acceptodds
Under review as a conference paper at ICLR 2027

Video Reasoning Generalizes to Perception, Embodiment, and Real-World Simulation

Abstract

Progress in video generation has been driven largely by visual fidelity. Currently, these models are being deployed as world models and embodied planners, where success depends on reasoning-like competences rather than on pixels. Whether such competences can be acquired by training, and whether they generalize beyond the domain they were trained on, remains unanswered. We introduce XVReason and XVReason-Bench, a video corpus that pairs training data with matched evaluation across four domains: abstract visual reasoning, low-level perception, embodied manipulation, and world simulation. On this foundation, we run a cross-domain study on two video backbones. Generalization turns out to be asymmetric. Abstract reasoning is the only fine-tuning domain that improves all four task domains on both backbones, with gains of +42.3/+20.6/+10.7% on the three application domains of Wan2.2. This generalization ability also works in different visual domains. Follow-up experiments show that generalization varies with the adaptation regime, the data scale, the denoising stage, and the task composition, and trace the gain to three competences: instruction following, goal completion, and scene stability. An expert-loading ablation recovers most of the reasoning gain through the high-noise expert. These findings point to a recipe, ReasonGraft, which blends a reasoning LoRA with an in-domain LoRA at a 3:7 ratio in weight space. It beats in-domain fine-tuning on all three application domains. Video Reasoning is an under-used general-purpose resource for training video generation models.

Primary area:
generative models
Keywords:
video reasoning, generalization, world models, video generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eYkTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether fine-tuning video generators on abstract reasoning tasks improves capabilities in other domains. It introduces XVReason and XVReason-Bench and conducts a complete four-domain transfer study on two backbones. Reasoning-data adaptation improves all four domain aggregates, including under visual shifts. Further ablations examine adaptation regime, data scale, denoising experts, and task composition. The proposed ReasonGraft blends reasoning and application-specific LoRA updates to improve downstream performance.

Critical strengths
  1. A valuable, systematic transfer study. The complete 4×44 \times 4 evaluation on two backbones, with matched sample budgets and shared inference settings, provides convincing evidence that synthetic task-solving data can benefit other generation tasks. The substantial perception and manipulation gains make this consequential beyond improvements on reasoning benchmarks alone.
  2. Informative diagnostic breadth. The adaptation, scale, expert-loading, task-composition, and visual-shift experiments help characterize when transfer occurs. In particular, the contrast between LoRA and full fine-tuning and the concentration of gains in the high-noise expert provide useful guidance for subsequent work.
Critical weaknesses
  1. The central claim is insufficiently isolated. Instruction-following-only training outperforms reasoning-only training on all three application domains (Figure 10). The results therefore leave unclear what reasoning contributes beyond improved instruction following.

  2. ReasonGraft’s added value is weakly supported. It uses standard adapter blending, while its gains over in-domain fine-tuning are only 0.2 points on perception and 0.3 on simulation, without uncertainty estimates (Table 3). These results do not establish a reliable general improvement.

Questions for authors
  1. How does ReasonGraft perform when the reasoning adapter is replaced by the instruction-following-only adapter? Can motion-controlled comparisons distinguish improved task completion from reduced dynamics?

  2. Was the blending ratio selected using separate validation data? Can paired confidence intervals and additional seeds establish the small merging gains?

  3. How do the 100 training task types relate to VBVR-Bench’s 50 ID and 50 supposedly held-out OOD types? Please provide the task mapping, exclusion rules, and training-step/checkpoint-selection details needed to verify the controlled comparison.

6 comments

We thank the reviewer for the supportive review and the concrete questions. We also thank the reviewer for recognizing our experiments as convincing evidence that reasoning data can benefit other generation tasks. We answer each point below and support the answers with new experiments.

W1. Instruction-following-only training outperforms reasoning-only training on all three application domains, so it is unclear what reasoning contributes beyond improved instruction following.

Reasoning data still generalizes after instruction following is removed. Sec. 5.4 was designed to strip instruction following out of the reasoning data, not to rank the two subsets. Every task whose answer is fully specified in the prompt went into the instruction-following subset, so all 47 remaining reasoning task types require inferring an unstated target state. Trained on these alone, Wan2.2 still improves all three application domains by a large margin (numbers from Fig. 10 / Table 7):

Training data (50K clips, same budget)PerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
Reasoning-only subset41.7 (+28.3%)47.4 (+16.5%)64.5 (+4.0%)51.2 (+13.6%)
Instruction-following subset47.149.968.555.2
Full corpus46.249.168.654.6

Data with no instruction-following task still produces large cross-domain gains, with perception up 28.3% and embodied manipulation up 16.5% over the baseline. The gain therefore cannot come from instruction execution alone. As stated in Sec. 6, we attribute this gain to empowering the model with the ability to find specific methods to achieve a goal under specific requirements, rather than simply rendering. We will rephrase Sec. 5.4 to make this intent explicit.

Two further controls show that the transfer comes from reasoning, not from explicitly stated targets or from VBVR itself. First, we retrained on the same VBVR tasks and videos with prompts rewritten so that the target is no longer stated and must be inferred. For example, the original sequence-completion prompt (G-131) states that the shapes keep the same shape and color and grow by a constant size step; the revised prompt only says that "the shapes form a sequence" and asks for the option that "continues the same pattern". Transfer stays at about the same level, and in several settings it is higher than with the original prompts. On Wan2.2, the three-domain average rises from 54.6 to 55.4:

BackboneModelVBVR-BenchPerceptionEmbodiedSimulation (cal.)Average (3 domains)
Wan2.2Baseline37.132.540.762.045.1
Wan2.2Reasoning FT (original prompts)61.846.249.168.654.6
Wan2.2Reasoning FT (revised prompts)62.246.451.168.755.4
LTX-2.3Baseline35.128.835.258.740.9
LTX-2.3Reasoning FT (original prompts)53.036.539.366.047.3
LTX-2.3Reasoning FT (revised prompts)52.039.836.265.747.2

Second, we fine-tuned Wan2.2 on about 50K clips from four video-reasoning sources other than VBVR (VideoRLVR, PhyWorld, OpenCoF, VR-Bench). The model that never sees VBVR still beats the baseline on all three application domains, with +6.0 on perception and +8.5 on embodied manipulation:

Wan2.2 training dataPerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
VBVR only (Reasoning FT)46.249.168.654.6
Full mixture (VBVR + 5 other sources)42.551.466.053.3
Non-VBVR sources only38.549.264.450.7

Together, these controls show that the generalization comes from the reasoning content of the data: it survives the removal of instruction-following tasks, the removal of explicitly stated targets, and the removal of VBVR altogether.

Q1a. How does ReasonGraft perform when the reasoning adapter is replaced by the instruction-following-only adapter?

ReasonGraft also works with the instruction-following-only adapter.

