acceptodds
Under review as a conference paper at ICLR 2027

Getting Space Right: Agentic Rewards for Image Generation

11 views

Abstract

Reinforcement learning (RL) is promising for spatial text-to-image generation, but existing work largely focuses on policy optimization, assuming accurate rewards and effective online rollouts. We observe in-domain gains with out-of-domain degradation, exposing two limitations: unreliable spatial rewards in open-world settings and rollouts lacking basic-to-complex progression. We propose a curriculum-guided RL framework. First, we introduce SpatialCritic, an agentic reward model equipped with a spatial perception harness that combines reusable verification guidance with specialized vision tools. We then construct spatial preference datasets through targeted counterfactual perturbations and perform agent–harness alignment to improve tool selection and evidence integration efficiently. Second, curriculum learning organizes shared foundational capabilities and their compositional dependencies in a capability graph for mastery-guided rollouts, periodic review, and targeted remediation. On SpaRW-Eval, SpatialCritic achieves 95.1% verification accuracy, outperforming 320B GLM-5.3-Flash with a smaller 8B backbone and raising average baseline scores from 78.4 to 83.9 on SD3.5-Medium and 83.2 to 86.5 on BAGEL.

Primary area:
generative models
Keywords:
Text-to-Image Generation, Spatial Intelligence, Agentic Reward Modeling, Curriculum Learning

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_1wPnOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies spatial text-to-image post-training through two components: SpatialCritic, a tool-augmented reward agent trained with outcome supervision, and a capability-graph curriculum incorporating mastery assessment, review, and remediation. Experiments evaluate preference judgment, candidate selection, and downstream reinforcement learning on SD3.5-Medium and BAGEL.

The approach contains identifiable methodological additions and reports improvements. However, the experiments establish the benefits of the complete system more clearly than the necessity or cost-effectiveness of its agentic reward component.

Critical strengths
  • The paper addresses a relevant problem: inaccurate spatial feedback and poorly organized training can limit generation quality beyond the training distribution.

  • The agent ablations distinguish tool-free judgment, fixed tool execution, and dynamic planning, both before and after RL. These provide direct evidence that adaptive evidence acquisition improves accuracy on SpaRW-Eval.

  • The curriculum ablation separately examines prompt data, graph-guided progression, and review/remediation. Evaluation on two generation backbones also provides useful evidence of applicability.

Critical weaknesses
  • The substantive advance over closely related reward models needs sharper delineation. SpatialReward already combines prompt decomposition, expert perception, and grounded VLM reasoning. The main addition here is learning adaptive evidence acquisition and integration. SpatialScore already uses spatially perturbed preference pairs and a specialized reward for online RL, so counterfactual preference construction is not itself a distinct contribution. The fixed-routine ablation supports dynamic planning, but does not substitute for a matched comparison with SpatialReward.

  • The downstream contribution of the new reward is not adequately isolated from the curriculum. Table 3 includes SpatialReward and SpatialScore on SD3.5-Medium, but does not combine them with the same curriculum used by the proposed system. Consequently, the results do not establish whether SpatialCritic is necessary for the final gains, or whether an existing spatial reward with the same prompt pool and scheduler would perform similarly. Differences in reward normalization or rank conversion should also be controlled.

  • The additional complexity is not accompanied by sufficient cost–benefit evidence. Without the curriculum, SpatialCritic achieves 81.9 versus 80.8 for SpatialScore on SD3.5-Medium. The complete system reaches 83.9, compared with 82.3 for DiffusionNFT with GenEval-2. These are positive numerical gains, but their practical value is difficult to assess against data construction, agent RL, tool inference, and curriculum assessment costs. Parameter counts, response-length reductions, and example-level latency do not establish end-to-end efficiency. The paper therefore leaves the return on this additional investment unresolved.

  • Some comparisons and uncertainty estimates are missing. SpatialReward is absent from the main preference-accuracy table; both spatial reward baselines are absent from the candidate-selection table and BAGEL results. Independent training runs or uncertainty estimates are not reported, limiting interpretation of smaller downstream differences. SpaRW-Eval also heavily weights attribute and existence examples, so its overall accuracy alone does not establish broad geometric verification performance.

Questions for authors
  • How do SpatialReward and SpatialScore perform with the identical curriculum, prompt pool, optimizer, generation budget, and reward transformation?

  • Does agent–harness alignment improve downstream generation relative to the unaligned SpatialCritic, beyond improving preference accuracy?

