acceptodds
Withdrawn from ICLR 2027

Sampling Headroom Is Not Selection Gain: A Compute-Value Audit of Test-Time Scaling for Video World Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time scaling (TTS) succeeds when additional computation improves the output a system returns. For video world models, however, a larger candidate pool can contain better videos while leaving selected quality nearly unchanged. We introduce the Compute-Value Audit (CVA), a sequential framework that traces this gap through four requirements: available improvement, informative inference-time signals, useful interventions, and gains that exceed matched uniform compute after generation and verification costs. On 192 Physics-IQ scenes, expanding the Wan2.2-5B pool from 4 to 16 candidates raises oracle quality by +9.23 IQ points (95% confidence interval [+7.44,+11.14]), while the estimated changes for denoising-stability, reconstruction-consistency, and learned-preference selectors range from −0.75 to +0.26 IQ. On VideoPhy2, perturb-and-refine sampling yields negative estimated changes in joint semantic and physical success across three fresh-seed replications at a matched budget of 60 denoiser evaluations; the intervals include zero. Reasoning controls expose the decisive distinction between predicting difficulty and identifying where additional sampling helps. An anchor–explorer policy beats matched uniform allocation on PRM800K, with gains concentrated in a small subset of problems, while privileged real-future references improve video selection. CVA thus shifts the scaling question from how much quality a pool contains to how much a realizable decision recovers at full cost.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Test-Time Scaling, Video World Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KBV4OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies test-time scaling (TTS) for video world models from a compute-value perspective. It introduces the Compute-Value Audit (CVA) framework, which decomposes the effectiveness of TTS into four requirements: opportunity, state confidence, action value, and compute cost. Through extensive experiments across video generation and reasoning benchmarks, the paper shows that increasing the number of generated candidates can substantially improve oracle quality without translating into meaningful improvements in the quality of the selected output.

Critical strengths

I appreciate the paper's perspective on test-time scaling. The authors distinguish between the potential improvement available in a candidate pool and the improvement that can actually be recovered by a practical selection mechanism. I find this distinction refreshing and important, especially in the context of video world models, where evaluating physical correctness remains challenging.

Critical weaknesses
  1. The contribution remains largely diagnostic rather than constructive. While I find the problem formulation interesting, the paper primarily focuses on evaluating and diagnosing the limitations of existing TTS mechanisms. It remains unclear how the proposed framework can directly inform the design of better TTS algorithms for video world models. In particular, the paper does not provide a convincing new mechanism that effectively closes the gap between sampling headroom and realized selection gain in this domain.

  2. The main conclusions are somewhat expected, despite the extensive experimental analysis. The central observations—that better candidates do not necessarily lead to better selected outputs, that imperfect verifiers can limit the benefits of additional sampling, and that adaptive computation must outperform compute-matched baselines—are largely consistent with existing insights about verifier alignment and test-time optimization.

Questions for authors

Can the authors demonstrate a case in video world models where the CVA framework leads to a new or improved TTS strategy, rather than merely diagnosing the failure of existing approaches?

4 comments

We thank the reviewer for recognizing the importance of distinguishing sampling headroom from deployable selection gain. CVA denotes the Compute-Value Audit, and TTS denotes test-time scaling. We provide a constructive video comparison motivated by the audit's diagnoses.

W1 & Q1: A constructive video TTS strategy derived from CVA

We introduce and evaluate a CVA-guided, state-aware compute allocation strategy that converts the audit's diagnostic findings into an improved video TTS policy.

The design follows directly from the original CVA results.

On Wan2.2-5B, increasing the candidate pool from B=4B=4 to B=16B=16 raises the Physics-IQ quality score (IQ) of the oracle from 35.0935.09 to 44.3244.32, but Native cycle selection improves only from 27.5327.53 to 27.7927.79. Thus, increasing sampling width alone has limited value when the additional candidates cannot be identified reliably.

The original G+M selector, which combines global history evidence (G) and the motion signal (M), improves the B=4B=4 selected quality to 30.8130.81 IQ, but its motion component is not universally reliable. On ABot-14B Dynamic scenes, M achieves 39.8639.86 IQ against a no-TTS baseline of 40.8040.80, whereas G reaches 42.2842.28. This motivates verification conditioned on the physical requirements of the scene.

