acceptodds
Under review as a conference paper at ICLR 2027

VERDI: Verified Contradictions for Continual World-Model Optimization

Abstract

World models serve different domains and are tuned to different metrics, from visual fidelity to physical consistency and efficiency, so automating their optimization is valuable. Their optimization operators—reusable training, representation, sampling, or inference changes—look portable, yet models agreeing on architecture class, size, and baseline score can respond to the same operator in opposite directions, so what repairs one degrades its nearest neighbor. Existing methods supply the search but not the answer: they rank candidates by observational descriptors, assume these predict the same intervention response, and discard failed transfers as noise. We argue that an operator's effect sign is fixed by the model's response along the direction the operator modifies, not by any descriptor, so similarity is not a sufficient statistic and no fixed description can repair itself. The repair signal is a verified contradiction—nearby models whose opposing outcomes for the same operator are established by probing and by paired experiments under a verifier the search loop cannot edit. VERDI turns this into recursive self-optimization through shared-probe Optimization Fingerprints, IRG-guided retrieval, frozen-verifier settlement, and probe admission gated by held-out replay. Across nine backbones, 27 leave-one-backbone-out campaigns, and 216 transfer decisions, VERDI predicts effect signs with 83% accuracy, reaches the first verified positive in 3.6 trials and 96 GPU-h (68% and 69% below cold start), and lowers negative-transfer risk from 0.34 to 0.06 at 47% coverage. Ablations confirm that the geometry, not permissive acceptance, does the work, and a small-scale Franka Panda study gains 17.5 percentage points on two tasks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
world models, transfer learning, continual learning, optimization, research agents, robotics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ekCZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how optimization experience can be reused across heterogeneous world models. VERDI characterizes each model with paired inference-time probe responses, retrieves compatible operators using Interventional Repair Geometry (IRG), and confirms their effects through target-side experiments under a frozen verifier with protected-metric checks. Reliable opposite outcomes for nearby models trigger additional probes, admitted through held-out replay. The paper reports leave-one-backbone-out transfer, selector and verification ablations, and model-specific video/control improvements, including a small real-robot study. Its central contribution is the proposed co-evolution of reusable operators and the behavioral representation used to select them.

Critical strengths
  1. The separation between transfer prioritization and target-side verification is well motivated. A favorable source prediction does not establish a target improvement, and utility gains cannot compensate for protected regressions. The paired confirmation, explicit support conditions, and retention of null/harmful outcomes make the proposed optimization loop more informative than simply collecting successful recipes. The mixed Ctrl-World and Cosmos-Predict 2 outcomes in Appendix C are useful examples of this distinction.

  2. Using cheap intervention responses to guide expensive operator trials is a promising design. The reported raw-response, metadata, fixed-executor, and verification controls provide useful initial evidence, while the simulated control results connect selected operators to downstream utility. I also appreciate that the small Franka experiment is presented with uncertainty and that the combined RL + self-forcing result is not attributed to either component alone.

Critical weaknesses
  1. The scope of the evidence supporting the main transfer claims is inconsistent. Table 3 and Section 4.5 describe 216 decisions over 27 campaigns, whereas Appendix B.12 explicitly attributes the same Table 3 point estimates and their confidence intervals to 72 decisions over nine campaigns on three backbones. Likewise, Table 5 is labeled as a nine-backbone study, but Table 8 presents the same ablation values on the three-backbone subset. Section 4.1 and Figure 5 also describe nine-backbone evaluation, while Appendix B generally assigns archive-prefix replay to the detailed subset. The paper needs to distinguish these results and their actual bootstrap units. Tables 6-7 list all nine targets, but the numerical per-target breakdown in Table 9 covers only three. This matters because transfer between heterogeneous interfaces is a central claim; the current reporting does not let me assess how much of the headline evidence comes from the other six targets.

  2. The experiments do not yet isolate the distinctive benefit of contradiction-driven representation refinement. Figure 7 compares four fixed probes, fourteen freely added probes, and seven gated probes; it therefore changes probe count and admission policy together. It supports careful admission, but does not establish that verified contradictions select better probes than another selection strategy using the same candidate pool, budget, and held-out gate. The three extra directions in Figure 8 and the locality example in Appendix C.8 do not resolve this comparison.

Questions for authors
  1. Which results in Tables 3/5 and Figures 5/6 use all nine backbones, and which use the three-backbone subset? Please reconcile Appendix B/B.12 with the main text and provide the missing per-target numerical results, together with the actual cluster counts used for each confidence interval.

  2. Does contradiction-driven probe selection improve held-out operator-selection regret over a control that uses the same candidate probes, admits the same number, and applies the same replay/calibration gate without selecting from verified opposing outcomes? Please specify the discovery, promotion, calibration, and final evaluation split sizes used for Figure 7 and report uncertainty for the comparison.

  3. How is the learned-embedding baseline in Table 3 trained and calibrated? Can you compare IRG with a response-informed transfer surrogate, such as RGPE or a Bayesian task embedding adapted to the compatible compiled-operator pool, while holding the archive, executor, verifier, and probing budget fixed? A risk comparison at matched transfer coverage would also help identify the contribution of the selection representation.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ekCZ (1/3): Scope, uncertainty, and matched-budget probe selection

  • Question 1: Which results use nine backbones versus the three-backbone subset? Please provide per-target statistics and the actual clusters used for confidence intervals.

Answer: Thank you for identifying the conflicting scope descriptions. The revised main transfer results cover all nine backbones; Appendix B.12 reports the original three-backbone subset.