  • What are the total reward-construction costs and average online evaluation costs, including tool calls? How do methods compare under equal GPU-hour or wall-clock budgets?

  • Can you provide repeated-run results and an independent human evaluation of the key downstream comparisons?

8 comments
AuthorsOpenReview ↗

Response to W1

[W1] [W1-1] The substantive advance over closely related reward models needs sharper delineation. SpatialReward already combines prompt decomposition, expert perception, and grounded VLM reasoning. The main addition here is learning adaptive evidence acquisition and integration. SpatialScore already uses spatially perturbed preference pairs and a specialized reward for online RL, so counterfactual preference construction is not itself a distinct contribution. [W1-2] The fixed-routine ablation supports dynamic planning, but does not substitute for a matched comparison with SpatialReward.

[Q2] Does agent–harness alignment improve downstream generation relative to the unaligned SpatialCritic, beyond improving preference accuracy?

Response. We thank the reviewer for this constructive comment on our contributions. We answer [W1-1] by clarifying the methodological differences, [W1-2] with matched comparisons, and [Q2] with a downstream ablation of agent–harness alignment.

Revision. In the main manuscript, we expanded Sec. 2 to clarify the differences from SpatialReward and SpatialScore. In the App., we added the downstream AHA ablation in App. A.3 and Tab. S3.


[W1-1] The substantive advance over closely related reward models needs sharper delineation. SpatialReward already combines prompt decomposition, expert perception, and grounded VLM reasoning. The main addition here is learning adaptive evidence acquisition and integration. SpatialScore already uses spatially perturbed preference pairs and a specialized reward for online RL, so counterfactual preference construction is not itself a distinct contribution.

[A1-1] In the revised Sec. 2, we clarify how SpatialCritic differs from SpatialReward and SpatialScore. SpatialCritic decides which evidence to acquire and how to integrate it, and agent–harness alignment trains this decision process with verifiable outcome rewards. We compare the methods below.

(1) Comparison with SpatialReward. SpatialReward [1] grounds decomposed spatial constraints with expert detectors and reasons over the resulting observations. Its verification procedure is determined prior to observation, following a pre-defined routine per task category [4]. SpatialReward uses category-specific verification routines, whereas our agent selects and composes tools according to the claims and uncertainties in each input. Under matched RL post-training without CL, SpatialCritic reaches an average of 81.9 versus 77.4 for SpatialReward in Tab. 3. (2) Comparison with SpatialScore. SpatialScore [2] trains a reward model on constructed spatial preference pairs, but it is essentially a regression-based reward model. It therefore inherits the main limitation of such models, as a learned score fits the training distribution and generalizes poorly to data outside it. SpatialScore learns preference scores from spatial pairs, whereas we use such pairs [2, 3] as a verifiable environment for learning tool selection and evidence integration. On the held-out 3D-Pose category, SpatialScore obtains only 61.7% versus 95.6% for SpatialCritic in Tab. 1.

[W1-2] The fixed-routine ablation supports dynamic planning, but does not substitute for a matched comparison with SpatialReward.

[A1-2] We add a matched comparison with SpatialReward in our response to W2 & Q1. SpatialReward, SpatialScore, and SpatialCritic are trained on SD3.5-Medium under the same curriculum, prompt pool, RL method, and rollout budget. SpatialCritic reaches an average of 83.9, compared with 79.8 for SpatialReward and 76.8 for SpatialScore, as shown in Tab. R2.

AuthorsOpenReview ↗

Response to Q2

[Q2] Does agent–harness alignment improve downstream generation relative to the unaligned SpatialCritic, beyond improving preference accuracy?

[A2] We train SD3.5-Medium under the same curriculum with SpatialCritic before and after AHA, as reported in App. A.3 and listed in Tab. R1. In the reward evaluation, AHA raises SpaRW-Eval accuracy from 92.9% to 95.1% in Tab. 1 and shortens responses by 38.3% in Fig. S2.

Table R1. Downstream effect of agent–harness alignment under the identical curriculum on SD3.5-Medium, corresponding to Tab. S3 in the revised manuscript.

MethodRewardBRARRR2D3DBasicSpatialIRPSHPSAvg
w/ Flow-GRPO with CLSpatialCritic w/o AHA84.182.973.488.082.793.569.10.8880.8500.30582.0
w/ Flow-GRPO with CLSpatialCritic w/ AHA86.283.974.193.385.395.069.50.9080.8570.30583.9

AHA raises the average from 82.0 to 83.9 and improves all three benchmarks, raising TIIF-Bench RR from 73.4 to 74.1, UniGenBench++ 2D from 88.0 to 93.3, and SpatialGenEval Spatial from 69.1 to 69.5. AHA therefore improves both preference accuracy and downstream generation.