Wan2.2PerceptionEmbodiedSimulation
In-domain FT78.748.869.4
In-domain LoRA + full reasoning adapter78.952.369.7
In-domain LoRA + instruction-following adapter78.752.170.0

Replacing the full reasoning adapter with the instruction-following-only adapter gives similar results: on par with the in-domain specialist on perception and simulation, and above it on embodied manipulation (52.1 vs. 48.8). This is consistent with our answer to W1 above: both components of the reasoning corpus transfer, so either one grafts well onto an in-domain adapter.

Q1b. Can motion-controlled comparisons distinguish improved task completion from reduced dynamics?

The gains are not explained by reduced motion. To test whether the gains come from reduced dynamics, we measured motion as the mean Farneback optical-flow magnitude over 17 uniformly sampled frames, and compared the reasoning model with the baseline on motion-matched test items, where the two videos' motion differs by at most a factor of 1.25 (strict) or 1.5 (loose):

DomainBackboneItemsGain, all itemsGain, matched ×1.25Gain, matched ×1.5
PerceptionWan2.2300+13.7+7.3 (n=30)+8.0 (n=66)
PerceptionLTX-2.3298+7.7+8.3 (n=19)+6.4 (n=43)
EmbodiedWan2.2250+8.4+5.6 (n=47)+7.1 (n=94)
EmbodiedLTX-2.3250+4.1+2.3 (n=77)+0.2 (n=123)
SimulationWan2.2200+3.1+0.2 (n=100)+1.5 (n=142)
SimulationLTX-2.3200+2.8+1.7 (n=54)+2.5 (n=81)

The reasoning model does move less on perception and embodied items. But on motion-matched subsets, the perception gain remains significant on both backbones (+8.0 and +6.4 at the 1.5 tolerance), and the embodied gain on Wan2.2 remains significant at the 1.5 tolerance (+7.1). On simulation, the reasoning model does not move less than the baseline. Improved task completion is therefore not explained by reduced dynamics.

W2 / Q2. ReasonGraft uses standard adapter blending, and its gains over in-domain fine-tuning are marginal, without uncertainty estimates. Was the blending ratio selected using separate validation data?

Held-out validation data select 3:7 on both backbones. We built a validation split for each application domain, disjoint from the test sets, and swept the ratio grid {1:9, 3:7, 5:5, 7:3, 9:1} (reasoning : in-domain) on both backbones.

  • Simulation: 200 instances stratified from WorldScore static/stylized, which the test set never uses.
  • Perception: 300 samples (50 per task) from the perception pool that were not used for training or evaluation.
  • Embodied: the 200-item embodiment half of RBench, which shares no image or prompt with the 5-task test half used in our paper. It is scored with the official RBench VQA rubric.

Each cell is the validation score. In each row the best ratio is in bold and the second best is marked with *.

Validation setBackbone1:93:75:57:39:1
SimulationWan2.266.467.666.9*64.762.4
SimulationLTX-2.366.666.5*65.659.157.9
PerceptionWan2.277.3*77.676.769.751.1
PerceptionLTX-2.367.263.8*55.346.838.8
EmbodiedWan2.274.577.175.677.0*76.8
EmbodiedLTX-2.368.470.569.6*68.061.4
Average—70.1*70.568.364.258.1

3:7 is the best ratio in four of the six settings and the second best in the other two, so it is clearly the strongest choice overall. It also has the highest average across the six settings (70.5). We will report the validation protocol and this table in the revision.

The same ratio holds on the test sets of both backbones. Test-set sweep from the reasoning-only endpoint (10:0) to the in-domain endpoint (0:10):

BackboneDomain10:07:35:53:70:10
Wan2.2Perception46.272.977.578.978.7
Wan2.2Embodied49.148.752.152.348.8
Wan2.2Simulation68.669.770.069.769.4
LTX-2.3Perception36.550.758.069.168.7
LTX-2.3Embodied39.344.342.648.840.9
LTX-2.3Simulation66.069.069.870.570.4
Average—51.059.261.764.962.8

3:7 is the best ratio in five of six columns and within 0.3 of the best in the sixth, and it is above both endpoints in every column, on both backbones. On average, 3:7 reaches 64.9, against 62.8 for in-domain fine-tuning alone. The LTX-2.3 5:5 and 7:3 models were run during the rebuttal.

The gains hold across generation seeds and paired confidence intervals. We add two kinds of uncertainty. (1) Generation seeds: the 3:7 models are regenerated on the full test sets with two more sampling seeds, giving mean ± s.d. over 3 seeds. (2) Item sampling: a paired bootstrap over test items (10,000 resamples, stratified by task) of the difference between ReasonGraft and the in-domain specialist.

BackboneDomainReasonGraft 3:7, seeds 0 / 1 / 2Mean ± s.d.In-domain FTΔ vs. in-domain, paired 95% CI (seed 0)
Wan2.2Perception78.9 / 78.2 / 78.278.4 ± 0.478.7+0.2 [−0.5, +1.0]
Wan2.2Embodied52.3 / 49.5 / 51.251.0 ± 1.448.8+3.5 [−0.5, +7.6]
Wan2.2Simulation69.7 / 69.2 / 69.169.4 ± 0.369.4—
LTX-2.3Perception69.1 / 68.5 / 69.068.9 ± 0.368.7+0.6 [−0.2, +1.1]
LTX-2.3Embodied48.8 / 44.8 / 43.445.7 ± 2.840.9+7.9 [+4.2, +11.6]
LTX-2.3Simulation70.5 / 70.6 / 69.970.3 ± 0.470.4—

Averaged over three seeds, ReasonGraft is on par with or above the in-domain specialist on perception and simulation. It is clearly above the specialist on embodied manipulation: +2.2 on Wan2.2 and +4.8 on LTX-2.3, higher for every seed, and on LTX-2.3 the paired 95% CI excludes zero. ReasonGraft therefore never falls meaningfully below a dedicated specialist and gains where the reasoning component matters most. It gets there without any training: the reasoning adapter is trained once, reused for every target domain, and blended at one fixed ratio that is not tuned per task or per backbone. (The paired CI for simulation is not reported because per-instance Simulation FT scores were not saved.)

The added value lies in what is merged. The blending operator is standard by design. What matters is which adapter is added (a reasoning adapter, trained once and reused for every target domain) and at what ratio. Table 3 of the paper shows that this plain ratio blend surpasses or matches more complex merging methods and clearly beats joint training on the mixed data.

Reviewer_NUq9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how fine-tuning a video generator in one domain affects its performance in other domains. It assembles XVReason and XVReason-Bench from reasoning, perception, embodied manipulation, and world-simulation resources. Separate fine-tuning runs on Wan2.2 and LTX-2.3 produce a four-by-four transfer matrix. Reasoning fine-tuning improves all four domain scores under the tested setup. The paper also studies adaptation, data scale, denoising experts, and task composition. Finally, ReasonGraft combines a reasoning LoRA with a target-domain LoRA using a 3:7 weighted sum. It reports gains over target-domain fine-tuning in three separate target-domain settings.

