Summary (Overview)

• Core thesis: VERDI challenges the assumption that retrieval of optimization strategies from one world model can be directly transferred to another. The paper argues that retrieval is not transfer—a strategy validated on one model is merely an optimization hypothesis for another, requiring target-side experimental validation before becoming reusable knowledge.
• Key contribution: Introduction of the Optimization Fingerprint, a representation of world models based on shared inference-time probes, and the Interventional Repair Geometry (IRG) for measuring optimization-relevant similarity between heterogeneous world models.
• Framework: VERDI operates as a four-stage continual loop—Characterize → Retrieve → Validate → Accumulate—where retrieved knowledge is treated as hypotheses, validated under frozen target-side verifiers, and only settled evidence (positive, null, harmful, or interaction outcomes) is accumulated in a World Model Optimization Memory (WMOM).
• Results: 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 across Ctrl-World, Cosmos, and RoboCoin.
• Novel mechanism: Probe evolution triggered by IRG collisions—when nearby fingerprints yield contradictory validated effects—continually refines the diagnostic representation itself, not just the candidate pool.

---

Introduction and Theoretical Foundation

Background

Foundation world models have become essential for embodied intelligence, enabling planning, policy learning, synthetic data generation, and simulation through predictive modeling of future observations. As models diversify in architecture, scale, and capability (e.g., NVIDIA's Cosmos family, Ctrl-World), improving pretrained world models toward user-specified objectives has become increasingly important. Recent autonomous research agents can automate much of the optimization pipeline—experiment generation, implementation, execution, and evaluation—but each campaign typically starts from scratch.

The Core Problem

The paper identifies a fundamental bottleneck: world model optimization is non-cumulative. Each campaign rediscovers optimization strategies, and the resulting knowledge rarely transfers to subsequent models. This is because:

• Optimization knowledge is inherently model-dependent
• Similar architectures do not necessarily exhibit similar optimization behaviors
• Architecturally different models may benefit from remarkably similar strategies

Theoretical Foundation

The key insight is that effective knowledge transfer requires comparing world models according to their optimization-relevant characteristics rather than implementation details. The authors hypothesize these characteristics can be actively measured through shared inference-time probes, yielding an Optimization Fingerprint for each model.

The paper formalizes a world model $M_\theta$ producing a rollout:

$$\hat{\tau}_{1:H} = \mathrm{Rollout}(M_\theta; \mathbf{c}, \mathbf{u}),\tag{1}$$

where $\mathbf{c}$ denotes context and $\mathbf{u}$ denotes optional actions/controls. An optimization campaign $\mathcal{C}_x = (M_x, \mathcal{D}_x, \mathcal{G}_x, V_x, B_x)$ fixes a target model, permitted data, user goal, frozen verifier, and nominal resource allowance.

Central to the framework is the Differential Repair Contract—a typed intervention defined by a paired change vector rather than an absolute score:

$$\Delta \mathbf{y}_x(a) = \mathbf{y}_x(a) - \mathbf{y}_x(0),\tag{2}$$

where $\mathbf{y}_x(a)$ collects target utility and protected outcomes after applying skill $a$, and $\mathbf{y}_x(0)$ denotes the matched control. A repair is accepted only when $\Delta U_x(a) > \tau_{\mathcal{G}_x}$ (the improvement threshold) and all protected constraints pass.

---

Methodology

Stage 1: Characterize

VERDI measures how each model responds to a shared library of typed, inference-only probes. A probe $q \in \mathcal{Q}$ applies a reversible semantic perturbation with dose $\alpha$ (e.g., modifying action magnitude, context retention, first-frame anchoring, or sampler noise) without changing model parameters or the optimization objective.

The local response for outcome $k$ under probe $q$ is:

$$J_x[k, q] = \mathbb{E}_{\xi, \omega}\left[\frac{z_k(M_x^{q,\alpha}; \xi, \omega) - z_k(M_x; \xi, \omega)}{\alpha}\right],\tag{3}$$

with bootstrap uncertainty $\Sigma_x$. The Optimization Fingerprint is then:

$$\phi_x = (J_x, G_x, \Sigma_x),\tag{5}$$

where $G_x = J_x^{\top} W_x J_x$ is an uncertainty-aware metric (Eq. 4) with $W_x$ standardizing outcome scale and discounting uncertain coordinates.

Stage 2: Retrieve

Interventional Repair Geometry (IRG) defines optimization similarity through shared responses to controlled interventions. VERDI aggregates source effects using IRG proximity, source evidence quality $\rho_s(a)$, and context overlap $\kappa_{ts}$:

$$\widehat{\Delta U}_t(a) = \frac{\sum_{s \in \mathcal{N}_t(a)} w_{ts}(a) r_s^U(a)}{\sum_{s \in \mathcal{N}_t(a)} w_{ts}(a)}, \qquad w_{ts}(a) \propto \frac{\rho_s(a) \kappa_{ts}}{\epsilon + \mathrm{dist}_{\mathrm{IRG}}(\phi_t, \phi_s)}.\tag{6}$$

The resulting ranked set $\mathcal{H}_t$ consists of optimization hypotheses, not transferable facts.