We thank the reviewer for this comment.

References

[1] SpatialReward: Verifiable Spatial Reward Modeling for Fine-Grained Spatial Consistency in Text-to-Image Generation. CVPR 2026.

[2] Enhancing Spatial Understanding in Image Generation via Reward Modeling. CVPR 2026.

[3] CREPE: Can Vision-Language Foundation Models Reason Compositionally? CVPR 2023.

[4] SpatialReward: Official Implementation. Code repository, 2026.

AuthorsOpenReview ↗

Response to W2 & Q1

[W2] The downstream contribution of the new reward is not adequately isolated from the curriculum. Table 3 includes SpatialReward and SpatialScore on SD3.5-Medium, but does not combine them with the same curriculum used by the proposed system. Consequently, the results do not establish whether SpatialCritic is necessary for the final gains, or whether an existing spatial reward with the same prompt pool and scheduler would perform similarly. Differences in reward normalization or rank conversion should also be controlled.

[Q1] How do SpatialReward and SpatialScore perform with the identical curriculum, prompt pool, optimizer, generation budget, and reward transformation?

Response. We thank the reviewer for this insightful suggestion. We answer [W2] and [Q1] together with a matched-curriculum comparison in [A1].

Revision. In the App., we added the matched-curriculum comparison with SpatialReward and SpatialScore in App. A.3 and Tab. S3.


[Q1] How do SpatialReward and SpatialScore perform with the identical curriculum, prompt pool, optimizer, generation budget, and reward transformation?

[A1] As discussed in our response to W1, SpatialScore fits preference scores to its training pairs and SpatialReward follows category-specific routines, and thus neither generalizes beyond the spatial relations and categories it was built for.

We train SD3.5-Medium with SpatialReward and SpatialScore under the same Full CL scheduler, prompt pool PCL\mathcal{P}_{\mathrm{CL}}, mastery probe, RL method Flow-GRPO, and rollout budget used for SpatialCritic. The two new rows with CL are reported in App. A.3 and listed in Tab. R2.

Table R2. Spatial rewards without and with the identical curriculum on SD3.5-Medium, corresponding to Tabs. 3 and S3 in the revised manuscript.

MethodRewardBRARRR2D3DBasicSpatialIRPSHPSAvg
w/ Flow-GRPOSpatialReward81.380.169.676.277.392.665.00.4510.8230.25977.4
w/ Flow-GRPOSpatialScore83.480.969.492.178.793.767.30.8850.8570.30680.8
w/ Flow-GRPOSpatialCritic83.380.369.893.385.394.467.00.8570.8550.30481.9
w/ Flow-GRPO with CLSpatialReward82.778.668.888.281.393.266.10.8620.8540.30379.8
w/ Flow-GRPO with CLSpatialScore83.275.871.081.468.092.665.80.4360.8200.24376.8
w/ Flow-GRPO with CLSpatialCritic86.283.974.193.385.395.069.50.9080.8570.30583.9

As shown in Tab. R2, SpatialCritic achieves 83.9 under the identical curriculum, outperforming SpatialReward at 79.8 and SpatialScore at 76.8. We attribute the gap to the limitations discussed in our response to W1. SpatialScore does not generalize to the curriculum prompts and drops from 80.8 to 76.8, while the category-specific routines of SpatialReward limit its gain to 77.4 to 79.8. The benefit of CL therefore depends on how well the reward model generalizes, and CL helps most when paired with a reward that generalizes well, such as SpatialCritic, which improves from 81.9 to 83.9.


We thank the reviewer for requesting this controlled ablation.

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

The paper improves spatial text-to-image generation through an agentic reward model and capability-guided curriculum learning. SpatialCritic combines prompt decomposition, specialized vision tools, and reinforcement learning for preference verification, while the curriculum schedules training through capability dependencies, mastery assessment, and targeted review. Experiments evaluate preference accuracy, image selection, and generator post-training on SD3.5-Medium and BAGEL.

Critical strengths
  1. The agent ablations support adaptive planning beyond simply providing tool access. Under matched controls, dynamic planning reaches 95.1% accuracy after reinforcement learning, compared with 92.9% for fixed routines and 88.3% without tools. This provides concrete evidence for the agentic design.