(1) The conflicting descriptions have been reconciled. Tables 3/5, Figures 5/6, and Appendix Tables 8/9 use nine backbones, 27 campaign-seed units, and 216 decisions. Table 9 now reports all nine targets, including per-target correct signs, search costs, accepted outcomes, coverage, and risk. Main-study totals are 179/216 correct signs and 102 accepted transfers, comprising 68 positive, 28 null, and 6 harmful outcomes.

(2) The confidence intervals now specify their actual clusters. Figure 5 and Appendix E.1 use nine backbone clusters, with Table 25 reporting full-study intervals. Appendix B.12 retains the original three-backbone, nine-campaign, 72-decision results and their intervals. Trials/GPU-h use the hierarchical average over tasks, repeated searches, successful campaign-seed units, and finally equally weighted backbones; completion counts retain all 27 units. These rules are stated in the introduction to Appendix B.

(3) Additional backbone visualizations are directly accessible in the revised manuscript. Appendix E.5 contains DreamerV3, V-JEPA 2, DINO-WM, and TD-MPC2 in Figures 16–19, indexed in Table 33 with results in Tables 34–35. Their evaluation units are specified separately, and none is added to the 216 main-study decisions. Real-robot results appear in Section 4.2 and Figure 4; additional paired FRANKA videos are available in the project page's Real robot section.

  • Question 2: Does contradiction-driven probe selection improve regret with the same candidate pool, probe budget, and admission gate? What are the split sizes and uncertainty?

Answer: Thank you for requesting a controlled test of contradiction-driven selection. At the same seven-probe budget, candidate pool, and replay/calibration gate, contradiction-driven selection lowers held-out regret relative to both random selection and replay-gain selection.

(1) The new comparison holds probe budget and admission conditions fixed. The original Fixed/Free/Gated settings change both probe count and admission policy. In Appendix E.2, all three selection rules use seven probes, the same candidate pool, admission thresholds, and final evaluation set. Table 26 reports regret/calibration error of 0.30/0.15 for random selection, 0.26/0.11 for replay-gain, and 0.21/0.08 for contradiction-driven selection.

Selection ruleProbe budgetRegretCalibration error
Random candidate selection70.300.15
Replay-gain70.260.11
Contradiction-driven selection70.210.08

(2) The regret contrasts have positive cluster-level intervals. Table 27 reports comparator-minus-contradiction differences. Relative to random selection, the regret reduction is 0.09, with 95% CI [0.078, 0.102]; relative to replay-gain, it is 0.05 [0.021, 0.070]. Calibration error improves by 0.07 [0.058, 0.087] versus random selection. Versus replay-gain, the calibration difference is 0.03 [−0.002, 0.061]; this interval includes zero. The intervals are unadjusted and resample nine backbone clusters.

Contrast (comparator minus contradiction-driven)MetricPoint estimate95% cluster-bootstrap CIPositive differences / 9 clusters
Random candidate selectionRegret0.09[0.078, 0.102]9/9
Random candidate selectionCalibration error0.07[0.058, 0.087]9/9
Replay-gainRegret0.05[0.021, 0.070]7/9
Replay-gainCalibration error0.03[−0.002, 0.061]6/9

(3) The revised protocol states the split sizes. Nine-fold outer LOBO allocates four source backbones to discovery, two to promotion, two to calibration, and one held-out target to final evaluation in each fold. With three seeds per backbone, the stages contain 12, 6, 6, and 3 campaign-seed units, respectively. Final evaluation spans nine backbone clusters. These details are provided in Appendix E.2 and referenced in the Figure 7 caption and main text.

Figure, table, and appendix references are to the revised manuscript; substantive revisions are marked in dark blue. Rollout videos are indexed on the anonymous project page.

AuthorsOpenReview ↗

Response to Reviewer ekCZ (2/3): Matched-coverage comparison and the RGPE baseline

  • Question 3: How is the learned-embedding baseline trained and calibrated, and how does IRG compare with response-informed RGPE at matched coverage?

Answer: Thank you for requesting the baseline details and a matched-coverage comparison. At matched coverage of 102/216 decisions, IRG has 5.9% negative-transfer risk, compared with 11.8% for response-informed RGPE and 14.7% for the learned embedding.

(1) The comparison holds the transfer and verification conditions fixed. IRG and RGPE share the initial historical archive, compatible compiled-operator pool, operator registry, probes and query budget, executor, frozen verifier, and stopping rule. Additional target observations consume the same search budget. Table 28 reports harmful counts of 6, 12, and 15 among 102 accepted transfers, sign accuracies of 179/216, 171/216, and 160/216, and trial means of 3.83, 4.25, and 4.60 for IRG, RGPE, and the learned embedding, respectively.

SelectorSign accuracyTrialsCoverageHarmful / acceptedProtected violations
IRG179/216 = 0.8293.83102/216 = 47.2%6/102 = 5.9%0.03
Response-informed RGPE171/216 = 0.7924.25102/216 = 47.2%12/102 = 11.8%0.08
Learned embedding160/216 = 0.7414.60102/216 = 47.2%15/102 = 14.7%0.10

(2) The response-informed RGPE configuration is specified in Table 30. Compatible historical source observations train up to eight source GPs per fold; a target GP uses only observations completed before the current prediction. Inputs include budgeted response descriptors, aligned operator encodings, and missing-value masks; the target is continuous utility change. Each GP uses a constant mean and ARD Matérn-5/2 kernel. Ensemble weights use minimum-ranking-loss probabilities estimated from 256 posterior samples, with leave-one-out target predictions; fewer than two target observations trigger equal source weights. Candidates are ranked by ensemble expected improvement within the same finite compatible pool. GP fitting costs are recorded separately, and the original frozen verifier determines final outcomes.