These observations motivate two coupled design changes.

1. State-aware verification.

The policy estimates motion requirements and verifier disagreement from observable prompts, conditioning frames, and generated candidate statistics. It chooses between the inexpensive G/M signals and additional learned-dynamics verification according to their calibrated conditional utility.

Rather than uniformly applying the same verification rule, the policy emphasizes motion-preserving evidence on dynamic interactions and avoids unnecessarily expensive verification when inexpensive signals already agree.

2. Value-aware compute allocation.

The policy objective balances expected quality improvement against incremental normalized inference cost.

For state ss and action aa, we define

U(a∣s)=E[ΔQ(a)∣s]−λC(a∣s)U(a\mid s)=\mathbb{E}[\Delta Q(a)\mid s]-\lambda C(a\mid s),

where ΔQ\Delta Q is the improvement over the current selected output, CC is the incremental normalized inference cost, and λ\lambda controls the quality–compute trade-off. This equation states the allocation objective; the frontier evaluates the combined policy, without isolating the effects of its components.

The action set contains accepting the current candidate, generating another candidate, and invoking the learned-dynamics verifier. Conditional utilities are calibrated using family-disjoint data, and all selection decisions are made without access to the current query's target IQ or privileged future frames.

The policy uses at most eight generated candidates per scene, with an average of 5.6 candidates and 4.0 learned-verifier calls. Its mean normalized cost is 144 Diffusion Transformer (DiT)-equivalent units, including 112 generation units and 32 verification units. The frontier uses fixed relative operation prices for generation and learned-verifier inference; the resulting cost ratios are distinct from strict same-model forward counts and measured latency or energy.

Constructive evaluation: quality versus complete inference cost

The CVA-guided policy improves selected quality over the best reported fixed-width uniform baseline under a common mean-cost cap, and uses less normalized compute than the wider learned-verifier baseline.

MethodMean candidate countTotal normalized costSelected IQ
No-TTS12025.07
G+M, B=448030.81
WMReward, B=4411231.62
Uniform WMReward, B=5514031.79
CVA-guided selection5.614432.36
G+M, B=8816031.43
WMReward, B=8822432.18
G+M, B=161632031.78
WMReward, B=161644832.76

At a mean cost of 144 units, CVA-guided selection achieves 32.3632.36 IQ. The best reported fixed-width uniform WMReward baseline under this mean-cost cap uses five candidates and costs 140 units, reaching 31.7931.79 IQ.

The resulting improvement is +0.57+0.57 IQ, with a paired 95% confidence interval (CI) of [+0.08,+1.11][+0.08,+1.11] over Physics-IQ event families. This interval applies to the displayed fixed-width comparison. The two policies share a mean-cost cap, but their actual expenditure differs; the cap does not constrain each scene's realized expenditure.

The proposed policy also exceeds the IQ point estimate of uniform WMReward at B=8B=8 (32.1832.18 IQ) while using 35.7% less normalized compute (144144 versus 224224 units).

This establishes a positive quality–cost trade-off against the displayed fixed-width baselines. State-independent randomized width policies and a matched-random allocation control are not reported for the new video policy. The result therefore does not complete the full allocation audit required by CVA or isolate the benefit of assigning compute according to state.

The policy and WMReward baseline use the same prompts, conditioning inputs, learned verifier, and candidate-generation model. The evaluated policy changes how computation and selection signals are used across scenes. The aggregate comparison does not attribute the gain separately to motion conditioning, conditional-utility estimation, or score combination. The revised appendix distinguishes this policy objective from a complete executable specification and does not supply unreported parameter values.

The original twelve adaptive-depth controls relied on the previously evaluated score-only state signals and did not pass G4. The new policy combines motion-state information, learned-dynamics verification, and explicit marginal-cost calibration. Its fixed-width comparison tests a richer state–action class while retaining the additional controls required for a full allocation-audit pass.

W2: Contribution beyond previously known verifier-alignment observations

The novelty of CVA is the operational separation of four conditions that are frequently conflated when evaluating inference-time scaling.

Prior TTS research establishes that search and verification are important. CVA determines whether the available opportunity, observable state, actionable intervention, and complete inference economics jointly support a deployable improvement.

The experiments establish that these properties are distinct.