  2. The improvements translate into useful generation outcomes. SpatialCritic improves Best-of-10 selection, and the complete framework benefits two generators across multiple benchmarks. Table S1 also separates prompt curation, curriculum progression, and review, supporting a scheduling contribution beyond changing the training data.

Critical weaknesses
  1. The headline verification comparison is sensitive to benchmark composition. Attribute and existence judgments constitute 70.9% of SpaRW-Eval. An equal-category average calculated from Table 1 gives approximately 90.5% for aligned SpatialCritic versus 91.0% for GLM-5.3-Flash, reversing their weighted ranking. Small geometric subsets and missing uncertainty estimates limit conclusions about broadly superior spatial verification.

  2. Important curriculum details remain unspecified. The construction of the curriculum prompt pool, the implementation of the absolute mastery probe, and the complete generator-training budget are insufficiently described. These details are necessary to reproduce the competence-driven progression and verify the matched-budget comparisons.

Questions for authors
  1. Do the verification conclusions persist on a category-balanced evaluation with confidence intervals, particularly for the small geometric subsets?
  2. Can you specify the curriculum prompt pool, the absolute mastery probe, and the total rollout budget, including how review and remediation consume that budget?
4 comments
AuthorsOpenReview ↗

Response to W1 & Q1 (1/2)

[W1] The headline verification comparison is sensitive to benchmark composition. Attribute and existence judgments constitute 70.9% of SpaRW-Eval. [W1-1] An equal-category average calculated from Table 1 gives approximately 90.5% for aligned SpatialCritic versus 91.0% for GLM-5.3-Flash, reversing their weighted ranking. [W1-2] Small geometric subsets and missing uncertainty estimates limit conclusions about broadly superior spatial verification.

[Q1] Do the verification conclusions persist on a category-balanced evaluation with confidence intervals, particularly for the small geometric subsets?

Response. We thank the reviewer for this careful analysis of the equal-category ranking. We answer [W1-1] on the equal-category ranking in [A1], and [W1-2] together with [Q1] on small geometric subsets and uncertainty in [A2].

Revision. In the main manuscript, we revised the abstract and Sec. 5.2 to state that SpatialCritic is comparable to GLM-5.3-Flash, and added equal-category Macro accuracy and 95% confidence intervals in Tab. 1. In the App., we added the five-run repeated evaluation in App. A.4 and Tab. S4, the real-photograph geometric evaluation in App. A.5 and Tab. S5, and the end-to-end cost comparison in App. A.6 and Tab. S6.


[W1-1] An equal-category average calculated from Table 1 gives approximately 90.5% for aligned SpatialCritic versus 91.0% for GLM-5.3-Flash, reversing their weighted ranking.

[A1] We agree that SpatialCritic is comparable to GLM-5.3-Flash rather than broadly superior, and we have revised the abstract and Sec. 5.2 accordingly. This conclusion rests on two results. (1) In the single reported run of Tab. R1, SpatialCritic is higher in Overall accuracy at 95.1% versus 94.5% and lower in Macro accuracy at 90.5% versus 91.0%, with overlapping 95% confidence intervals. (2) Over five repeated runs in Tabs. R2 and R3, SpatialCritic is slightly ahead on both metrics, at 95.2% versus 94.6% Overall and 91.5% versus 90.5% Macro, indicating that the small Macro gap in a single run comes from random differences between runs. Tab. R1 shows the updated Tab. 1, where Overall weights categories by their number of pairs and Macro averages the 13 categories equally.

Table R1. Accuracy on SpaRW-Eval with 95% confidence intervals, corresponding to Tab. 1 in the revised manuscript.