RGPE componentExperimental setting
Source models and candidate spaceOne GP for each remaining source backbone with compatible historical observations, up to 8 per fold; evaluate only operator configurations within the same task-goal and response schema that pass the shared compiler's compatibility checks
Inputs and regression targetBudgeted response descriptors, semantically aligned operator types and parameter encodings, and missing-value masks; predict continuous utility changes under the original goal contract with consistent utility direction; scaling uses available training data only
GP and noiseSingle-output exact GP, constant mean, ARD Matérn-5/2 kernel with an output-scale parameter, and float64; use measured repeat variance when available, otherwise learn Gaussian-likelihood noise
FittingMaximize the exact marginal log likelihood with L-BFGS-B, up to 200 iterations, and record convergence status
WeightsEstimate the probability of minimum ranking loss using 256 posterior samples; use leave-one-out predictions for the target model and split tied contributions equally
Cold startWith fewer than 2 valid target observations, combine compatible source models with equal weights without estimating weights from target ranking loss
RankingRank by ensemble expected improvement relative to a predefined positive-utility threshold, selecting only within the fixed finite candidate pool
Verification and calibration boundaryThe original verifier determines positive/null/harmful outcomes and protected metrics; posterior probability of exceeding the utility threshold does not itself establish confirmation, and any additional probability calibration is restricted to independent source calibration data
Resources and sweepObservation queries count against the common budget and GP fitting costs are recorded separately; use scores saved before outcome revelation for an offline risk–coverage sweep at accepted counts of 43, 72, 102, 144, 180, and 216

Figure, table, and appendix references are to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer ekCZ (3/3): Learned embedding and risk-coverage results

  • Question 3 (continued): How is the learned-embedding baseline trained and calibrated, and how does IRG compare with response-informed RGPE at matched coverage?

Answer: Thank you for requesting the baseline details and a matched-coverage comparison. IRG has fewer harmful transfers at every partial-coverage operating point; all three selectors accept the same outcomes at full coverage.

(1) The learned embedding predicts utility-change direction from historical transfer metadata. Table 31 lists binned historical GPU-h cost, source family, campaign time intervals, cumulative transfer success/failure counts, and discrete backbone metadata as features. Training uses utility-change direction labels and inverse-frequency-weighted BCEWithLogitsLoss. Temperature scaling uses an 8:2 training/validation split within historical training data. All 216 held-out decisions are excluded from fitting and calibration. The logit sign predicts effect direction; the verifier's positive/null/harmful labels remain distinct from this prediction target.

ComponentProtocol
Input featuresBinned historical transfer cost in GPU-h, source-family encoding, campaign time intervals, cumulative transfer success/failure counts, and discrete backbone metadata; encoded features are concatenated into a fixed-length vector
Training targetPredict the effect direction of an individual transfer; labels are derived from the direction of observed utility changes under the target contract, and the logit sign gives the predicted direction
LossBCEWithLogitsLoss with inverse-frequency class weights computed from the training set
CalibrationTemperature scaling using an 8:2 training/validation split within the training data; only a global temperature parameter is fitted
Data separationHistorical time-window split; all 216 held-out decisions form a separate test set, and training and calibration exclude their samples, labels, and features

(2) IRG has fewer harmful transfers at every partial-coverage operating point. Table 29 reports accepted counts of 43, 72, 102, 144, 180, and 216. IRG harmful counts are 2, 5, 6, 18, 45, and 73; RGPE counts are 4, 7, 12, 29, 53, and 73; learned-embedding counts are 5, 11, 15, 39, 61, and 73. At full coverage, all methods accept the same 216 decisions and therefore share 73 harmful outcomes. Scores are saved before the corresponding outcomes are revealed; thresholds are not fitted to test outcomes. Accepted null outcomes remain in the risk denominator. The full protocol and comparisons appear in Appendix E.3.

Accepted decisionsIRG harmfulRGPE harmfulLearned embedding harmful
43245
725711
10261215
144182939
180455361
216737373

Figure, table, and appendix references are to the revised manuscript.

Reviewer_54AWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VERDI is a framework for reusing optimization experience across pretrained world models. It characterizes each model through small, reversible inference-time perturbations and records the resulting responses in an Optimization Fingerprint. These fingerprints guide retrieval of potentially useful training or inference modifications from an archive. Candidate modifications are tested in paired target experiments under a frozen evaluator that checks both improvement and protected metrics. When models with similar fingerprints respond oppositely to the same modification, the system proposes additional diagnostic probes and retains them only when held-out replay improves selection. The evaluation includes eight physical simulation environments, nine world-model backbones, and robot manipulation experiments

Critical strengths
  1. A useful approach to learning from failed transfer. Response-based fingerprints provide a concrete alternative to selecting optimization methods from architecture metadata or static embeddings. Using reliable disagreements between nearby models to refine the probes is particularly interesting.
  2. Explicit verification of automated experiments. Separating proposal, implementation, and evaluation, while retaining null and harmful outcomes, makes the optimization process easier to audit.
Critical weaknesses
  1. Inconsistent experimental scope and unlinked case-study records. For example, Table 5 describes results across nine backbones and 27 campaigns, but its shared rows reproduce Table 8, whose caption explicitly identifies a three-backbone subset. Similarly, Table 3 and Appendix B,12 look inconsistent.
  2. The breadth and total cost of transfer remain insufficiently documented. The reported 47% transfer coverage does not reveal how often useful transfer occurs between substantially different architectural familie
Questions for authors

Please add the details identified in the weaknesses and clarify apparent contradictions.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 54AW (1/3): Scope corrections, case mapping, and videos

  • Question 1: Which results cover all nine backbones, and how are the case studies linked to the reported results?