Critical strengths
  1. The full transfer matrices on two backbones provide a useful comparison. Testing every domain-specific model across all four domains is more informative than reporting only improvements on its training domain. The repeated direction of the main reasoning-transfer result makes this an observation worth studying.

  2. The paper makes a useful effort to test the limits of the effect. The adaptation, data-scale, expert-loading, and task-composition experiments go beyond a single successful fine-tuning run. The combined training and evaluation suite could also help others study transfer more systematically.

Critical weaknesses

1. The contribution is not yet strong enough as either a study of transfer or a new method.

As a transfer study, the paper broadens an existing observation: VBVR already reported improved camera-motion alignment and scene consistency after reasoning fine-tuning [1]. Section 5.4 also shows that the instruction-following subset transfers better than the reasoning subset on all three application domains. Gains from the reasoning subset do not isolate their cause, since these examples also train instruction execution, goal completion, and scene preservation. The benefit may therefore reflect other properties of VBVR rather than reasoning itself.

To address this, compare matched tasks with inferred versus explicitly supplied targets, keeping visual content and training budgets comparable. Test another reasoning-data source or narrow the conclusion to VBVR, and distinguish the new findings from prior VBVR results.

As a method, ReasonGraft uses a standard weighted sum of LoRA updates. Simplicity alone is not a weakness, but Table 3 evaluates three separate reasoning-plus-target models, not one model combining all target domains. This does not establish that the method remains useful as more domains are combined. Existing merging baselines also match or exceed it in some columns, leaving a limited method contribution beyond applying the transfer observation.

To address this, combine one, two, and three target domains and evaluate each model on every domain. Compare adding the reasoning adapter with target-only merging, existing merging methods, and joint training under matched budgets. Select weights on held-out data and report average performance and the largest domain-specific drop. Otherwise, limit the method claim to the tested two-adapter setting.

2. Missing directly relevant baselines and limited robustness checks weaken the experimental conclusions.

VBVR-Wan2.2 [1] is the clearest omission: it shares the backbone, training-data source, and reasoning benchmark, yet Table 9 imports other models’ scores while omitting this checkpoint. Relevant embodied-generation baselines include Cosmos3-Nano, announced as the RBench open-source leader on June 1, 2026 [2], and Wan2.6 and Seedance 1.5 Pro, respectively the leader and a strong competitor in the Cosmos 3 report’s RBench comparison [3]. All were public before June 26, 2026; the latter two were released in December 2025 [4, 5]. They fall outside the three-month comparison exemption before the September 26 deadline.

To address this, evaluate VBVR-Wan2.2 on reasoning and transfer, and Cosmos3-Nano and Wan2.6 on embodied generation, with Seedance 1.5 Pro as another relevant comparison. Match prompts, subsets, scoring, and generation settings where feasible; explain any access or compatibility barriers. RBench aggregate scores cannot substitute for evaluation under Table 11’s task-completion protocol. Robustness is also unclear. In Tables 12–13, Perception FT raises Wan2.2’s original simulation score from 63.5 to 65.4 but lowers its calibrated score from 62.0 to 61.4, so some transfer directions depend on aggregation. Each prompt is generated once with a fixed seed, while Table 3’s ReasonGraft gains are only 0.2 points for perception and 0.3 for simulation. Without uncertainty estimates, these gains do not establish a reliable advantage.

To address this, report sensitivity to the original metrics and calibration weights. Repeat key comparisons across generation seeds, report paired confidence intervals, and repeat training seeds where feasible. Prioritize these checks over further small ablations.

Sources cited above:

[1] A Very Big Video Reasoning Suite. February 23, 2026; original version, Sections 5.3–5.4.

[2] RBench official repository. June 1, 2026 announcement for Cosmos3-Nano.

[3] Cosmos 3: Omnimodal World Models for Physical AI. June 1, 2026; Table 12.

[4] Alibaba announcement of Wan2.6. December 16, 2025.

[5] ByteDance announcement of Seedance 1.5 Pro. December 16, 2025.

Questions for authors
  1. Clarification: What evidence separates reasoning-specific transfer from improved instruction following, given that the instruction-following subset performs better on all three application domains? Please identify which existing control supports that distinction.

  2. Scope clarification: Is ReasonGraft intended only for separate target-domain specialists, or for a single model combining several domains? Are there existing experiments for the latter? The domain-count experiment suggested above is an additional experiment, not a request to relabel the existing pairwise results.

  3. Evaluation clarification: Why is the public VBVR-Wan2.2 checkpoint absent, and which training or evaluation differences explain the gap between its published results and the corresponding runs here? Please also explain the exclusion of the RBench competitors listed above.

  4. Robustness clarification: Which central conclusions remain true under the original WorldScore aggregation, and what uncertainty estimates support the small ReasonGraft gains? Please explain how the calibration weights and the 3:7 merge ratio were selected, including whether test results informed those choices.

7 comments

We thank the reviewer for the detailed and constructive review. We also thank the reviewer for recognizing that the generalization result is an observation worth studying, and that our adaptation, data-scale, expert-loading and task-composition experiments are solid. We answer each point below and support the answers with the new experiments the reviewer asked for.

W1. Distinguish the characteristics of reasoning from other properties of the data

W1 (prior work). VBVR already reported improved camera-motion alignment and scene consistency after reasoning fine-tuning [1]. The new findings should be distinguished from the prior VBVR results.

VBVR offers valuable first observations; our paper studies these abilities more completely and systematically, and adds new findings. VBVR [1] Sec. 5.3 qualitatively analyzes the fine-tuned model on held-out reasoning tasks, and Sec. 5.4 compares one checkpoint with the base model on VBench-I2V, reporting a higher Video-Text Camera Motion score. Building on these observations, our paper measures camera control and scene consistency directly and under controls: camera control as executing a specified multi-step camera path, scored against the estimated camera trajectory, and scene consistency through 3D re-projection, photometric, and style consistency (WorldScore). Beyond measuring these abilities, Sec. 6 analyzes what changes after reasoning fine-tuning and classifies the new behavior into three competences. We tie these competences to the quantitative gains in 3 downstream benchmarks. These findings are beyond what was already covered by VBVR.

W1 / Q1 (role of reasoning). Sec. 5.4 shows that the instruction-following subset transfers better than the reasoning subset on all three application domains. Which existing control separates reasoning-specific transfer from improved instruction following?

The transfer comes from reasoning, not from instruction following or VBVR-specific properties. We address this with three controls. (1) is the existing control in the paper; (2) and (3) are the two experiments the reviewer suggested.