G1: Opportunity. On Physics-IQ, increasing Wan2.2-5B width from B=4B=4 to B=16B=16 creates +9.23+9.23 oracle IQ points. Thus, failure to improve selected quality is not explained by the absence of better candidates.

G2: State confidence. On MMLU-Pro, disagreement identifies harder prompts, producing a +15.20+15.20 percentage-point state separation, but this predictive signal alone does not establish useful additional sampling.

G3: Action value. Anchor–explorer on PRM800K demonstrates a positive action-value case, improving by +3.93+3.93 percentage points over matched random allocation and +2.64+2.64 over matched uniform allocation. Conversely, the original video refinement experiments fail to produce reliable gains.

G4: Complete entry fee. None of the twelve original video adaptive-depth policies exceeds matched uniform computation; they recover only 42–69% of the measured fee.

The constructive experiment demonstrates how these diagnoses translate into algorithm design. Since G1 passes but existing video verification and allocation are insufficient, the new policy improves the information used for selecting actions and explicitly prices every additional inference operation.

What the comparisons show

The results compare verifier choice, sampling width, and the combined allocation policy.

First, the learned-dynamics verifier improves fixed-pool selection: WMReward reaches 31.6231.62 IQ at B=4B=4, compared with 30.8130.81 for G+M. This establishes the value of stronger physical verification.

Second, the multi-budget results demonstrate diminishing selected-quality returns from increasing uniform sampling width. G+M improves from 30.8130.81 at B=4B=4 to only 31.7831.78 at B=16B=16, despite the substantial increase in oracle quality.

Third, the CVA-guided policy combines these observations to allocate additional computation selectively, reaching 32.3632.36 IQ at a mean cost substantially below uniform B=8B=8 learned-verifier inference.

Thank you once again for taking the time to review our manuscript. We have carefully addressed all the concerns raised in your review and provided detailed responses to fully resolve each of the points raised.

As the discussion period is coming to a close, we would like to kindly follow up on our previous response in case you have not yet had a chance to review it. We appreciate the many demands on your time and would be grateful for any opportunity you may have to review our responses.

Thank you once again for your time and consideration!

Thank you for the clarifications and additional experiments. The new video results partially address my concern about CVA’s practical value. However, as acknowledged in Appendix P, the new policy still lacks important implementation details, and the experiments do not clearly show how much state-aware allocation contributes to the improvement. These issues leave my main concern only partially addressed, so I maintain my rating.

Reviewer_unP2OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether test-time scaling (best-of-B selection and iterative refinement) improves video world models once generation and verification compute are fairly counted. The authors propose an offline four-gate evaluation checklist (G1-G4, the Compute-Value Audit) and evaluate three video generators (Wan2.2-5B, ABot-14B, Cosmos-Nano) across five video benchmarks alongside two LLM reasoning tasks. They show that while larger candidate pools (B=4 to B=16 on Physics-IQ) substantially increase oracle quality, existing verifiers and Predict-and-Perturb (P&P) refinement fail to convert this headroom into actual gains, and they propose a lightweight history-plus-low-motion selector (CVA-Select).

Critical strengths

The compute-calibrated audit of video test-time scaling provides a much-needed reality check against uncalibrated oracle best-of-B claims. By locking candidate pools and accounting for all generator and verifier DiT forward passes in Section 4 (Tables 1 and 5), the paper convincingly shows that existing self-scoring signals (Proprio, cycle reconstruction, VideoReward) and in-sampler P&P refinement recover almost none of the physical reasoning headroom.

The experimental controls cleanly isolate where candidate selection signals actually come from. Using shuffled-history baselines and ordinary pairwise reward-model controls in Section 4.1 (Table 1), the authors separate genuine physical trajectory alignment from generic visual quality or prompt priors.

Separating state uncertainty (G2) from action value (G3) offers a clear diagnostic framework for adaptive compute allocation. The contrast between MMLU-Pro and PRM800K in Section 4.2 (Table 3; Appendix A, Tables 8 and 9) nicely illustrates that answer disagreement identifies hard prompts without guaranteeing that extra sampling will fix them.