Answer: Thank you for pointing out the inconsistent scope and asking us to link the cases to the results. The revised results cover all nine backbones, and four additional visualizations extend the original Ctrl-World and Cosmos case studies to control, manipulation, and navigation tasks.

(1) The scope corrections now appear in the manuscript itself. Tables 3–5 and 8 use nine backbones and 27 campaign-seed units; Table 9 covers all nine targets and 216 decisions. Appendix B.12 explicitly retains its historical three-backbone subset. The existing case records are linked in Appendix E.5: Ctrl-World's latent-motion prior is documented in Appendix C.3, Tables 20–21 and Figure 10; Cosmos-Predict 2's DINO-guided REP in Appendix C.4, Table 22 and Figure 11; and Cosmos3's joint RL/self-forcing intervention in Appendices C.4–C.6, Tables 23–24 and Figures 3(b), 4 and 12.

BackboneOperator or configuration in the original paperExisting evidence location
Ctrl-WorldLaMo-style latent-motion priorAppendix C.3; Tables 20–21; Figure 10
Cosmos-Predict 2DINO-feature-guided REPAppendix C.4; Table 22; Figure 11
Cosmos3Joint RL + self-forcingAppendix C.4–C.6; Figures 3(b), 4, and 12; Tables 23–24
Backbone and taskFigure locationVERDI interventionCorresponding results and evaluation unit
DreamerV3 · Walker WalkFigure 16: (a) Scene 1; (b) left plotTeacher-guided policy distillation with DAgger data aggregationWalker Walk row in the task-results table; cumulative return of the displayed trajectory: 476.5 → 655.4
DreamerV3 · Reacher HardFigure 16: (a) Scene 2; (b) right plotReacherIK with PD-teacher supervisionReacher Hard row in the task-results table; mean fingertip–target distance of the displayed trajectory: 7.4 → 0.3 cm
V-JEPA 2 · Pick and PlaceFigure 17: (a) Task 1; bottom-left terminal-state comparisonAction-MoECorresponding row in the task-results table; 100 episodes per method, successful episodes: 40 → 75
V-JEPA 2 · Tool UseFigure 17: (b) Task 2; bottom-middle terminal-state comparisonAction-MoECorresponding row in the task-results table; 100 episodes per method, successful episodes: 35 → 42
V-JEPA 2 · Open and CloseFigure 17: (c) Task 3; bottom-right terminal-state comparisonAction-MoECorresponding row in the task-results table; 100 episodes per method, successful episodes: 32 → 56
V-JEPA 2 · Push and PullQuantitative results only; not shown in the figureAction-MoECorresponding row in the task-results table; 100 episodes per method, successful episodes: 21 → 34
DINO-WM · Wall NavigationFigure 18: (a) Cases 1 and 2; (b) corresponding distance curvesNavigation-policy optimizationSix paired conditions, evaluation identifiers 10–15; 6 trajectories per method, successful trajectories: 1 → 3
TD-MPC2 · Nut AssemblyFigure 19: (a) Layouts 1 and 2Trust-region-constrained MPPINut Assembly row in the task-results table; qualitative trajectory comparisons for two layouts

(2) The case index links each visualization to its intervention and evaluation results. Appendix E.5 groups all four figures, Table 33 maps tasks to interventions, figure panels, and evaluation units, and Tables 34–35 give the corresponding results. All four additions are standalone evaluations, excluded from the 216 LOBO decisions and the main-study overhead estimate. The four backbones belong to the latent dynamics/features family; the separate family-level analysis below quantifies cross-family transfer. The project page also provides seven corresponding comparison videos: Walker Walk and Reacher Hard for DreamerV3; Pick and Place, Tool Use, and Open and Close for V-JEPA 2; Wall Navigation for DINO-WM; and Nut Assembly for TD-MPC2.

BackboneCampaign-seed unitsHeld-out decisionsCorrect signsGated acceptedGated trials / GPU-hNegative-transfer risk
V-JEPA 232420/2410/244.54 / 1372/10 = 20.0%
DINO-WM32420/2413/243.7 / 1010/13 = 0.0%
DreamerV332422/2412/243.0 / 761/12 = 8.3%
TD-MPC232419/2412/244.5 / 1281/12 = 8.3%

Figure, table, and appendix references are to the revised manuscript; substantive revisions are marked in dark blue. Rollout videos are indexed on the anonymous project page.

AuthorsOpenReview ↗

Response to Reviewer 54AW (2/3): Additional control, manipulation, and navigation results

  • Question 1 (continued): Which results cover all nine backbones, and how are the case studies linked to the reported results?

Answer: Thank you for pointing out the inconsistent scope and asking us to link the cases to the results. The additional evaluations show improved DreamerV3 task trajectories, higher V-JEPA 2 success rates, better aggregate DINO-WM navigation outcomes, and paired TD-MPC2 manipulation behavior.

(1) DreamerV3 illustrates improvements in task behavior after the selected interventions. VERDI selects teacher-guided policy distillation with DAgger for Walker and ReacherIK with PD-teacher supervision for Reacher. In the displayed trajectories, Walker return increases from 476.5 to 655.4 (+37.5%), while Reacher mean fingertip–target distance decreases from 7.4 to 0.3 cm. Figure 16 shows the paired rollouts and metric curves, with trajectory-level results in Table 34.