(1) Reasoning data still generalizes after instruction following is removed. Sec. 5.4 was designed to strip instruction following out of the reasoning data, not to rank the two subsets. Every task whose answer is fully specified in the prompt went into the instruction-following subset, so all 47 remaining reasoning task types require inferring an unstated target state. Trained on these alone, Wan2.2 still improves all three application domains by a large margin (numbers from Fig. 10 / Table 7):

Training data (50K clips, same budget)PerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
Reasoning-only subset41.7 (+28.3%)47.4 (+16.5%)64.5 (+4.0%)51.2 (+13.6%)
Instruction-following subset47.149.968.555.2
Full corpus46.249.168.654.6

Data with no instruction-following task still produces large cross-domain gains, with perception up 28.3% and embodied manipulation up 16.5% over the baseline. The gain therefore cannot come from instruction execution alone. As stated in Sec. 6, we attribute this gain to empowering the model with the ability to find specific methods to achieve a goal under specific requirements, rather than simply rendering. We will rephrase Sec. 5.4 to make this intent explicit.

(2) Transfer persists when targets must be inferred rather than stated. Following the reviewer's suggestion, we retrained on the same VBVR tasks and videos, with the same training budget, with prompts rewritten so that the target is no longer stated and must be inferred. An example of the rewrite:

TaskOriginal prompt: rule and target statedRevised prompt: rule must be inferred
Two-step analogy (O-12)“The scene shows an analogy A→B→C :: D→?→? with two rows of shapes. Top row: color_042 rectangle at extra_large becomes color_073 rectangle at extra_large, then color_073 rectangle at large. This shows two rules: first change color from color_042 to color_073, then change size from extra_large to large. Bottom row: minus starts with color_042 and extra_large. Apply the same transformation: first change color to color_073, then change size to large.”“The scene shows an analogy A→B→C :: D→?→? with two rows of shapes. The top row shows a two-step transformation. Apply the same two-step transformation to the bottom row's shape.”
BackboneModelVBVR-BenchPerceptionEmbodiedSimulation (cal.)Average (3 domains)
Wan2.2Baseline37.132.540.762.045.1
Wan2.2Reasoning FT (original prompts)61.846.249.168.654.6
Wan2.2Reasoning FT (revised prompts)62.246.451.168.755.4
LTX-2.3Baseline35.128.835.258.740.9
LTX-2.3Reasoning FT (original prompts)53.036.539.366.047.3
LTX-2.3Reasoning FT (revised prompts)52.039.836.265.747.2

Overall, rewriting the prompts so that targets must be inferred leaves the transfer at about the same level, and in some settings it is even higher than with the original prompts, e.g. on VBVR-Bench, perception and embodied manipulation for Wan2.2, and on perception for LTX-2.3. On Wan2.2, the three-domain average even rises from 54.6 to 55.4. The generalization does not come from a specific property of the original VBVR prompts, nor from the instruction-following signal they contain: once the explicitly stated targets are removed, the model still generalizes as well or better.

(3) Reasoning data without any VBVR still generalizes. We assembled a 200K-clip video-reasoning mixture from six sources:

SourceClipsTask types
VBVR114,300100
VBVR-multistep36,00036
VideoRLVR18,0003
PhyWorld18,0003
OpenCoF7,4058
VR-Bench6,3365
Total200,041155

We fine-tune Wan2.2 on the full mixture and on the four non-VBVR sources alone (VideoRLVR, PhyWorld, OpenCoF, VR-Bench; ~50K clips).

Wan2.2 training dataPerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
VBVR only (Reasoning FT)46.249.168.654.6
Full mixture (VBVR + 5 other sources)42.551.466.053.3
Non-VBVR sources only38.549.264.450.7

The model that never sees VBVR still beats the baseline on all three application domains. Perception improves by +6.0 and embodied manipulation by +8.5. Both gains are significant. Transfer is therefore not a property specific to VBVR, and we do not need to narrow the conclusion to VBVR.

W1 (method). ReasonGraft as a method

W1 (method) / Q2. ReasonGraft uses a standard weighted sum of LoRA updates, and Table 3 evaluates three separate reasoning-plus-target models rather than one model combining all target domains.

  • One merged model improves all three target domains. As submitted, ReasonGraft produces one specialist per target domain. We have now merged the reasoning adapter with all three in-domain adapters into a single adapter. On Wan2.2, one merged model is significantly above the baseline on perception, embodiment, and simulation. Its largest drop relative to the three specialists is under 3 points, and it exceeds the embodied and simulation specialists (details below).
  • The contribution lies in what is merged, not in the operator. We agree the operator is not where the contribution lies: linear merging, RobustMerge and LoRA-LEGO give similar results in Table 3 and in the multi-domain merge. What matters is which adapter is added and at what ratio, which held-out validation data support (see below).

A single merged model matches or exceeds all 3 specialists. We merged the reasoning LoRA with all three in-domain LoRAs (perception, embodied, simulation) into one adapter and evaluated that single model on every application domain, using two merging operators on Wan2.2: linear task arithmetic with coefficient 0.5 per adapter, and RobustMerge (default settings).

Wan2.2, single merged modelPerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
Linear 0.5 × 476.8*49.0*71.1*65.6*
RobustMerge76.049.771.365.7
In-domain FT for each column78.748.869.465.6*

In each column, the best result is in bold and the second best is marked with *.

With either operator, one merged model significantly improves all three target domains over the baseline. Relative to the three per-domain specialists, the linear 0.5 merge is −1.9 on perception, +0.2 on embodied and +1.7 on simulation; RobustMerge is −2.7, +0.9 and +1.9. On average over the three domains, the single merged model (65.7) matches the three separate specialists (65.6).

Held-out validation data select 3:7. We built a validation split for each application domain, disjoint from the test sets, and swept the ratio grid {1:9, 3:7, 5:5, 7:3, 9:1} (reasoning : in-domain) on both backbones.

  • Simulation: 200 instances stratified from WorldScore static/stylized, which the test set never uses.
  • Perception: 300 samples (50 per task) from the perception pool that were not used for training or evaluation.
  • Embodied: the 200-item embodiment half of RBench, which shares no image or prompt with the 5-task test half used in our paper. It is scored with the official RBench VQA rubric.

Each cell is the validation score. In each row the best ratio is in bold and the second best is marked with *.

Validation setBackbone1:93:75:57:39:1
SimulationWan2.266.467.666.9*64.762.4
SimulationLTX-2.366.666.5*65.659.157.9
PerceptionWan2.277.3*77.676.769.751.1
PerceptionLTX-2.367.263.8*55.346.838.8
EmbodiedWan2.274.577.175.677.0*76.8
EmbodiedLTX-2.368.470.569.6*68.061.4
Average—70.1*70.568.364.258.1

3:7 is the best ratio in four of the six settings and the second best in the other two, so it is clearly the strongest choice overall. It also has the highest average across the six settings (70.5). We will report the validation protocol, this table, and the multi-domain merge in the revision.