Critical weaknesses

The proposed constructive selector (CVA-Select, G+M) is only evaluated at B=4 and is missing from the B=16 scaling experiments in Section 4.4 (Table 5; Appendix E.2, Table 7). Because other history-based selectors suffer negative recovery as the pool grows from B=4 to B=16 (e.g., Retrieved history drops by -0.61 IQ), omitting CVA-Select at B=16 leaves its scalability unverified.

On Physics-IQ at B=4 (Section 4.1, Table 2), a degenerate baseline that simply picks the video with the lowest pixel motion (M) already achieves 36.38 IQ and beats every existing verifier. Combining global history matching with low motion (G+M, 36.60 IQ) barely improves over G alone (36.55 IQ) in a post-hoc comparison, and it remains unclear whether CVA-Select degrades on dynamic scenes like the low-motion heuristic does in Appendix F (Table 22).

The label "CVA-Select" does not refer to a single consistent method across benchmarks. It denotes conditioning-history template matching plus low motion on Physics-IQ (Section 4.1, Table 2), whereas on PAI-Bench-robot (Section 4.3, Table 4) the selectors switch to model-specific retrospective visual rerankers without conditioning history.

The presentation is difficult to follow and relies heavily on the appendix for core definitions. Figure 1 depicts G1-G4 as sequential steps underneath the generator (which looks like an online per-prompt routing pipeline rather than an offline dataset audit), and key formulas and notation referenced in the main text (such as Eq. 11 in Table 5 and c00 in Section 3.3) are only defined in the appendix.

Questions for authors

Can you evaluate Global history (G) and CVA-Select (G+M) on the B=16 Wan2.2-5B pool in Section 4.4 (Table 5 and Appendix E.2, Table 7)? Since the B=16 candidate videos and low-motion scores are already present, computing G and G+M would directly test whether CVA-Select holds up at larger budgets.

In Appendix F (Table 22), the low-motion heuristic works well on Calm scenes but drops below no-TTS on Dynamic scenes for ABot-14B (39.86 vs. 40.80). How do G and CVA-Select (G+M) perform on that same Calm vs. Dynamic split?

For the negative P&P and Proprio baselines on Wan2.2-5B (Section 4.1, Table 1; Appendix C, Table 17), did you retune the P&P perturbation schedule for Wan2.2-TI2V-5B, and why was Proprio only evaluated via static 17-frame scoring (Appendix E.7) without its refinement stage?

4 comments

We thank the reviewer for the detailed assessment of the compute-calibrated evaluation and for identifying the missing budget-scaling, motion-stratified, and refinement controls. CVA denotes the Compute-Value Audit, and TTS denotes test-time scaling; IQ is the Physics-IQ quality score. We address each concern below.

W1 & Q1: CVA-Select scalability at B=16

We extend the original CVA-Select evaluation to B=8B=8 and B=16B=16 while preserving its frozen scoring rule and candidate identities.

The original Physics-IQ experiment reports global history evidence (G), the motion signal (M), and their sum G+M at B=4B=4. The expanded analysis reconstructs the original global-history scores using the same leave-one-family-out (LOFO) reference pool and evaluates all three selectors on nested Wan2.2-5B candidate pools. Each selector uses the same score orientation and within-pool normalization across budgets, without target IQ feedback.

SelectorB=4 IQB=8 IQB=16 IQ
G30.5930.9731.31
M30.3330.5530.77
G+M30.8131.4331.78
Native cycle27.5327.6827.79
Oracle35.0939.8644.32

G+M improves by +0.97+0.97 IQ from B=4B=4 to B=16B=16, compared with +0.72+0.72 for G and +0.44+0.44 for M. The paired 95% confidence interval (CI) for the G+M budget gain is [−0.81,+2.74][-0.81,+2.74]. Its advantage over G grows from +0.22+0.22 at B=4B=4 to +0.47+0.47 IQ at B=16B=16, with a paired interval of [−0.36,+1.35][-0.36,+1.35] at the largest budget.

Thus, G+M maintains its descriptive advantage over G on Wan2.2-5B, but the additional recovery is small and remains statistically uncertain. The much larger oracle gain (+9.23+9.23 IQ) confirms that widening the pool creates more improvement opportunity than the existing selector can reliably recover.

We also preserve the distinction between the original exploratory selection of G+M and the subsequent evaluation of that fixed rule. The B=4B=4 comparison is not treated as an independent confirmation of the combination.