(2) V-JEPA 2 provides an aggregate downstream success-rate improvement of 19.75 percentage points. The baseline is an existing checkpoint adapted to our RoboTwin data. VERDI selects Action-MoE and iteratively refines routing, expert count, shared experts, and router design without manual hyperparameter tuning by the authors, before freezing the final configuration. With 100 paired evaluation episodes per method in each of four task categories, success increases from 40 to 75 in Pick and Place, 35 to 42 in Tool Use, 21 to 34 in Push and Pull, and 32 to 56 in Open and Close. Overall success rises from 128/400 (32.0%) to 207/400 (51.75%). Table 34 reports all four tasks; Figure 17 shows representative rollouts for Pick and Place, Tool Use, and Open and Close.

Backbone, task, and metricBaselineVERDIChange or result
V-JEPA 2 · Pick and Place40/100 = 40%75/100 = 75%+35 pp
V-JEPA 2 · Tool Use35/100 = 35%42/100 = 42%+7 pp
V-JEPA 2 · Push and Pull21/100 = 21%34/100 = 34%+13 pp
V-JEPA 2 · Open and Close32/100 = 32%56/100 = 56%+24 pp
V-JEPA 2 · RoboTwin overall128/400 = 32.0%207/400 = 51.75%+19.75 pp, approximately 61.7% relative improvement
DINO-WM · Wall Navigation (Successful trajectories / evaluated trajectories)1/6 = 16.7%3/6 = 50.0%+33.3 pp; 2 additional successful trajectories
DINO-WM · Wall Navigation (Mean final distance to goal)13.289 grid units8.135 grid units38.8% reduction
DINO-WM · Wall Navigation (Mean distance to goal over the trajectory)19.247 grid units12.841 grid units33.3% reduction
DreamerV3 · Walker WalkCumulative return 476.5Cumulative return 655.4+178.9 (approximately +37.5%)
DreamerV3 · Reacher HardMean fingertip-target distance 7.4 cm0.3 cm95.9% reduction
TD-MPC2 · Nut AssemblyBefore rolloutVERDI rolloutQualitative trajectory comparison in Layouts 1 and 2

(3) DINO-WM improves the aggregate outcome across six paired navigation conditions. Successful trajectories increase from 1/6 to 3/6; mean final distance decreases from 13.289 to 8.135 grid units, and mean trajectory distance from 19.247 to 12.841. Each condition uses one trajectory per method with the same checkpoint, start–goal states, and planning budget. Success requires final distance below 4.5 grid units at control step 100. Figure 18 shows two improved cases, while Table 35 reports all six conditions, including the failures and two conditions with slightly larger final distances after intervention.

Evaluation condition (seed)Baseline final distanceVERDI final distanceBaseline successVERDI success
1014.70715.831NoNo
1124.7270.337NoYes
125.3061.640NoYes
1314.43915.037NoNo
1418.85814.990NoNo
151.6950.973YesYes
Mean final distance / success count13.2898.1351/63/6

(4) TD-MPC2 provides a further qualitative manipulation comparison. VERDI selects trust-region-constrained MPPI for Nut Assembly. Figure 19 compares baseline and VERDI trajectories in two layouts, with their alignment curves. The baseline remains misaligned over the displayed interval in Layout 1 and aligns later in Layout 2.

Figure, table, and appendix references are to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 54AW (3/3): Real-robot results, cross-family transfer, and compute

  • Question 1 (continued): Which results cover all nine backbones, and how are the case studies linked to the reported results?

Answer: Thank you for pointing out the inconsistent scope and asking us to link the cases to the results. Real-robot evaluation shows an average success-rate gain of 17.5 percentage points across two manipulation tasks.

(1) Real-robot evaluation shows an average success-rate gain of 17.5 percentage points across two manipulation tasks. Section 4.2 and Figure 4 report COSMOS3 with the combined RL + self-forcing operator on a Franka Panda. With 20 trials per method and task, pick-and-place success increases from 11/20 (55%) to 14/20 (70%), and drawer-opening success from 7/20 (35%) to 11/20 (55%). The project page's Real robot section also provides paired baseline and improved-policy videos for block stacking, tablecloth folding, and moving an occluded object from a bowl into a drawer before closing it. These qualitative demonstrations show a completed stack, a more direct folding path, and a completed pick–place–close sequence; the videos are presented at 4× speed.

Real-robot taskBaseline COSMOS3COSMOS3 with VERDI-selected operatorImprovement
Pick-and-place11/20 = 55%14/20 = 70%+15 pp
Drawer opening7/20 = 35%11/20 = 55%+20 pp
Mean across the two tasks45.0%62.5%+17.5 pp
  • Question 2: How much useful transfer occurs across architectural families, and what is the total computational cost?

Answer: Thank you for asking us to document transfer breadth and computational cost. The added family-level analysis identifies 40 verified positive cross-family transfers, while the compute analysis distinguishes measured component costs from the extrapolated study-scale overhead.

(1) VERDI yields 40 verified positive cross-family transfers with 4.8% conditional risk. Table 32 groups Ctrl-World, Cosmos-Predict 2, and Cosmos3 as visual prediction; IRASim and OccWorld as action–observation/occupancy; and DreamerV3, TD-MPC2, DINO-WM, and V-JEPA 2 as latent dynamics/features. Among 156 cross-family candidates, 62 are accepted: 40 positive, 19 null, and 3 harmful. Coverage is 39.7%, the positive-outcome fraction among accepted transfers is 64.5%, and conditional risk is 4.8%. Verified positives account for 40/156 = 25.6% of all cross-family candidates. Within-family transfer accepts 40/60 candidates, with 28 positive, 9 null, and 3 harmful outcomes. Appendix E.4 reports the complete source–target matrix.