The same ratio holds on the test sets of both backbones. Test-set sweep from the reasoning-only endpoint (10:0) to the in-domain endpoint (0:10):

BackboneDomain10:07:35:53:70:10
Wan2.2Perception46.272.977.578.978.7
Wan2.2Embodied49.148.752.152.348.8
Wan2.2Simulation68.669.770.069.769.4
LTX-2.3Perception36.550.758.069.168.7
LTX-2.3Embodied39.344.342.648.840.9
LTX-2.3Simulation66.069.069.870.570.4
Average—51.059.261.764.962.8

3:7 is the best ratio in five of six columns and within 0.3 of the best in the sixth, and it is above both endpoints in every column, on both backbones. On average, 3:7 reaches 64.9, against 62.8 for in-domain fine-tuning alone. Table 3 of the paper further shows that this plain ratio blend surpasses or matches more complex merging methods and clearly beats joint training on the mixed data. A fixed-ratio blend therefore works across tasks and backbones, which makes it both a meaningful finding and a practical method.

Reviewer_jUUNOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates whether training video generation models on abstract reasoning tasks improves their capabilities in other domains. The authors construct XVReason and XVReason-Bench, covering reasoning, perception, embodied manipulation, and world simulation, and conduct a 4 × 4 cross-domain transfer study on Wan2.2 and LTX-2.3. They also introduce ReasonGraft, which combines reasoning and in-domain LoRA adapters through weighted parameter averaging. The experiments show positive transfer from reasoning training to all four domains on both backbones.

Critical strengths

S1. A meaningful research question. The paper examines how training on one capability domain affects other capabilities of video generation models. Understanding these transfer relationships could help guide data selection and adaptation strategies for general-purpose video generators. S2. Systematic cross-domain experiments. The complete 4 × 4 transfer matrices on two backbones provide useful empirical evidence. On Wan2.2, reasoning fine-tuning improves perception, embodied manipulation, and calibrated simulation by 42.3%, 20.6%, and 10.7%, respectively, relative to the baseline. It is also the only training source that improves all four domains on both backbones. S3. Useful analyses of transfer behavior. The paper examines adaptation regimes, data scale, denoising stages, task composition, and visual distribution shifts. These analyses help characterize the observed transfer effects and their conditions. S4. A simple and practical adapter combination strategy. ReasonGraft reuses a reasoning LoRA trained once and requires no additional training during merging. Compared with in-domain fine-tuning, it improves embodied manipulation scores by 3.5 points on Wan2.2 and 7.9 points on LTX-2.3.

Critical weaknesses

W1. The specific contribution of reasoning to cross-domain transfer remains unclear. The task-composition analysis in Section 5.4 shows that instruction-following training outperforms reasoning-only training on all three application domains, including perception (47.1 vs. 41.7) and embodied manipulation (49.9 vs. 47.4) on Wan2.2. Although reasoning-only training also produces positive transfer, the results leave open how much of the full corpus's benefit comes from abstract reasoning itself. This limits the interpretation of the paper's central finding. W2. Limited methodological novelty and uncertain small gains for ReasonGraft. ReasonGraft uses a straightforward weighted combination of LoRA updates (Eq. 3). Its contribution is primarily empirical. While the embodied manipulation gains are meaningful, improvements over in-domain fine-tuning are only 0.2 points on perception and 0.3 points on calibrated simulation in Table 3, without uncertainty estimates. Moreover, the 3:7 ratio is selected from a sweep across the Wan2.2 target domains, leaving its broader applicability less established. W3. The world simulation gains depend strongly on the calibrated evaluation metric. On Wan2.2, the uncalibrated mean of the seven WorldScore aspects improves from 63.5 to 64.7, while the calibrated score increases from 62.0 to 68.6 (Tables 12–13). Under Eq. 5, the nonlinear camera-score transformation accounts for most of the larger gain. Since the evaluation is restricted to static scenes, the evidence primarily supports improved camera-motion control. A clearer discussion of the calibration's effect and the scope of the simulation-related conclusions would strengthen the paper.

Questions for authors

Q1. Given that instruction-following training transfers better than reasoning-only training on the three application domains, how do the authors interpret the specific role of abstract reasoning in the observed generalization? Q2. Could the authors provide uncertainty estimates for the smaller ReasonGraft improvements using the existing evaluation samples? Also, was the 3:7 ratio selected using the same target-domain evaluation sets reported in Table 3?

5 comments

We thank the reviewer for the positive assessment and the precise questions. We also thank the reviewer for recognizing our research question as meaningful and the complete generalization matrices on two backbones as useful empirical evidence. We answer each point below and support the answers with new experiments.

W1 / Q1. Sec. 5.4 shows that instruction-following training outperforms reasoning-only training. How much of the transfer comes from abstract reasoning itself, and how do the authors interpret its role?

Reasoning data still generalizes after instruction following is removed. Sec. 5.4 was designed to strip instruction following out of the reasoning data, not to rank the two subsets. Every task whose answer is fully specified in the prompt went into the instruction-following subset, so all 47 remaining reasoning task types require inferring an unstated target state. Trained on these alone, Wan2.2 still improves all three application domains by a large margin (numbers from Fig. 10 / Table 7):

Training data (50K clips, same budget)PerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
Reasoning-only subset41.7 (+28.3%)47.4 (+16.5%)64.5 (+4.0%)51.2 (+13.6%)
Instruction-following subset47.149.968.555.2
Full corpus46.249.168.654.6

Data with no instruction-following task still produces large cross-domain gains, with perception up 28.3% and embodied manipulation up 16.5% over the baseline. The gain therefore cannot come from instruction execution alone. As stated in Sec. 6, we attribute this gain to empowering the model with the ability to find specific methods to achieve a goal under specific requirements, rather than simply rendering. We will rephrase Sec. 5.4 to make this intent explicit.

Two further controls show that the transfer comes from reasoning, not from explicitly stated targets or from VBVR itself. First, we retrained on the same VBVR tasks and videos with prompts rewritten so that the target is no longer stated and must be inferred. For example, the original sequence-completion prompt (G-131) states that the shapes keep the same shape and color and grow by a constant size step; the revised prompt only says that "the shapes form a sequence" and asks for the option that "continues the same pattern". Transfer stays at about the same level, and in several settings it is higher than with the original prompts. On Wan2.2, the three-domain average rises from 54.6 to 55.4:

BackboneModelVBVR-BenchPerceptionEmbodiedSimulation (cal.)Average (3 domains)
Wan2.2Baseline37.132.540.762.045.1
Wan2.2Reasoning FT (original prompts)61.846.249.168.654.6
Wan2.2Reasoning FT (revised prompts)62.246.451.168.755.4
LTX-2.3Baseline35.128.835.258.740.9
LTX-2.3Reasoning FT (original prompts)53.036.539.366.047.3
LTX-2.3Reasoning FT (revised prompts)52.039.836.265.747.2

Second, we fine-tuned Wan2.2 on about 50K clips from four video-reasoning sources other than VBVR (VideoRLVR, PhyWorld, OpenCoF, VR-Bench). The model that never sees VBVR still beats the baseline on all three application domains, with +6.0 on perception and +8.5 on embodied manipulation:

Wan2.2 training dataPerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
VBVR only (Reasoning FT)46.249.168.654.6
Full mixture (VBVR + 5 other sources)42.551.466.053.3
Non-VBVR sources only38.549.264.450.7

Together, these controls show that the generalization comes from the reasoning content of the data: it survives the removal of instruction-following tasks, the removal of explicitly stated targets, and the removal of VBVR altogether.