W2 & Q2: Calm versus Dynamic scenes

The motion-stratified analysis establishes that G mitigates the failure of low-motion selection on Dynamic scenes, while the original G+M combination retains a residual motion penalty.

We use the same Calm/Dynamic partition as the original appendix, covering 33 Calm and 33 Dynamic event families, with 99 scenes in each group. This 198-scene diagnostic population is kept separate from the 192-scene main Physics-IQ evaluation.

ABot-14B selectorCalm IQDynamic IQ
No-TTS40.2440.80
M46.9339.86
G45.9842.28
G+M46.9541.07

The low-motion heuristic performs strongly on Calm scenes, reaching 46.9346.93 IQ, but decreases Dynamic-scene quality from the no-TTS baseline of 40.8040.80 to 39.8639.86.

G avoids this failure, reaching 42.2842.28 IQ on Dynamic scenes, or +1.48+1.48 over no-TTS. G+M achieves 41.0741.07, improving over M by +1.21+1.21 IQ and over no-TTS by +0.27+0.27, but remaining 1.211.21 IQ below G alone.

This difference is important. The global G+M combination is not uniformly optimal across scene types. Motion reduction helps in relatively static scenes but can suppress physically necessary movement in dynamic interactions.

The analysis motivates verification conditioned on the physical requirements of the scene. The stronger Dynamic result of G identifies a reason to test state-aware selection; this stratified comparison alone does not establish the benefit of a runtime allocation policy.

W3: Consistency of CVA-Select across benchmarks

CVA-Select uses a shared G+M decision interface, while the underlying observables remain explicitly benchmark-dependent.

For Physics-IQ, G represents train-global LOFO history evidence and M penalizes excessive frame-difference motion. For PAI-Bench-robot, G combines the three non-motion quality terms, while M measures positive motion adequacy.

These different signal definitions reflect the observables available in each benchmark. They do not constitute identical feature implementations or independently validate the same physical heuristic.

We therefore distinguish the common scoring interface from its benchmark-specific instantiations and present all component definitions explicitly.

The submitted PAI-Bench-robot results remain unchanged:

ModelNo-TTSCVA-Select
Wan2.2-5B0.75940.7698
ABot-14B0.76890.7725
Cosmos3-Nano0.77300.7758

We additionally retain the negative OpenS2V result: CVA-Select achieves 0.39920.3992, below the best fixed candidate source at 0.41270.4127. This establishes that the shared interface does not guarantee cross-benchmark improvement and that each instantiation must pass its own CVA evaluation.

W4: Presentation and notation

We clarify that G1–G4 constitute an offline evaluation protocol rather than four runtime operations inside the video generator.

The revised presentation separates two layers: the online generation and selection procedure, and the offline CVA analysis of the resulting decisions.

The main text also explicitly defines the variables used in the principal compute comparisons, including the terms previously introduced only in the appendix. In particular, c00c_{00} denotes the initial candidate used for the no-TTS comparison, whereas the complete candidate set is used for best-of-B selection.

The revised Figure 1 distinguishes the candidate-generation pipeline from the audit gates and makes clear that CVA itself neither generates videos nor assigns selection scores.

Q3: Wan-specific Predict-and-Perturb (P&P) tuning and Proprio refinement

We provide a Wan2.2-specific perturbation-schedule comparison under the same constraint of 60 denoiser function evaluations (NFE) and distinguish static Proprio scoring from its refinement implementation.

The original P&P experiment uses 40 base denoising steps plus 20 refinement evaluations. We compare alternative allocations while retaining the same 179 VideoPhy2-hard prompts, frozen evaluator, initial latents, and total denoiser budget.

ConfigurationBase NFERefinement NFEJoint success
Uniform FlowMatch6005.03%
P&P, stronger perturbation30303.91%
P&P, original40203.35%
P&P, moderate refinement45154.47%
P&P, short refinement50105.03%
Proprio-guided refinement60 totalIncluded4.47%

The best P&P schedule reaches 5.03%5.03\% Joint success, matching uniform FlowMatch but providing no measurable improvement. The original configuration obtains 3.35%3.35\%, consistent with the submitted experiment's primary result.

The Proprio-guided refinement control reaches 4.47%4.47\% under the same 60-NFE budget. Its lower Joint-success point estimate shows no observed improvement in the tested Wan2.2-5B configuration under matched computation; it does not establish statistically reliable harm.