Stage 3: Validate

The Compatibility Certificate determines whether a skill receives a transfer-prioritized trial (never certifying target-side effect):

$$\begin{array}{l} C_t(a) = \mathbb{I}[\text{compile} = 1] \mathbb{I}[\text{overlap} \geq \rho_{min}] \mathbb{I}[N_{\text{eff}} \geq n_{\text{min}}] \\ \quad \cdot \mathbb{I}[e_{\text{align}} \leq \epsilon_{\text{align}}] \mathbb{I}[\mathrm{sign\_agree} \geq \gamma] \mathbb{I}[L_a(\phi_t) > \delta_{\mathcal{G}_t}] \\ \quad \cdot \prod_j \mathbb{I}[\hat{v}_{t,j}(a) \leq 0], \end{array}\tag{7}$$

The six terms correspond to failure modes: compilation failure, insufficient support overlap, insufficient evidence, response misalignment, sign disagreement, and insufficient gain over cold start.

All candidates undergo target-side experimental validation under the frozen verifier: low-cost paired screens first, then held-out confirmation with increasing rollout horizons, optimization steps, and random seeds.

Stage 4: Accumulate

IRG Collision Discovery identifies when two nearby campaigns yield incompatible validated effects:

$$\operatorname{dist}_{\mathrm{IRG}}\left(\phi_x, \phi_y\right) \leq \epsilon_{\mathrm{IRG}}, \quad \operatorname{LCB}\left(-r_x^U(a) r_y^U(a)\right) > 0.\tag{8}$$

Probe Evolution then proposes new probes to resolve collisions, selecting:

$$q^* = \arg \max_{q \in \mathcal{Q}_{\text{cand}}} \frac{\operatorname{Regret}(\mathcal{Q}) - \operatorname{Regret}(\mathcal{Q} \cup \{q\})}{\operatorname{Cost}(q)},\tag{9}$$

subject to a frozen regression battery and minimum information-gain threshold.

---

Empirical Validation / Results

Failure-Specific Repair on ACWM-Phys

VERDI selects mechanism-specific repairs for different failure families: guidance scheduling for cube/contact environments, self-forcing for articulated motion, and next-forcing for deformable/granular/fluid cases. Largest improvements include 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

Table 1: Concrete failure-specific repairs across world models

Table 2: Ctrl-World on WorldArena

Counterexample-Driven Probe Evolution

An initial action-dimension anisotropy probe was local in 5/8 environments. A counterexample on cloth_move (locality residual 3.88) triggered a typed proposal for the action-embedding temporal mix probe, reducing residual to 0.0196. The successor was not globally substituted (on reacher, residual was 0.7032), demonstrating conditional atlas expansion.

Table 3: Atlas-evolution policies on nested replay

Cross-Backbone Transfer

Table 4: Held-out cross-backbone transfer

Certified reuse matches unconditional speed while reducing negative transfer from 0.34 to 0.06. The complete intervention fingerprint achieves 0.83 sign agreement, 3.6 trials to positive, 0.06 negative transfer, and 0.03 protected violation.

---

Theoretical and Practical Implications

Theoretical Implications

1. Retrieval ≠ Transfer: The paper establishes a principled distinction between retrieving optimization knowledge and transferring it. Retrieved strategies are hypotheses requiring target-side validation—a safeguard against harmful "unconditional reuse."

2. Optimization-Relevant Representation: Optimization Fingerprints based on inference-time probes provide a more reliable basis for relating heterogeneous world models than architectural similarity, since similar architectures don't guarantee similar optimization behaviors.

3. Continual Representation Refinement: IRG collisions serve as counterevidence that the probe basis omits transfer-relevant distinctions, driving probe evolution that improves the diagnostic representation itself—not just the candidate pool.

4. Differential Effect Vectors: Representing repairs as paired, multi-coordinate effects (rather than absolute scores) makes trade-offs explicit before transfer decisions, preventing protected-coordinate regressions from being hidden by averaged scores.

Practical Implications

1. Cost Reduction: VERDI reduces search cost by 68% and GPU cost by 69%, making world model optimization substantially more efficient across campaigns.

2. Reliable Transfer: The compatibility certificate with explicit abstention prevents harmful transfer (negative transfer reduced from 0.34 to 0.06) while maintaining speed gains.

3. Knowledge Accumulation: The WMOM/IEM archive enables systematic reuse of settled positive, null, harmful, and interaction outcomes across heterogeneous world model families.

4. Autonomous Optimization: VERDI provides an autonomous framework that can be deployed for continual world model improvement without requiring per-campaign manual effort.

---

Conclusion

VERDI introduces a verifiable differential-repair architecture for continual world model optimization. The framework characterizes world models through Optimization Fingerprints, relates them via Interventional Repair Geometry, retrieves prior experience as ranked hypotheses, validates every candidate under frozen target-side verifiers, and accumulates settled evidence in a World Model Optimization Memory.

Key experimental findings demonstrate:

• Failure-specific repair across ACWM-Phys environments and full backbone codebases
• Held-out certified transfer with 83% sign accuracy and 0.06 negative transfer
• Counterexample-driven probe evolution that improves the diagnostic representation conditionally

The authors acknowledge limitations: guarantees are conditional on complete goal contracts and valid evaluators; fingerprints are local and must be recomputed after major architectural changes; and early archives inherit limits of observational priors and single-laboratory coverage.

Future directions include broader model families, uncertainty calibration, and contradiction-aware scheduling. The work suggests a paradigm shift in which world model optimization evolves from isolated campaigns into a continually improving knowledge system, where transfer is licensed by evidence rather than assumed by proximity.