ModelParamsObj-AttrObj-ExistCountBgComp3D-PoseDepthOrientPos-2DTopoLayoutOcclDistOverallMacro
Aesthetic predictor0.4B54.351.446.846.745.047.848.142.451.947.546.755.451.351.6±1.348.9±2.8
CLIP0.4B72.785.884.452.066.776.552.862.050.663.366.763.159.075.3±1.165.8±2.7
PickScore1.0B82.386.485.349.365.079.372.278.350.672.573.361.571.881.1±1.171.4±2.5
HPSv21.0B84.394.090.840.055.075.467.672.846.878.376.052.379.583.1±1.070.2±2.5
UnifiedReward7B85.791.489.961.373.387.163.971.757.084.281.366.289.785.7±0.977.1±2.3
SpatialScore7B89.198.492.798.778.361.776.984.886.195.090.780.092.387.5±0.986.5±1.8
Qwen3-VL-8B8B89.582.385.358.770.091.058.367.448.182.581.367.779.584.9±1.074.0±2.5
DeepSeek-V4.1-Flash†552B92.294.095.470.788.391.070.479.374.791.786.772.376.990.9±0.883.3±2.2
Qwen3-VL-235B235B95.394.693.688.081.795.768.576.167.195.090.770.892.393.3±0.785.3±2.0
Qwen3-Omni-30B35B95.395.595.469.388.395.174.181.568.491.789.380.089.793.5±0.785.7±2.0
Qwen3-VL-30B30B95.395.992.773.383.394.484.378.369.695.885.380.094.993.7±0.786.4±2.0
GLM-5.3-Flash†320B93.996.799.177.393.397.087.092.474.795.894.786.294.994.5±0.691.0±1.6
SpatialCritic8B94.293.690.890.788.392.688.082.679.793.393.381.589.792.9±0.789.1±1.8
SpatialCritic w/ AHA8B96.096.298.284.088.395.690.785.981.095.094.781.589.795.1±0.690.5±1.8
SpatialCritic30B96.696.095.489.395.095.988.090.291.197.590.778.592.395.5±0.692.0±1.6
Doubao-Seed-2.0-Lite-86.092.393.668.080.090.665.776.168.491.788.061.576.987.2±0.979.9±2.3
Gemini 3.5 Flash-90.393.192.770.765.088.356.576.163.390.081.352.376.988.4±0.976.7±2.4
Qwen3.8-Flash-Next†-97.097.999.184.081.796.676.994.679.796.797.376.992.395.9±0.690.1±1.8
Grok 4.6†-96.698.398.288.091.795.493.591.392.495.894.784.697.496.3±0.593.7±1.4
Doubao-Seed-2.0-Pro-96.998.597.293.396.796.684.395.783.595.897.383.1100.096.6±0.593.8±1.3
Gemini 3.8 Flash†-96.498.998.286.795.097.491.787.097.595.094.787.794.996.7±0.593.9±1.4
AuthorsOpenReview ↗

Response to W1 & Q1 (2/2)

To check whether the single-run gap is due to chance, we rerun both judges five times with the same prompts and settings in App. A.4 and list the results in Tabs. R2 and R3.

Table R2. Five repeated runs of GLM-5.3-Flash on SpaRW-Eval, corresponding to the upper block of Tab. S4 in the revised manuscript.

SettingObj-AttrObj-ExistCountBgComp3D-PoseDepthOrientPos-2DTopoLayoutOcclDistOverallMacro
Reported93.996.799.177.393.397.087.092.474.795.894.786.294.994.591.0
Repeat 194.096.899.176.091.797.486.192.476.095.894.781.594.994.590.5
Repeat 294.097.299.173.393.397.686.192.474.795.894.783.194.994.790.5
Repeat 394.296.899.176.090.097.285.291.376.095.893.383.194.994.690.2
Repeat 494.196.999.174.795.097.486.192.478.594.293.381.594.994.690.6
Repeat 594.596.899.177.391.797.486.192.474.795.894.784.694.994.890.8
μ±σ\mu^{\pm\sigma}94.2±0.296.9±0.299.1±0.075.5±1.592.3±1.997.4±0.185.9±0.492.2±0.575.9±1.695.5±0.794.1±0.782.8±1.394.9±0.094.6±0.190.5±0.2
μ±CI95\mu^{\pm\text{CI}_{95}}94.2±0.296.9±0.299.1±0.075.5±1.992.3±2.497.4±0.285.9±0.592.2±0.675.9±1.995.5±0.994.1±0.982.8±1.694.9±0.094.6±0.190.5±0.2

Table R3. Five repeated runs of 8B SpatialCritic w/ AHA on SpaRW-Eval, corresponding to the lower block of Tab. S4 in the revised manuscript.