The original 17-frame Proprio experiment has a narrower interpretation: it measures the quality of its static video-ranking signal. The refinement experiment separately tests whether that signal supplies a beneficial trajectory intervention. The two evaluations are therefore reported as distinct evidence rather than treating static scoring as a complete implementation of Proprio refinement.

Across all tested configurations, the additional refinement steps do not produce a net benefit over the matched uniform sampler. This conclusion remains specific to the Wan-compatible implementations and tested compute budgets.

Summary

The additional analyses directly address the reviewer's concerns. G+M exhibits limited descriptive scaling to B=16B=16, while the growing oracle gap demonstrates persistent selection inefficiency. Motion-stratified results show lower quality under unconditional low-motion scoring on Dynamic scenes and motivate state-dependent verification. The multi-schedule P&P and Proprio controls show no observed improvement over uniform sampling under the tested denoising budgets.

Thank you once again for taking the time to review our manuscript. We have carefully addressed all the concerns raised in your review and provided detailed responses to fully resolve each of the points raised.

As the discussion period is coming to a close, we would like to kindly follow up on our previous response in case you have not yet had a chance to review it. We appreciate the many demands on your time and would be grateful for any opportunity you may have to review our responses.

Thank you once again for your time and consideration!

Thanks for running the extra experiments. The P&P schedule sweep and the Proprio refinement run on VideoPhy2-hard answer my question on Q3.

That said, the new numbers confirm my main worries about CVA-Select. Going from B=4 to B=16 on Wan2.2-5B, G+M only moves from 30.81 to 31.78 (vs. +9.23 for the oracle), its small edge over G alone is not statistically significant, and off-the-shelf WMReward beats it at every budget. On the ABot-14B Dynamic split, adding the low-motion term actually hurts performance compared to G alone (41.07 vs. 42.28), and CVA-Select is still a different hand-crafted rule on each benchmark. The new CVA-guided policy added in the revision is also missing key implementation details and a random-width baseline, so I will keep my score at 2.

Reviewer_8vweOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Compute-Value Audit (CVA) to study how additional test-time computation improves the output of video generation models (mainly regarding physics/video consistency and plausibility).

It examines candidate selection, compute allocation, and sampling refinement. The main finding is that larger candidate pools can contain better videos while the selected output shows little improvement.

Critical strengths

The main experiment provides useful evidence of the gap between sampling headroom and selection gain. Using shared candidate pools and paired uncertainty estimates, increasing the pool from 4 to 16 raises oracle quality by 9.23 IQ points, while reconstruction-based selection improves by only 0.26. The matched-compute refinement experiments and additional controls also strengthen the empirical analysis.

Critical weaknesses

My main area of expertise is test-time scaling for video diffusion models, and the comments below primarily assess the paper’s contributions from this perspective.

I find the main message quite familiar from the diffusion TTS literature. The importance of verifier alignment, search strategy, and compute allocation has already been studied [1, 2]. The paper provides useful measurements on these video tasks, but I am less convinced that CVA offers substantial new insight beyond organizing these existing ideas.

The support for CVA-Select is also incomplete. Its main results use B=4, and Appendix E.6 states that the records needed for B=16 are unavailable. In Table 2, CVA-Select improves the cross-model average over global history alone by only 0.05 IQ points (36.60 vs. 36.55). Appendix E.6 also states that this combination was highlighted after examining the results. The broader video scaling evidence mainly concerns fixed-pool reranking and one refinement configuration, with limited coverage of methods that actively explore denoising trajectories, such as beam search and cyclic diffusion search [2, 3].

For a paper centered on compute value, I expected a more complete cost–performance analysis. The paper reports selected budget comparisons and a compute ledger, but does not present a clear end-to-end cost–performance frontier. The appendix’s references to scripts, missing records, and separate historical variants also make the evidence difficult to follow.

[1] Scaling Inference Time Compute for Diffusion Models. CVPR, 2025.

[2] Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search. NeurIPS, 2025.

[3] Adaptive Inference-Time Scaling via Cyclic Diffusion Search. NeurIPS, 2025.

Questions for authors

What new prediction or practical decision does CVA provide beyond existing analyses of verifier alignment and compute allocation?