Architecture familyBackbone
Visual predictionCtrl-World, Cosmos-Predict 2, Cosmos3
Action-observation / occupancyIRASim, OccWorld
Latent dynamics / featureDreamerV3, TD-MPC2, DINO-WM, V-JEPA 2
Source family → target familyVisual predictionAction-observation / occupancyLatent dynamics / featureTotal
Visual prediction18 / 9 / 6 / 2 / 118 / 8 / 6 / 2 / 036 / 11 / 6 / 4 / 172 / 28 / 18 / 8 / 2
Action-observation / occupancy18 / 8 / 5 / 3 / 06 / 4 / 3 / 1 / 024 / 9 / 4 / 4 / 148 / 21 / 12 / 8 / 1
Latent dynamics / feature36 / 15 / 12 / 2 / 124 / 11 / 7 / 4 / 036 / 27 / 19 / 6 / 296 / 53 / 38 / 12 / 3
Total72 / 32 / 23 / 7 / 248 / 23 / 16 / 7 / 096 / 47 / 29 / 14 / 4216 / 102 / 68 / 28 / 6

(2) Measured overhead totals 399 GPU-h for the three detailed cases. Appendix B.10 and Table 16 report fingerprinting/screening/confirmation costs of 9/51/36 GPU-h for Ctrl-World, 11/74/88 for Cosmos-Predict 2, and 8/60/62 for Cosmos3, giving totals of 96, 173, and 130 GPU-h. Screening is the largest component for Ctrl-World; confirmation is largest for the Cosmos cases.

BackboneFingerprintingScreeningConfirmationTotal
Ctrl-World9513696
Cosmos-Predict 2117488173
Cosmos386062130
Measured total399 GPU-h

(3) Estimated study-scale overhead is 3,759 GPU-h. This combines the measured 399 GPU-h with an assumed 140 GPU-h for each of the remaining 24 campaigns: 399 + 24 × 140 = 3,759 GPU-h. Varying the assumed per-campaign overhead from 115 to 165 GPU-h yields a sensitivity range of 3,159–4,359 GPU-h. The estimate excludes standalone evaluations. Appendix B.10 separately reports cost to the first verified positive, full search cost, and overhead.

Figure, table, and appendix references are to the revised manuscript.

AuthorsOpenReview ↗

A Gentle Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable comments and constructive feedback.

As the rebuttal discussion period is approaching its deadline, we would like to kindly follow up on our previous responses. We would sincerely appreciate it if you could take a moment to review our clarifications and let us know whether any concerns remain.

We fully understand your busy schedule and greatly appreciate your time and consideration.

Best regards,
The Authors

Reviewer_8iiNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VERDI accumulates evidence about optimization operators across pretrained world models. It characterizes models using paired inference-time probes, retrieves operator outcomes through an Interventional Repair Geometry, and prioritizes transfers using compatibility and uncertainty checks. An independent frozen verifier confirms effects through paired target experiments while checking protected outcomes. Opposite effects in nearby models motivate additional diagnostic probes, which must pass held-out replay before admission. The paper reports a nine-backbone leave-one-backbone-out study, selector and control-scaffold ablations, model-specific improvements, and small robot experiments. Its headline results include 83% effect-sign accuracy and a reduction in mean trials to a verified positive from 11.2 to 3.6.

Critical strengths
  1. The separation between predicting a useful transfer and verifying its actual target effect is well motivated. VERDI retains null and harmful outcomes, checks protected metrics, and freezes the evaluator and registry within a campaign. This is a useful design for scientific agents whose scalar improvements can conceal regressions. The appendices explicitly retain mixed outcomes for Ctrl-World and Cosmos-Predict 2 rather than treating every improved coordinate as a verified overall gain.
  2. The paper includes relevant comparisons of the selection representation and control scaffold. The metadata, raw-response, and learned-embedding selectors, together with fixed-executor controls, aim to test whether the response geometry adds value beyond a generic agent loop. The contradiction-driven probe-admission mechanism also has a concrete downstream criterion: improved selection regret and calibration on replay, rather than descriptive separation alone. These experiments could support a substantial empirical contribution if their scope and aggregation were unambiguous.
Critical weaknesses

I only have one point.

  1. The stated scope of the central experiments is inconsistent. Main Table 3 labels the selector results as 216 decisions over 27 campaign–seed units; Appendix B.12 explicitly associates the Table 3 point estimates and confidence intervals with three backbones, nine campaigns, and 72 decisions. Main Table 5 describes nine backbones and 27 campaigns, whereas Appendix Table 8 assigns the same fixed-executor values to the detailed three-backbone subset. Appendix B also describes subset-based replay, while Figure 5 describes nine-backbone replay. A backbone inventory does not resolve which observations generated each result. This directly affects the claimed breadth.
Questions for authors

Please give an explicit scope and denominator for each result in Figures 5–7 and Tables 3–5, resolve the conflicts with Appendix B and Tables 8–9, and provide per-backbone sufficient statistics for all nine targets. Which headline numbers change after this reconciliation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8iiN (1/3): Corrected scope and manuscript changes

Question: Experimental scope, denominators, per-backbone results, and changes to the headlines

Answer: Thank you for identifying the inconsistencies between the stated experimental scope and the reported results.We would like to explain it as follows.The revised main results consistently cover all nine backbones, and the corrected headline values are 82.9% sign accuracy, 3.83 trials / 105 GPU-h, 47.2% coverage, and 5.9% negative-transfer risk.