SettingObj-AttrObj-ExistCountBgComp3D-PoseDepthOrientPos-2DTopoLayoutOcclDistOverallMacro
Reported96.096.298.284.088.395.690.785.981.095.094.781.589.795.190.5
Repeat 196.096.698.286.788.395.191.785.982.393.396.089.289.795.291.5
Repeat 296.395.996.389.390.094.991.784.881.095.093.389.292.395.291.5
Repeat 396.196.596.389.390.094.690.785.983.594.292.087.789.795.291.3
Repeat 495.996.097.289.390.095.493.587.083.594.296.087.789.795.292.0
Repeat 596.296.596.392.088.394.293.585.979.894.293.384.689.795.191.1
μ±σ\mu^{\pm\sigma}96.1±0.296.3±0.396.9±0.889.3±1.989.3±0.994.8±0.492.2±1.285.9±0.882.0±1.694.2±0.694.1±1.887.7±1.990.3±1.195.2±0.091.5±0.3
μ±CI95\mu^{\pm\text{CI}_{95}}96.1±0.296.3±0.496.9±1.089.3±2.389.3±1.194.8±0.592.2±1.585.9±1.082.0±2.094.2±0.794.1±2.287.7±2.390.3±1.495.2±0.191.5±0.4

Across all five runs, SpatialCritic exceeds GLM-5.3-Flash on both Overall and Macro accuracy, and the two judges are comparable overall.

[W1-2] Small geometric subsets and missing uncertainty estimates limit conclusions about broadly superior spatial verification.

[Q1] Do the verification conclusions persist on a category-balanced evaluation with confidence intervals, particularly for the small geometric subsets?

[A2] Yes. The conclusion persists for the small geometric subsets. Ten categories contain 39 to 120 pairs, as listed in Tab. S9, and consequently have wider sampling intervals than the larger categories. Across the five runs in Tabs. R2 and R3, however, the per-category standard deviation ranges from 0.2 to 1.9 for SpatialCritic and from 0.0 to 1.9 for GLM-5.3-Flash. The results are therefore stable overall, and the conclusion that the two judges are comparable does not change across runs.

To test geometric relations beyond SpaRW-Eval, App. A.5 evaluates 820 real-photograph pairs from What'sUp [1] that differ only in left and right, on and under, or front and behind.

Table R4. Accuracy on 820 real-photograph pairs from What'sUp, corresponding to Tab. S5 in the revised manuscript.

ModelParamsWhat'sUp
Qwen3-VL-8B8B64.4
GLM-5.3-Flash320B98.9
SpatialCritic8B98.3

In Tab. R4, SpatialCritic raises accuracy from 64.4% for its tool-free policy Qwen3-VL-8B to 98.3%, on par with 98.9% for the 320B GLM-5.3-Flash. The small remaining gap arises because SpatialCritic uses only generated images in its post-training, whereas GLM-5.3-Flash is pretrained on large-scale real photographs. Reaching accuracy on par with GLM-5.3-Flash thus shows that SpatialCritic transfers to real photographs, while using 5 GPUs versus 16 in Tab. S6.


We thank the reviewer for raising this statistical point.

References

[1] What's "up" with vision-language models? Investigating their struggle with spatial reasoning. EMNLP 2023.

AuthorsOpenReview ↗

Response to W2 & Q2 (1/2)

[W2] Important curriculum details remain unspecified. The construction of the curriculum prompt pool, the implementation of the absolute mastery probe, and the complete generator-training budget are insufficiently described. These details are necessary to reproduce the competence-driven progression and verify the matched-budget comparisons.

[Q2] Can you specify [Q2-1] the curriculum prompt pool, the absolute mastery probe, and [Q2-2] the total rollout budget, including how review and remediation consume that budget?

Response. We thank the reviewer for pointing out these missing details. We added these specifications in Apps. D.3 to D.5. We answer [Q2-1] with the prompt-pool construction and mastery-probe implementation, and [Q2-2] with the rollout accounting.

Revision. We documented the curriculum prompt pool PCL\mathcal{P}_{\mathrm{CL}} in App. D.5 and Tab. S16, the absolute mastery probe Pv(I,c)\mathcal{P}_v(\mathbf{I}, c) and the rollout budget in Apps. D.3 and D.4, the curriculum runtime overhead in App. A.7 and Tabs. S7 and S8, and the end-to-end reward inference cost in App. A.6 and Tab. S6.


[Q2-1] Can you specify the curriculum prompt pool, the absolute mastery probe, […]

[A2-1] We documented the construction of PCL\mathcal{P}_{\mathrm{CL}} in App. D.5, and summarize the pool in Tab. R5.

Table R5. Composition of the curriculum prompt pool, corresponding to Tab. S16 in the revised manuscript.

