- 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
- A valuable, systematic transfer study. The complete 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.
- 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
-
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.
-
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
-
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?
-
Was the blending ratio selected using separate validation data? Can paired confidence intervals and additional seeds establish the small merging gains?
-
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
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):
| Domain | Backbone | Items | Gain, all items | Gain, matched ×1.25 | Gain, matched ×1.5 |
|---|---|---|---|---|---|
| Perception | Wan2.2 | 300 | +13.7 | +7.3 (n=30) | +8.0 (n=66) |
| Perception | LTX-2.3 | 298 | +7.7 | +8.3 (n=19) | +6.4 (n=43) |
| Embodied | Wan2.2 | 250 | +8.4 | +5.6 (n=47) | +7.1 (n=94) |
| Embodied | LTX-2.3 | 250 | +4.1 | +2.3 (n=77) | +0.2 (n=123) |
| Simulation | Wan2.2 | 200 | +3.1 | +0.2 (n=100) | +1.5 (n=142) |
| Simulation | LTX-2.3 | 200 | +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 set | Backbone | 1:9 | 3:7 | 5:5 | 7:3 | 9:1 |
|---|---|---|---|---|---|---|
| Simulation | Wan2.2 | 66.4 | 67.6 | 66.9* | 64.7 | 62.4 |
| Simulation | LTX-2.3 | 66.6 | 66.5* | 65.6 | 59.1 | 57.9 |
| Perception | Wan2.2 | 77.3* | 77.6 | 76.7 | 69.7 | 51.1 |
| Perception | LTX-2.3 | 67.2 | 63.8* | 55.3 | 46.8 | 38.8 |
| Embodied | Wan2.2 | 74.5 | 77.1 | 75.6 | 77.0* | 76.8 |
| Embodied | LTX-2.3 | 68.4 | 70.5 | 69.6* | 68.0 | 61.4 |
| Average | — | 70.1* | 70.5 | 68.3 | 64.2 | 58.1 |
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):
| Backbone | Domain | 10:0 | 7:3 | 5:5 | 3:7 | 0:10 |
|---|---|---|---|---|---|---|
| Wan2.2 | Perception | 46.2 | 72.9 | 77.5 | 78.9 | 78.7 |
| Wan2.2 | Embodied | 49.1 | 48.7 | 52.1 | 52.3 | 48.8 |
| Wan2.2 | Simulation | 68.6 | 69.7 | 70.0 | 69.7 | 69.4 |
| LTX-2.3 | Perception | 36.5 | 50.7 | 58.0 | 69.1 | 68.7 |
| LTX-2.3 | Embodied | 39.3 | 44.3 | 42.6 | 48.8 | 40.9 |
| LTX-2.3 | Simulation | 66.0 | 69.0 | 69.8 | 70.5 | 70.4 |
| Average | — | 51.0 | 59.2 | 61.7 | 64.9 | 62.8 |
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.
| Backbone | Domain | ReasonGraft 3:7, seeds 0 / 1 / 2 | Mean ± s.d. | In-domain FT | Δ vs. in-domain, paired 95% CI (seed 0) |
|---|---|---|---|---|---|
| Wan2.2 | Perception | 78.9 / 78.2 / 78.2 | 78.4 ± 0.4 | 78.7 | +0.2 [−0.5, +1.0] |
| Wan2.2 | Embodied | 52.3 / 49.5 / 51.2 | 51.0 ± 1.4 | 48.8 | +3.5 [−0.5, +7.6] |
| Wan2.2 | Simulation | 69.7 / 69.2 / 69.1 | 69.4 ± 0.3 | 69.4 | — |
| LTX-2.3 | Perception | 69.1 / 68.5 / 69.0 | 68.9 ± 0.3 | 68.7 | +0.6 [−0.2, +1.1] |
| LTX-2.3 | Embodied | 48.8 / 44.8 / 43.4 | 45.7 ± 2.8 | 40.9 | +7.9 [+4.2, +11.6] |
| LTX-2.3 | Simulation | 70.5 / 70.6 / 69.9 | 70.3 ± 0.4 | 70.4 | — |
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.
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.
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):
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:
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:
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.
ReasonGraft also works with the instruction-following-only adapter.
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.