W2 / Q2. The gains of ReasonGraft over in-domain fine-tuning in Table 3 are marginal and without uncertainty estimates. The 3:7 ratio was selected from a sweep across the Wan2.2 target domains, so its broader applicability is less established. Was the 3:7 ratio selected on the same evaluation sets reported in Table 3?

The contribution lies in what is merged, and the simple blend matches more complex merging methods. We agree that the operator is simple. What matters is which adapter is added and at what ratio. Table 3 of the paper shows that this plain ratio blend surpasses or matches more complex merging methods (LoRA-LEGO and RobustMerge) and clearly beats joint training on the mixed data.

The gains hold across generation seeds. We add two kinds of uncertainty. (1) Generation seeds: the 3:7 models are regenerated on the full test sets with two more sampling seeds, giving mean ± s.d. over 3 seeds. (2) Item sampling: a paired bootstrap over test items (10,000 resamples, stratified by task) of the difference between ReasonGraft and the in-domain specialist.

BackboneDomainReasonGraft 3:7, seeds 0 / 1 / 2Mean ± s.d.In-domain FTΔ vs. in-domain, paired 95% CI (seed 0)
Wan2.2Perception78.9 / 78.2 / 78.278.4 ± 0.478.7+0.2 [−0.5, +1.0]
Wan2.2Embodied52.3 / 49.5 / 51.251.0 ± 1.448.8+3.5 [−0.5, +7.6]
Wan2.2Simulation69.7 / 69.2 / 69.169.4 ± 0.369.4—
LTX-2.3Perception69.1 / 68.5 / 69.068.9 ± 0.368.7+0.6 [−0.2, +1.1]
LTX-2.3Embodied48.8 / 44.8 / 43.445.7 ± 2.840.9+7.9 [+4.2, +11.6]
LTX-2.3Simulation70.5 / 70.6 / 69.970.3 ± 0.470.4—

Averaged over three seeds, ReasonGraft is on par with or above the in-domain specialist on perception and simulation. It is clearly above the specialist on embodied manipulation: +2.2 on Wan2.2 and +4.8 on LTX-2.3, higher for every seed, and on LTX-2.3 the paired 95% CI excludes zero. ReasonGraft therefore never falls meaningfully below a dedicated specialist and gains where the reasoning component matters most. It gets there without any training: the reasoning adapter is trained once, reused for every target domain, and blended at one fixed ratio that is not tuned per task or per backbone. (The paired CI for simulation is not reported because per-instance Simulation FT scores were not saved.)

Held-out validation data select 3:7 on both backbones. We built a validation split for each application domain, disjoint from the test sets, and swept the ratio grid {1:9, 3:7, 5:5, 7:3, 9:1} (reasoning : in-domain) on both backbones.

  • Simulation: 200 instances stratified from WorldScore static/stylized, which the test set never uses.
  • Perception: 300 samples (50 per task) from the perception pool that were not used for training or evaluation.
  • Embodied: the 200-item embodiment half of RBench, which shares no image or prompt with the 5-task test half used in our paper. It is scored with the official RBench VQA rubric.

Each cell is the validation score. In each row the best ratio is in bold and the second best is marked with *.

Validation setBackbone1:93:75:57:39:1
SimulationWan2.266.467.666.9*64.762.4
SimulationLTX-2.366.666.5*65.659.157.9
PerceptionWan2.277.3*77.676.769.751.1
PerceptionLTX-2.367.263.8*55.346.838.8
EmbodiedWan2.274.577.175.677.0*76.8
EmbodiedLTX-2.368.470.569.6*68.061.4
Average—70.1*70.568.364.258.1

3:7 is the best ratio in four of the six settings and the second best in the other two, so it is clearly the strongest choice overall. It also has the highest average across the six settings (70.5). We will report the validation protocol and this table in the revision.

The same ratio holds on the test sets of both backbones. Test-set sweep from the reasoning-only endpoint (10:0) to the in-domain endpoint (0:10):

BackboneDomain10:07:35:53:70:10
Wan2.2Perception46.272.977.578.978.7
Wan2.2Embodied49.148.752.152.348.8
Wan2.2Simulation68.669.770.069.769.4
LTX-2.3Perception36.550.758.069.168.7
LTX-2.3Embodied39.344.342.648.840.9
LTX-2.3Simulation66.069.069.870.570.4
Average—51.059.261.764.962.8

3:7 is the best ratio in five of six columns and within 0.3 of the best in the sixth, and it is above both endpoints in every column, on both backbones. On average, 3:7 reaches 64.9, against 62.8 for in-domain fine-tuning alone. The LTX-2.3 5:5 and 7:3 models were run during the rebuttal. A fixed-ratio blend therefore works across tasks and backbones, which makes it both a meaningful finding and a practical method.

W3. The world-simulation gain depends strongly on calibration.

Near-static videos hack the original WorldScore benchmark, which is why calibration is needed. We built videos that contain no world simulation at all, and scored them with the full WorldScore pipeline on 200 test instances. freeze repeats the input image; jitter adds sub-pixel shake and mild noise to it.

VideoCameraObjectContentSubjective3DPhotometricStyleMeanCalibrated
freeze (input image repeated)19.855.337.564.896.998.3100.067.557.2
jitter (input + small shake and noise)19.843.737.265.494.297.697.165.054.6
Wan2.2 baseline34.739.755.860.580.787.586.663.665.2
Wan2.2 Reasoning FT37.041.656.559.780.488.487.664.568.3

Under the absolute protocol, a frozen image outscores every real model (67.5 vs. 64.5 for Reasoning FT), because three of the seven aspects (3D, photometric, style consistency) are frame-to-frame stability measures that a static clip maximizes. Calibration ranks both static videos last. The uncalibrated mean therefore understates a model that actually moves the camera as instructed, and the small change in the uncalibrated mean is not evidence that the gain is weak.

The gain is not only camera control. Splitting the calibrated score into its non-camera parts (Wan2.2, Table 12):

Wan2.2Camera scoreAbility without cameraConsistency
Baseline32.553.683.6
Reasoning FT36.853.984.8

Both non-camera groups are above the baseline (e.g., consistency 84.8 vs. 83.6), so improved camera control is not the only change.

The conclusion does not depend on the calibration weights. We recomputed the simulation score under the original absolute mean, and under calibrated variants with ability weight w ∈ {0.5, 0.6, 0.7, 0.8} and logistic temperature τ ∈ {1.5, 2.5, 5, 10} (16 settings per backbone). Reasoning FT > baseline holds under the absolute mean and in all 32 calibrated settings. The weights (w = 0.7, τ = 2.5) were fixed before any fine-tuned model was evaluated.