PoolSourceCapabilitiesPrompts
Single-capabilityValidated reference prompts over 13 capabilities110,027
CompositionalValidated multi-relation prompts2–44,031
Audited LLM-synthesized compositions2–41,164
Compositional subtotal2–45,195
Total PCL\mathcal{P}_{\mathrm{CL}}Combined training pool1–415,222

PCL\mathcal{P}_{\mathrm{CL}} is divided into two pools matching the two stages of our curriculum, with 10,027 single-capability prompts for foundational progression, replay, and remediation, and 5,195 multi-capability prompts for periodic review. We construct the pool in the following steps.

(1) Single-capability prompts. The prompts used to train SpatialCritic in Sec. 3.2 provide 8,307 prompts covering 11 capabilities, and 1,720 further prompts that pass text-level validation cover Object-Existence and 3D-Pose. A prompt that describes several spatial relations is used once for each of them, each time with a single target relation, and duplicate prompts within each capability are removed. Each prompt records its target relation, the objects involved, and the reference frame, and the mastery probe turns this relation into a verification question. Each capability has 739 to 873 prompts, which allows every capability to accumulate the 200 images required for mastery.

(2) Capability combinations. We enumerate all 1,079 combinations of 2 to 4 capabilities and include all 78 pairs, since pairs are the smallest unit of composition. Among the 286 triples and 715 quadruples, we remove combinations whose capabilities share no compatible scene, as well as combinations containing one of five pairs that rarely appear together in a short prompt, unless existing validated prompts already contain them. From the triples and quadruples that appear in existing validated prompts, we greedily select 60 triples and 40 quadruples, preferring combinations with more existing prompts and more shared scenes, capabilities connected in CCG, and capabilities from different parts of CCG, with each capability appearing in at least 8 triples and 5 quadruples. The resulting 78 + 60 + 40 = 178 combinations form the core combinations.

(3) Compositional prompts. We take 4,031 existing prompts that pass text-level validation and describe exactly the relations of their stated capabilities. For core combinations with too few prompts, DeepSeek-V4-Flash writes new prompts of 12 to 28 words with one explicitly stated relation for each capability, and a second DeepSeek-V4-Flash call checks capability coverage, relation clarity, reference frames, physical plausibility, and length. Only prompts that pass all five checks are kept, which adds 1,164 prompts and covers all 178 core combinations.

The mastery probe is described in App. D.4 and its cost in App. A.7. The probe runs on the policy VLM of the reward model, turns the target relation of each prompt into a yes or no question, and reads the probabilities of the two answer tokens to score an image between 0 and 1. It scores 4 images per prompt group and updates a running average of each capability with Eq. S5.

AuthorsOpenReview ↗

Response to W2 & Q2 (2/2)

[Q2-2] Can you specify […] the total rollout budget, including how review and remediation consume that budget?

[A2-2] The curriculum adds 29.3 s, or 0.79%, to a 3,704 s training step of SD3.5-Medium on 8 GPUs, as measured in App. A.7. All compared Flow-GRPO runs share the same rollout budget of 576 images per iteration, review and remediation replace forward steps within this budget, and capability assessment scores the four images already sampled in each rollout. The curriculum therefore adds no rollouts, and its only added cost is the probe time below.

(1) Added operations. After each rollout, the curriculum (1) generates a verification question from the target relation phrase, (2) scores each image with the mastery probe, (3) takes the minimum score over the capabilities of each prompt, (4) gathers the scores across GPUs, and (5) updates the capability states. Image sampling and reward computation are unchanged.

(2) Cost per call. Probe scoring reads the probabilities of the yes and no tokens and generates a single token, and a scoring call thus takes 0.08 s, compared with 8 s for one reward comparison. In each rollout, every GPU generates one question and issues four scoring calls, one per image, which takes 2.78 s in total, with 1.1 s for question generation, 0.1 to 0.3 s for the four scoring calls, and 1.4 s for image encoding and transfer.

Table R6. Cost of one judge call, corresponding to Tab. S7 in the revised manuscript. Input and output are token counts.

CallInputOutputLatency
(1) Question generation603451.13 s
(2) Probe scoring29410.08 s
Reward comparison4,8004008 s

(3) Cost per training step. Tab. R7 lists the time the curriculum adds to one 3,704 s training step. Of the 29.3 s, 12.6 s is GPU synchronization that image sampling already requires and that operation (4) moves earlier in the step. Reward batches take 348 s with concurrent probe calls and 350 s without them, indicating that probe calls do not slow reward computation.

Table R7. Time added by the curriculum to one 3,704 s training step, corresponding to Tab. S8 in the revised manuscript.