1. We have separated the full study from the historical three-backbone subset. The full LOBO study contains nine held-out targets, three campaign-seed units per target, and 24 transfer decisions per target: 27 units and 216 decisions in total. Figures 5–7, Tables 3–5, and Appendix Tables 8–9 cover all nine backbones; their result-specific denominators are reported below. Appendix B.12 reports the original three-backbone retrieval controls and intervals for nine campaigns and 72 decisions. The conflicting scope descriptions have been corrected in the introduction to Appendix B (Ablation and Audit Details).

Figure, table, or sectionIssue in the original manuscriptRevision
Abstract / main-text headline resultsUsed means from the three case-study targets in Appendix Table 9Recompute over all nine backbones
Figure 5: archive-prefix replayThe caption states nine backbones and 27 campaigns, whereas Audit Details assigns archive-prefix replay to the detailed three-backbone subsetReconcile the nine-backbone scope, metric denominators, and intervals based on 9 backbone clusters between the caption and appendix; retain the figure number and replay design
Figure 6(a,b)The caption specifies nine backbones and 27 campaigns, but the headline means came from three backbonesRecompute over nine backbones, with 27 campaign-level observations
Figure 6(c)The main text specifies 216 decisionsRetain the 216-decision denominator and report the numerators and denominators for coverage and risk
Figure 7: probe-growth ablationScope and evaluation denominator require clarificationSpecify nine LOBO targets and 27 campaign-seed evaluation units; report the matched-budget extension separately in Appendix E.2, Tables 26–27
Table 3The caption specifies 216 decisions / 27 units; B.12 assigns the point estimates and intervals to three backbones, 72 decisions / 9 campaignsReport completion counts, accuracy, trials, and protected violations over nine backbones; report risk separately at matched coverage
Table 4Completion counts such as 24/27 already cover nine backbones and 27 campaigns; VERDI's 3.6 trials used the three-backbone meanRetain completion counts and update VERDI's trials to the nine-backbone macro-average of 3.83
Table 5 / Appendix Table 8Table 5 claims nine backbones but duplicates the three-backbone results in Table 8; the original Table 8 caption agrees with its subset valuesReport both tables over nine backbones and update the scope stated in Table 8 accordingly
Appendix Table 9Lists only three case-study targetsExtend to all nine held-out targets
the introduction to Appendix B (Ablation and Audit Details)Mixes three-backbone point estimates with 27-campaign trials/cost averages without explaining the treatment of unsuccessful campaignsSpecify equal-weight averaging across nine backbones and the treatment of unsuccessful campaigns; report completion counts for each configuration

Figure, table, and appendix references are to the revised manuscript; substantive revisions are marked in dark blue. Rollout videos are indexed on the anonymous project page.

AuthorsOpenReview ↗

Response to Reviewer 8iiN (2/3): Denominators, uncertainty, and search cost

Gated reuse reduces trials by approximately 66.5% and GPU-h by approximately 67% under the corrected nine-backbone averaging scheme.

1. Each result now states its own denominator and uncertainty unit. Figure 5 evaluates every chronological archive prefix on 27 campaign-seed units and 216 decisions, with uncertainty resampled over nine backbone clusters. Figure 6(a,b) reports search means over successful campaigns, while Figure 6(c) uses 216 decisions for coverage and accepted, executed transfers for risk. Tables 3–5 and 8 report completion counts out of 27 alongside their search results. Figure 7 evaluates nine held-out backbones with 27 campaign-seed units; the added matched-budget contrasts use nine backbone clusters for their intervals (Appendix E.2, Tables 26–27).

ResultEvaluation scopeCampaign-seed unitsDecisions or other statistical denominatorResampling unit for intervals
Figure 5, each archive prefixNine LOBO held-out backbones27216 transfer decisions per prefix9 backbone clusters
Figure 6(a,b)Nine LOBO held-out backbones2727 campaign-level observations; means use successful campaignsPoint estimates in the figure
Figure 6(c)Nine LOBO held-out backbones27216 transfer decisions; risk is conditional on accepted, executed transfersPoint estimates in the figure
Figure 7 and the added matched-budget comparisonNine LOBO held-out backbones2727 campaign-seed evaluation unitsPoint estimates in the original figure; added comparison intervals use 9 backbone clusters
Table 3Nine LOBO held-out backbones27216 transfer decisions; trial counts use successful campaigns9 backbone clusters
Table 4Nine LOBO held-out backbones2727 campaign-level system comparisonsPoint estimates
Table 5Nine LOBO held-out backbones2727 campaign-level ablation evaluationsPoint estimates
Appendix Table 8Nine LOBO held-out backbones2727 fixed-executor campaign evaluationsPoint estimates
Appendix Table 9Nine LOBO held-out backbones27216 transfer decisionsAggregate intervals use 9 backbone clusters

2. The corrected search results preserve a substantial reduction in trials and compute. Cold start, unconditional reuse, and gated VERDI require 11.41/318, 4.21/123, and 3.83/105 trials/GPU-h, respectively. Using unrounded backbone macro-means, gated reuse reduces trials by approximately 66.5% and GPU-h by approximately 67% relative to cold start. Trials and compute are first averaged over tasks and repeated searches within each campaign-seed unit, then over successful units within each backbone, and finally equally across backbones. Unsuccessful campaigns remain in the completion denominator. Full VERDI completes 24/27 campaigns; the revised selector and fixed-executor comparisons appear in Tables 3–5 and 8.