Scope. In the revision, we will use the calibrated score as the single simulation score, explain why it is needed, and restrict the simulation conclusions to camera-controlled rollout of static scenes.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response (1/2)

We thank all reviewers for their thoughtful reviews. We are glad they found the question meaningful, the cross-domain study systematic and convincing, and the analyses informative.

We believe most concerns stem from misunderstandings of parts of our design, in particular the purpose of the task-composition experiment and the role of calibration in WorldScore. We clarify all of them below and support the clarifications with new experiments. Three concerns are shared: (i) whether the gain comes from reasoning or from instruction following (G1); (ii) whether ReasonGraft's ratio and gains are reliable (G2, G3); and (iii) why we implemented calibration in WorldScore, and whether it influences the judgment of performance gain (G4). Each per-reviewer response below is self-contained and gives the full answer to that reviewer's questions.

G1. Isolated from instruction following, reasoning itself generalizes

Reasoning data still generalizes after instruction following is removed. Sec. 5.4 was designed to strip instruction following out of the reasoning data, not to rank the two subsets. Every task whose answer is fully specified in the prompt went into the instruction-following subset, so all 47 remaining reasoning task types require inferring an unstated target state. Trained on these alone, Wan2.2 still improves all three application domains by a large margin (numbers from Fig. 10 / Table 7):

Training data (50K clips, same budget)PerceptionEmbodiedSimulationAverage
Baseline32.540.762.045.1
Reasoning-only subset41.7 (+28.3%)47.4 (+16.5%)64.5 (+4.0%)51.2 (+13.6%)
Instruction-following subset47.149.968.555.2
Full corpus46.249.168.654.6

Data with no instruction-following task still produces large cross-domain gains, with perception up 28.3% and embodied manipulation up 16.5% over the baseline. The gain therefore cannot come from instruction execution alone. As stated in Sec. 6, we attribute this gain to empowering the model with the ability to find specific methods to achieve a goal under specific requirements, rather than simply rendering. We will rephrase Sec. 5.4 to make this intent explicit.

G2. The 3:7 ratio was selected on held-out validation data and holds across backbones and domains

Held-out validation data select 3:7. We built a validation split for each application domain, disjoint from the test sets, and swept the ratio grid {1:9, 3:7, 5:5, 7:3, 9:1} (reasoning : in-domain) on both backbones.

  • Simulation: 200 instances stratified from WorldScore static/stylized, which the test set never uses.
  • Perception: 300 samples (50 per task) from the perception pool that were not used for training or evaluation.
  • Embodied: the 200-item embodiment half of RBench, which shares no image or prompt with the 5-task test half used in our paper. It is scored with the official RBench VQA rubric.

Each cell is the validation score. In each row the best ratio is in bold and the second best is marked with *.

Validation setBackbone1:93:75:57:39:1
SimulationWan2.266.467.666.9*64.762.4
SimulationLTX-2.366.666.5*65.659.157.9
PerceptionWan2.277.3*77.676.769.751.1
PerceptionLTX-2.367.263.8*55.346.838.8
EmbodiedWan2.274.577.175.677.0*76.8
EmbodiedLTX-2.368.470.569.6*68.061.4
Average—70.1*70.568.364.258.1

3:7 is the best ratio in four of the six settings and the second best in the other two, so it is clearly the strongest choice overall. It also has the highest average across the six settings (70.5). We will report the validation protocol and this table in the revision.

The same ratio holds on the test sets of both backbones. Test-set sweep from the reasoning-only endpoint (10:0) to the in-domain endpoint (0:10):

BackboneDomain10:07:35:53:70:10
Wan2.2Perception46.272.977.578.978.7
Wan2.2Embodied49.148.752.152.348.8
Wan2.2Simulation68.669.770.069.769.4
LTX-2.3Perception36.550.758.069.168.7
LTX-2.3Embodied39.344.342.648.840.9
LTX-2.3Simulation66.069.069.870.570.4
Average—51.059.261.764.962.8

3:7 is the best ratio in five of six columns and within 0.3 of the best in the sixth. 3:7 is above both endpoints in every column, on both backbones. On average, 3:7 reaches 64.9, against 62.8 for in-domain fine-tuning alone. The LTX-2.3 5:5 and 7:3 models were run during the rebuttal. Table 3 of the paper further shows that this plain ratio blend surpasses or matches more complex merging methods (LoRA-LEGO and RobustMerge) and clearly beats joint training on the mixed data. A fixed-ratio blend therefore works across tasks and backbones, which makes it both a meaningful finding and a practical method.

Due to the character limit, G3–G4 continue in Part 2 below.

AuthorsOpenReview ↗

General Response (2/2)

(Continued from Part 1.)

G3. Uncertainty estimates for ReasonGraft

We add two kinds of uncertainty. (1) Generation seeds: the 3:7 models are regenerated on the full test sets with two more sampling seeds, giving mean ± s.d. over 3 seeds. (2) Item sampling: a paired bootstrap over test items (10,000 resamples, stratified by task) of the difference between ReasonGraft and the in-domain specialist.

BackboneDomainReasonGraft 3:7, seeds 0 / 1 / 2Mean ± s.d.In-domain FTΔ vs. in-domain, paired 95% CI (seed 0)
Wan2.2Perception78.9 / 78.2 / 78.278.4 ± 0.478.7+0.2 [−0.5, +1.0]
Wan2.2Embodied52.3 / 49.5 / 51.251.0 ± 1.448.8+3.5 [−0.5, +7.6]
Wan2.2Simulation69.7 / 69.2 / 69.169.4 ± 0.369.4—
LTX-2.3Perception69.1 / 68.5 / 69.068.9 ± 0.368.7+0.6 [−0.2, +1.1]
LTX-2.3Embodied48.8 / 44.8 / 43.445.7 ± 2.840.9+7.9 [+4.2, +11.6]
LTX-2.3Simulation70.5 / 70.6 / 69.970.3 ± 0.470.4—

Averaged over three seeds, ReasonGraft is on par with or above the in-domain specialist on perception and simulation. It is clearly above the specialist on embodied manipulation: +2.2 on Wan2.2 and +4.8 on LTX-2.3, higher for every seed, and on LTX-2.3 the paired 95% CI excludes zero. ReasonGraft therefore never falls meaningfully below a dedicated specialist and gains where the reasoning component matters most. It gets there without any training: the reasoning adapter is trained once, reused for every target domain, and blended at one fixed ratio that is not tuned per task or per backbone. (The paired CI for simulation is not reported because per-instance Simulation FT scores were not saved.)

G4. World simulation: the importance of calibration and improvements beyond camera motion

Near-static videos hack the original WorldScore benchmark. We built videos that contain no world simulation at all, and scored them with the full WorldScore pipeline on 200 test instances. freeze repeats the input image; jitter adds sub-pixel shake and mild noise to it.