Can you evaluate the same frozen CVA-Select rule and its strongest components across multiple budgets, with paired uncertainty estimates?

Can you report cost–performance curves including generation and verification costs, and include a compatible trajectory-search baseline under matched budgets?

Since the main video experiments focus on physical plausibility, can the authors compare against a verifier or guidance method that uses a learned dynamics prior, such as WMReward [4], under matched information and compute budgets? This would help test how much the observed selection bottleneck depends on the chosen verifier family.

[4] Inference-time Physics Alignment of Video Generative Models with Latent World Models. arXiv (Jan 2026)

5 comments

We thank the reviewer for recognizing the rigorous measurement of the gap between sampling headroom and realized selection gain. We address the concerns regarding novelty, selector scalability, complete compute accounting, trajectory search, and learned-dynamics verification. CVA denotes the Compute-Value Audit, and TTS denotes test-time scaling; IQ is the Physics-IQ quality score.

W1 & Q1: Novelty and practical decisions beyond existing TTS analyses

CVA provides an actionable, compute-calibrated diagnosis of TTS that separates the existence of better candidates from the ability to identify and exploit them at an acceptable inference cost.

We agree that verifier alignment and compute allocation have been studied in previous diffusion TTS literature. Our contribution is not the isolated observation that imperfect verifiers limit scaling. CVA formalizes four independently testable requirements: G1 identifies available improvement, G2 determines whether observable signals represent informative states, G3 measures the value of interventions conditioned on those states, and G4 establishes whether the resulting gain exceeds matched uniform computation after accounting for all inference costs.

This separation produces concrete decisions.

First, wider sampling has limited value when the verifier is the bottleneck. On Physics-IQ, expanding Wan2.2-5B from B=4B=4 to B=16B=16 increases oracle quality from 35.0935.09 to 44.3244.32 IQ (+9.23+9.23, 95% confidence interval (CI) [+7.44,+11.14][+7.44,+11.14]), while Native cycle selection improves only from 27.5327.53 to 27.7927.79 (+0.26+0.26 IQ).

Second, state confidence does not establish action value. The MMLU-Pro control demonstrates that disagreement predicts difficulty without identifying prompts benefiting from additional sampling, whereas anchor–explorer achieves a +2.64+2.64 percentage-point gain over matched uniform allocation on PRM800K.

Third, an effective intervention must also pay its complete entry fee. The original twelve adaptive-depth video policies recover only 42–69% of the measured fee and do not outperform matched uniform computation.

These findings identify different algorithmic requirements: improve verification when G1 succeeds but selection recovery is weak; improve the state–action mapping when G2 succeeds but G3 fails; and reduce or reallocate inference computation when G4 fails.

W2 & Q2: Frozen CVA-Select across sampling budgets

The multi-budget evaluation establishes the behavior of the same G, M, and G+M rules on the nested Wan2.2-5B candidate pools.

We reconstruct the original train-global history evidence for every additional candidate while preserving the same leave-one-family-out (LOFO) exclusion, candidate ordering, score orientation, and within-pool normalization. G denotes global history evidence and M denotes the motion signal. They retain their original definitions, and G+M uses the same fixed equal-weight combination across all budgets. No target IQ labels enter scoring or candidate selection.

SelectorB=4 IQB=8 IQB=16 IQGain, B=4→16
G30.5930.9731.31+0.72
M30.3330.5530.77+0.44
G+M30.8131.4331.78+0.97
Native cycle27.5327.6827.79+0.26
Oracle35.0939.8644.32+9.23

The G+M gain from B=4B=4 to B=16B=16 is +0.97+0.97 IQ, with a paired 95% CI of [−0.81,+2.74][-0.81,+2.74]. The corresponding gain over G alone increases from +0.22+0.22 at B=4B=4 to +0.46+0.46 at B=8B=8 and +0.47+0.47 at B=16B=16. At B=16B=16, the paired G+M-minus-G interval is [−0.36,+1.35][-0.36,+1.35].

These results show that the fixed G+M rule does not collapse with increasing pool width, but also does not establish statistically reliable scaling benefits. Its +0.97+0.97 IQ improvement recovers only a limited part of the +9.23+9.23 IQ increase in oracle quality.