StrategyTrials to first verified positiveGPU-hCoverageNegative-transfer risk
Cold start11.41 [10.54, 12.33]318 [295.4, 342.3]——
Unconditional reuse4.21 [3.83, 4.60]123 [110.2, 136.1]216/216 = 100%73/216 = 33.8%
Gated VERDI3.83 [3.49, 4.17]105 [92.4, 118.4]102/216 = 47.2% [44.9%, 50.0%]6/102 = 5.9% [2.0%, 10.1%]

3. The full-study uncertainty estimates are now distinct from the subset intervals. Appendix E.1 and Table 25 report descriptive 95% percentile-bootstrap intervals obtained by resampling nine complete backbone clusters. The intervals are [77.8%, 87.5%] for sign accuracy, [3.49, 4.17] for gated trials, [92.4, 118.4] GPU-h for gated compute, [44.9%, 50.0%] for coverage, and [2.0%, 10.1%] for risk. Sign accuracy measures agreement in utility-change direction; positive/null/harmful are the frozen verifier's final classifications, which also account for gain thresholds and protected metrics.

Figure, table, and appendix references are to the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 8iiN (3/3): Complete nine-backbone results and ablations

Table 9 now reports all nine held-out targets; Full VERDI also achieves the highest completion count and the fewest trials in the selector, system, and fixed-executor comparisons.

1. Table 9 provides complete results for all nine held-out targets. It reports each target's sign accuracy, trials/GPU-h under all three strategies, and positive/null/harmful outcomes. Totals are 179/216 correct signs and 102 accepted transfers, comprising 68 positive, 28 null, and 6 harmful outcomes. Thus, coverage is 102/216 = 47.2% and conditional risk is 6/102 = 5.9%, compared with 73/216 = 33.8% under unconditional reuse.

Held-out targetCorrect signs / 24Cold trials / GPU-hUnconditional trials / GPU-hGated trials / GPU-hAccepted (positive / null / harmful)CoverageRisk
Ctrl-World17 / 2410.1 / 2863.8 / 1043.3 / 8411 (8 / 2 / 1)11/24 = 45.8%1/11 = 9.1%
Cosmos-Predict 221 / 2412.4 / 3444.5 / 1323.9 / 10810 (7 / 3 / 0)10/24 = 41.7%0/10 = 0.0%
Cosmos322 / 2411.1 / 3064.0 / 1183.6 / 9611 (8 / 2 / 1)11/24 = 45.8%1/11 = 9.1%
IRASim21 / 2410.7 / 2983.6 / 1023.5 / 9112 (9 / 3 / 0)12/24 = 50.0%0/12 = 0.0%
OccWorld17 / 2413.2 / 3614.8 / 1424.4 / 12611 (7 / 4 / 0)11/24 = 45.8%0/11 = 0.0%
DreamerV322 / 249.6 / 2673.3 / 943.0 / 7612 (6 / 5 / 1)12/24 = 50.0%1/12 = 8.3%
TD-MPC219 / 2413.8 / 3855.2 / 1534.5 / 12812 (8 / 3 / 1)12/24 = 50.0%1/12 = 8.3%
DINO-WM20 / 2411.6 / 3204.0 / 1173.7 / 10113 (9 / 4 / 0)13/24 = 54.2%0/13 = 0.0%
V-JEPA 220 / 2410.17 / 2944.7 / 1464.54 / 13710 (6 / 2 / 2)10/24 = 41.7%2/10 = 20.0%
Total / macro mean179 / 216 = 82.9%11.41 / 3184.21 / 1233.83 / 105102 (68 / 28 / 6)102/216 = 47.2%6/102 = 5.9%

2. Full VERDI achieves the highest completion count and the fewest trials across the selector, system, and fixed-executor comparisons. The selector comparison uses 216 decisions; completion counts use 27 campaign-seed units. Shared fixed-executor configurations match across Tables 5 and 8; the frequency prior is the additional configuration in Table 8.

SelectorSuccessful campaigns / 27Sign accuracyTrialsProtected violations
Random150.500 (108/216)8.900.33
Metadata180.648 (140/216)5.200.13
Raw response200.708 (153/216)4.300.09
Learned embedding220.741 (160/216)4.600.10
Full fingerprint / IRG240.829 (179/216)3.830.03
SystemSuccessful campaigns / 27Trials
VERDI24/273.83
Codex20/274.60
Claude Code19/274.90
Human experts19/277.40
ConfigurationSuccessful campaigns / 27TrialsGPU-hRiskProtected violations
w/o eligibility checks214.401320.210.14
w/o IRG204.901470.180.11
Bare LLM agent186.001890.310.26
Generic research-agent loop166.602100.360.29
Frequency prior195.301650.290.22
Full VERDI243.831050.0590.03

The additional task evidence for DreamerV3, V-JEPA 2, DINO-WM, and TD-MPC2 is in Appendix E.5, Figures 16–19 and Tables 33–35. These standalone evaluations are separate from the 216 LOBO decisions. Section 4.2 reports the real-robot results, with additional paired FRANKA videos in the project page's Real robot section.

Figure, table, and appendix references are to the revised manuscript.

AuthorsOpenReview ↗

A Gentle Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable comments and constructive feedback.

As the rebuttal discussion period is approaching its deadline, we would like to kindly follow up on our previous responses. We would sincerely appreciate it if you could take a moment to review our clarifications and let us know whether any concerns remain.

We fully understand your busy schedule and greatly appreciate your time and consideration.

Best regards,
The Authors

Comments on the paper

1 comment

The author has responded to everyone's comments. Please reply when it is convenient for you.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.