VideoCameraObjectContentSubjective3DPhotometricStyleMeanCalibrated
freeze (input image repeated)19.855.337.564.896.998.3100.067.557.2
jitter (input + small shake and noise)19.843.737.265.494.297.697.165.054.6
Wan2.2 baseline34.739.755.860.580.787.586.663.665.2
Wan2.2 Reasoning FT37.041.656.559.780.488.487.664.568.3

Under the absolute protocol, a frozen image outscores every real model (67.5 vs. 64.5 for Reasoning FT), because three of the seven aspects (3D, photometric, style consistency) are frame-to-frame stability measures that a static clip maximizes. Calibration ranks both static videos last. The same mechanism explains the Perception FT case raised by Reviewer NUq9. Perception FT produces nearly static clips, so its absolute score rises while its calibrated score falls. In the revision, we will use the calibrated score as the single simulation score and state the necessity.

The gain is not only camera control. Splitting the calibrated score into its non-camera parts (Wan2.2, Table 12):

Wan2.2Camera scoreAbility without cameraConsistency
Baseline32.553.683.6
Reasoning FT36.853.984.8

Both non-camera groups are above the baseline (e.g., consistency 84.8 vs. 83.6), so improved camera control is not the only change.

We also recomputed the simulation score under the original absolute mean, and under calibrated variants with ability weight w ∈ {0.5, 0.6, 0.7, 0.8} and logistic temperature τ ∈ {1.5, 2.5, 5, 10} (16 settings per backbone). Reasoning FT > baseline holds under the absolute mean and in all 32 calibrated settings. The weights (w = 0.7, τ = 2.5) were fixed before any fine-tuned model was evaluated.

AuthorsOpenReview ↗

Summary of paper revisions

We thank all reviewers again for their constructive feedback. We have uploaded a revised version of the paper. All new or changed text is shown in blue, and the main text stays within the 10-page limit. The main changes are:

  1. More details about the calibration of WorldScore (Sec. 2.1, Appendix E.6). Sec. 2.1 now explains why calibration is needed. Under WorldScore's uniform mean of seven aspects, a clip that merely repeats the input image, or adds slight shake to it, outscores every real model. This happens because three of the aspects measure frame-to-frame stability. The new Appendix E.6 reports these static and jittered clips under both aggregations. All simulation results in the paper use the calibrated score.

  2. Task composition is reframed and extended (Sec. 5.4, Appendix G.1). Instruction following is known to help video generation models complete tasks. Sec. 5.4 now states explicitly that the experiment removes it from the reasoning data to test whether the remaining reasoning tasks still transfer.

    • The reasoning-only subset improves all three application domains over the baseline.
    • On the out-of-domain split of VBVR-Bench, each subset mainly improves its own task type (Fig. 10b), which indicates that the two components train largely independent capabilities.
    • We also show that the transfer is not specific to VBVR. A mixture of VBVR with five other video-reasoning sources still improves all three domains, and so does a LoRA trained only on four non-VBVR sources (VideoRLVR, PhyWorld, OpenCoF, VR-Bench). The new Appendix G.1 gives the data sources and full results.
  3. The 3:7 ratio of ReasonGraft is selected on held-out validation data (Sec. 7.3, Appendix H). We built a validation split for each application domain, disjoint from all test sets, and swept the ratio on both backbones. 3:7 is the best ratio in four of the six backbone–domain settings and the second best in the other two. Appendix H further reports the test-set sweep on both backbones: 3:7 is the best ratio or matches the best in every domain, which makes the ratio a valuable finding and method.

  4. A single merged model serves all three domains (Sec. 7.3, Table 3, Appendix H). We merge the reasoning LoRA with all three in-domain LoRAs into one adapter. This single model improves every application domain over the baseline and matches the three separate in-domain specialists on average (65.7 vs. 65.6). Plain linear task arithmetic and RobustMerge give nearly the same result. Both merged models are now in Table 3.

We hope these revisions address the reviewers' concerns, and we are happy to answer any further questions.

AuthorsOpenReview ↗

Summary of rebuttal and changes for AC

Dear Area Chair,

We thank the reviewers for their constructive reviews. We are glad that they found the research question meaningful, the cross-domain study systematic and convincing, and the diagnostic analyses informative. We believe most concerns about our paper stemmed from misunderstandings of specific designs in our paper. We addressed all of the concerns with clarifications and new experiments in the rebuttal, summarized below.

  1. Reasoning vs. instruction following (jUUN, NUq9, eYkT). Instruction following is known to benefit video generation. Our task-composition experiment (Sec. 5.4) was therefore designed to remove instruction following from the reasoning data and test whether what remains still generalizes. It does: trained without any instruction-following task, Wan2.2 still improves perception by 28.3% and embodied manipulation by 16.5%. On the VBVR-Bench OOD task types, the reasoning subset improves more on unseen reasoning tasks, while the instruction subset gains more on unseen instruction tasks. Reasoning is thus a largely separate ability, not a by-product of instruction following. Following NUq9’s suggestions, we also ran two new controls. (i) Rewriting the prompts so that targets must be inferred keeps transfer at the same level; the three-domain average even rises from 54.6 to 55.4. (ii) Training on reasoning data with no VBVR at all still improves perception by +6.0 and embodied manipulation by +8.5.

  2. Ratio selection and uncertainty (jUUN, NUq9, eYkT). On held-out validation splits disjoint from the test sets, 3:7 is the best ratio in four of six backbone × domain settings, and the best on average. Over three generation seeds, ReasonGraft is on par with or above the in-domain specialist.

  3. WorldScore calibration (jUUN, NUq9). We showed that a frozen copy of the input image outscores every real model under the uncalibrated mean, which is why calibration is necessary; this also explains the Perception FT case. Reasoning FT beats the baseline under the uncalibrated mean and in all 32 alternative calibration settings, and it also improves the non-camera aspects.

  4. ReasonGraft as a method (NUq9). A single model merging the reasoning adapter with all three in-domain adapters matches the three separate specialists on average (65.7 vs. 65.6).

  5. Missing baselines (NUq9). We evaluated the public VBVR-Wan2.2 checkpoint. It also transfers to all three application domains, but less than our 50K model, and we explained the gap to its published numbers. On RBench, ReasonGraft (52.3) is above Sora 2, Veo 3.1, Kling 2.6, and Seedance 1.5 Pro (51.9).

  6. Additional analyses (eYkT). ReasonGraft works equally well with the instruction-following-only adapter. The gains persist on motion-matched subsets, so they are not explained by reduced dynamics. We also provided the ID/OOD task mapping and the checkpoint-selection details.

We believe these responses address all, or at least most, of the concerns given by the reviewers, including every experiment requested. In the revision, we (i) explained why the calibrated WorldScore is necessary and reported the static and jittered clips under both aggregations; (ii) reframed the task-composition experiment as removing instruction following from the reasoning data, and added the non-VBVR controls; (iii) added the held-out validation protocol for the 3:7 ratio and the test-set sweeps on both backbones; and (iv) added the single merged model serving all three domains to Table 3.

As the discussion period closes soon, we will greatly appreciate it if you encourage the reviewers to take a look at our responses and engage in further discussion.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.