprint is then related to previously optimized world models through the Interventional Repair Geometry (IRG), allowing the system to retrieve candidate Optimization Skills as optimization hypotheses. Rather than assuming retrieved knowledge is transferable, VERDI validates every hypothesis through target-side optimization experiments under a frozen verifier. Only experimentally settled evidence, including confirmed positive, null, harmful, and interaction outcomes, is incorporated into the World Model Optimization Memory (WMOM).

The paper defines a differential repair contract: A differential repair is a typed intervention whose claim is defined by a paired change vector rather than by an absolute score. A repair is accepted only when the paired utility improvement exceeds the threshold and all protected constraints pass. The Optimization Fingerprint is defined as φx = (Jx, Gx, Σx), where Jx is the response Jacobian, Σx is the bootstrap covariance, and Gx is an uncertainty-aware metric. IRG defines optimization similarity through shared responses to controlled inference interventions rather than implementation similarity.

The central safeguard is that no retrieved skill becomes knowledge merely because it is close in IRG. The compatibility certificate checks six conditions: compilation, support overlap, evidence sufficiency, response alignment, sign agreement, and gain over cold start. The certificate prioritizes an experiment but never accepts a target-side effect as valid under the target-side verifier.

Key Results:

• Failure-specific repair on ACWM-Phys: The selected repairs are mechanism-specific, with the largest metric changes occurring for robot arm (+1.08 dB PSNR), cloth move (+0.94 dB), pour water (+0.85 dB), and push sand (+0.64 dB).

• Concrete repair slices: On Ctrl-World, the repair improves motion smoothness from 68.75 to 81.53 while reducing dynamic degree from 45.32 to 43.78. On Cosmos-Predict 2, depth accuracy improves from 95.35 to 97.05 while background consistency drops from 85.35 to 82.73.

• Counterexample-driven probe evolution: A counterexample on cloth move triggers a typed proposal for the action-embedding temporal mix probe, reducing locality residual from 3.88 to 0.0196, and increasing local-chart coverage from 5/8 to 6/8. Collision-driven growth admits only three additional directions and improves both held-out selection regret and calibration.

• Cross-backbone transfer: Certified reuse matches unconditional speed (3.6 trials to positive vs. 4.1) while reducing negative transfer from 0.34 to 0.06, and reducing GPU cost from 312 to 96 GPU-h. The complete intervention fingerprint reaches 0.83 sign agreement, 3.6 trials to positive, 0.06 negative transfer, and 0.03 protected violation.

Limitations:

VERDI's guarantees are conditional on a complete goal contract, a valid evaluator, and a correct evidence policy. Fingerprints are local and must be recomputed after major architectural changes. Early archives inherit the limits of observational priors and single-laboratory coverage. RoboCoin covers one dataset family and base model; it does not establish adaptation to arbitrary embodiments. H/I/E single-axis controls reduce confounding but do not by themselves establish causal identification.

Conclusion:

The authors introduced VERDI, a verifiable differential-repair architecture. IRG ranks typed repair hypotheses across backbones; compilation, the six-term certificate, and target-side verification decide whether transfer is allowed. Settled positive, null, harmful, and interaction outcomes update WMOM, while failed transfers refine the diagnostic basis. They demonstrated failure-specific repair, held-out certified transfer, and RoboCoin adaptation. Broader model families, uncertainty calibration, and contradiction-aware scheduling remain important directions for future work.

Improvements to AI Systems:

1. Evidence-Gated Transfer Mechanism: Implement a system where any retrieved optimization strategy from a prior model is treated as a hypothesis, not a recipe. The system must run a target-side validation experiment under a frozen verifier before accepting the strategy as knowledge. This prevents negative transfer (e.g., reducing negative transfer from 0.34 to 0.06) and ensures that only experimentally settled outcomes—positive, null, harmful, or interaction—are added to a persistent memory.

2. Optimization Fingerprint for Model Characterization: Build a diagnostic module that computes a shared set of inference-time probes (response Jacobian, bootstrap covariance, uncertainty-aware metric) for any new model. This fingerprint enables the system to relate heterogeneous models by optimization-relevant behavior rather than architecture, allowing it to retrieve relevant prior skills even when models differ structurally.

3. Interventional Repair Geometry (IRG) for Skill Ranking: Use IRG to rank candidate optimization skills by shared responses to controlled interventions, not implementation similarity. This improves retrieval precision and reduces search cost by 68% and GPU cost by 69% compared to unconditional reuse.

4. Six-Term Compatibility Certificate: Before any transfer, the system checks: compilation, support overlap, evidence sufficiency, response alignment, sign agreement, and gain over cold start. This certificate prioritizes experiments but never assumes transfer validity, ensuring that only compatible repairs are attempted.

5. Counterexample-Driven Probe Evolution: When a retrieved skill fails on the target model, the system automatically proposes new diagnostic probes (e.g., action-embedding temporal mix) to refine the fingerprint. This reduces locality residual from 3.88 to 0.0196 and increases local-chart coverage, enabling the system to continually improve its understanding of model behavior and avoid repeated failures.

6. Failure-Specific Repair Selection: The system can identify mechanism-specific repairs (e.g., improving robot arm PSNR by +1.08 dB, cloth move by +0.94 dB) rather than applying generic fixes. This allows for targeted optimization of sub-objectives within a world model, such as improving motion smoothness while controlling dynamic degree.

7. Trade-off-Aware Optimization: The system explicitly tracks and reports trade-offs (e.g., depth accuracy improves from 95.35 to 97.05 while background consistency drops from 85.35 to 82.73). This enables users to make informed decisions about which objectives to prioritize and when to stop optimization.

8. Contradiction-Aware Memory Management: The system detects contradictions among nearby fingerprints and triggers probe evolution, preventing the archive from accumulating stale or misleading knowledge. This maintains the reliability of the World Model Optimization Memory over time.

9. Calibrated Uncertainty in Transfer Predictions: With 83% sign accuracy in predicting transfer outcomes, the system can forecast whether a retrieved skill will help or harm before running expensive experiments, allowing it to allocate compute resources more efficiently.

10. Continual Learning Without Catastrophic Forgetting: By accumulating settled evidence (positive, null, harmful, interaction) in a structured memory, the system improves over time—each new optimization campaign benefits from prior campaigns' validated knowledge, making the overall process cumulative rather than starting from scratch.

What the Improved AI System Can Do:

• Optimize a pretrained world model toward a user-specified objective with 68% less search and 69% less GPU cost, while reducing negative transfer by 82% (from 0.34 to 0.06).

• Automatically decide when to reuse a prior optimization strategy and when to run a new experiment, based on a rigorous compatibility certificate.

• Adapt to new world models (e.g., different backbones or datasets) by computing their optimization fingerprint and retrieving only relevant, validated skills.

• Detect and correct its own diagnostic limitations by evolving probes when counterexamples arise, improving its ability to characterize novel models.

• Provide transparent trade-off reports (e.g., accuracy vs. consistency) so users can make informed decisions about optimization targets.

• Maintain a growing, evidence-based memory of optimization knowledge that becomes more reliable and comprehensive with each campaign, enabling faster and safer optimization on future tasks.

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.