Importantly, G+M remains a retrospectively selected combination in the original B=4B=4 analysis. We do not reinterpret the original comparison as independent confirmation or claim statistically established synergy.

W3 & Q3: Complete cost–performance frontier and trajectory search

The complete inference-cost frontier confirms that verification quality and compute allocation jointly determine realized TTS value.

For Wan2.2-5B, each standard candidate generation requires 20 Diffusion Transformer (DiT) forward evaluations. G+M requires no additional world-model forward evaluation after its candidate statistics are available, whereas Native cycle requires 10 additional evaluations per candidate. The frontier assigns each WMReward call 8 DiT-equivalent units. These relative operation prices define the reported normalization; its cost ratios are distinct from strict same-model forward counts and measured latency or energy.

The following comparison includes fixed-pool reranking, trajectory search, learned-dynamics verification, and CVA-guided allocation.

MethodTotal cost (DiT-equivalent units)Selected IQ
No-TTS2025.07
G+M, B=48030.81
Trajectory search, 80-unit budget8029.98
Native cycle, B=412027.53
WMReward, B=411231.62
Uniform WMReward, B=514031.79
CVA-guided selection14432.36
G+M, B=816031.43
Trajectory search, 160-unit budget16031.17
WMReward, B=822432.18
G+M, B=1632031.78
Trajectory search, 320-unit budget32031.89
WMReward, B=1644832.76
Native cycle, B=1648027.79

The trajectory-search baseline performs intermediate latent branching under matched total computation, using only legal inference-time signals. Its selected IQ increases from 29.9829.98 to 31.8931.89 as the budget increases from 80 to 320 units. At the largest budget, it exceeds G+M by only 0.110.11 IQ, demonstrating limited additional recovery rather than a substantial improvement from trajectory exploration.

The stronger result comes from CVA-guided allocation. At a mean cost of 144 normalized compute units, it reaches 32.3632.36 IQ, compared with 31.7931.79 for the best reported fixed-width uniform WMReward baseline at 140 units. The paired improvement is +0.57+0.57 IQ, with a 95% CI of [+0.08,+1.11][+0.08,+1.11].

Both policies fall under the same mean-cost cap, but their actual expenditure differs. This comparison covers the displayed fixed-width baseline. It does not establish superiority over state-independent randomized width policies. The new video policy has no reported matched-random allocation control, so this result does not complete the full allocation audit required by CVA.

The comparison includes all candidate-generation steps, intermediate trajectory evaluations, and learned-verifier calls. Its efficiency comparisons apply under the stated normalization and include the additional cost of verification and search.

W4 & Q4: Comparison with a learned-dynamics verifier

The learned-dynamics comparison identifies a measurable verifier-family effect and establishes that the selection bottleneck is not intrinsic to all video verifiers.

We evaluate WMReward on the identical Wan2.2-5B candidate pools, with the same prompts, conditioning frames, and candidate identities. Neither WMReward nor the existing selectors receives target IQ, privileged future frames, or evaluation labels.

SelectorB=4 IQB=8 IQB=16 IQ
Native cycle27.5327.6827.79
G30.5930.9731.31
G+M30.8131.4331.78
WMReward31.6232.1832.76

WMReward improves over G+M by +0.81+0.81, +0.75+0.75, and +0.98+0.98 IQ at B=4,8,16B=4,8,16, respectively. The additional dynamics prior therefore improves selection quality across all evaluated budgets.

However, the result does not eliminate the selection bottleneck: even at B=16B=16, WMReward achieves 32.7632.76 IQ against an oracle of 44.3244.32. Moreover, its additional verification cost prevents raw IQ improvement from directly implying superior inference efficiency.

This distinction supports CVA's central contribution. Stronger dynamics-informed verification improves recoverable quality, while G4 determines whether the improvement remains valuable after complete computation is charged.

Summary

The expanded evidence establishes three conclusions. First, CVA distinguishes the opportunity, verification, action, and cost constraints on TTS. Second, G+M shows limited descriptive scaling with budget, while learned-dynamics verification produces stronger but still incomplete recovery. Third, the normalized cost frontier shows that a CVA-guided strategy improves selected quality over the best reported fixed-width uniform baseline under a common mean-cost cap.

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

Related papers

Discussion (0)

Sign in to comment.