ComponentTimeShare of step
(1) to (3) Question generation and scoring12.8 s0.35%
(4) GPU synchronization after scoring3.9 s0.11%
(4) GPU synchronization after sampling12.6 s0.34%
Total29.3 s0.79%

(4) Review and remediation. The hyperparameters of review and remediation are listed in Tab. R8.

Table R8. Curriculum and capability-assessment hyperparameters, corresponding to Tab. S15 in the revised manuscript.

ParameterSetting
Capability assessment
Mastery threshold τ\tau0.60
Minimum assessed images Nmin⁡N_{\min}200
EMA factor η\eta0.10
Mastery revocation criterionEvaluated by Eq. (2) (Sˉv<τ\bar{\mathcal{S}}_v < \tau); no separate margin
Capability development
Fallback assessment limit Nmax⁡N_{\max}240 assessed images
Assessed images per prompt group4
Replay probability for mastered capabilities0.10
Periodic review
Review intervalEvery 3 newly mastered capabilities
Review duration20 rollout iterations (counted within the shared rollout budget)
Single-capability prompt share50%
Compositional prompt share50%
Targeted remediation
Training budget per attempt3 rollout iterations (counted within the shared rollout budget)
Maximum retries per capability2

(5) Reward inference cost. The reward model itself is also cheap to run during training. We compare the deployment cost per pair of SpatialCritic and GLM-5.3-Flash at a concurrency of 1, as reported in App. A.6.

Table R9. End-to-end cost of SpatialCritic and GLM-5.3-Flash at a concurrency of 1, corresponding to Tab. S6 in the revised manuscript.

SpatialCriticGLM-5.3-Flash
Architecture8B dense + 16 toolsMoE
Total parameters8.8B + 6.0B tools320B
Active parameters per token8.1B†18B‡
Weight precisionBF16FP8
Policy weight memory18.2 GB326.7 GB
Tool server memory4 × 13.5 GB-
GPUs for the policy116
GPUs end to end516
Latency per pair30.2 s153.5 s
Policy LLM latency27.5 s-
Tool latency2.6 s-
GPU-seconds per pair1512,456
Accuracy, mean ± std95.2 ± 0.094.6 ± 0.1

† denotes the 8.0B language model of the policy plus 0.1B tool parameters per call, averaged over call frequencies. ‡ denotes 8 routed experts and 1 shared expert out of 288.

As shown in Tab. R9, SpatialCritic achieves accuracy comparable to GLM-5.3-Flash at 95.2% versus 94.6% while reducing latency from 153.5 s to 30.2 s, end-to-end GPUs from 16 to 5, policy weight memory from 326.7 GB to 18.2 GB, and active parameters from 18B to 8.1B.


We thank the reviewer for requesting these specifications.

Reviewer_bb6POpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a curriculum-guided reinforcement learning framework for improving spatial faithfulness in text-to-image generation. Its first component, SpatialCritic, is an agentic reward model that decomposes prompts into verifiable claims, dynamically plans tool use, and integrates specialized vision-tool outputs into spatial judgments. It is trained with spatial preference data and agent–harness alignment using verifiable rewards. The second component organizes training rollouts through a Compositional Capability Graph, using mastery estimation, curriculum progression, review, and remediation. SpatialCritic is evaluated on SpaRW-Eval, which contains 5,572 preference pairs rendered by held-out generators, with non-overlapping training and evaluation prompt pairs. It achieves 95.1% overall verification accuracy, compared with 84.9% for the tool-free backbone. Downstream RL post-training improves average generation scores from 78.4 to 83.9 on SD3.5-Medium and from 83.2 to 86.5 on BAGEL. Ablations examine tool planning, reward alignment, preference data, curriculum construction, and remediation.

Critical strengths
  1. SpatialCritic provides grounded and adaptive spatial verification rather than relying only on a VLM’s ungrounded judgment. The improvement from 84.9% tool-free accuracy to 95.1% with the complete method is substantial. The fixed-routine and dynamic-planning ablations further show that both tool access and the policy for using tools matter. Evaluation with held-out generators and non-overlapping prompt pairs strengthens the evidence that the reward model generalizes beyond memorizing specific training examples.

  2. The paper connects reward-model improvements to downstream generation gains and provides meaningful component ablations. Improvements on both SD3.5-Medium and BAGEL indicate that the approach is not limited to a single generator. The curriculum ablation on SD3.5-Medium—separating preference data, capability-graph construction, and review/remediation—provides useful evidence that the downstream gains are not attributable solely to adding more training data.

Critical weaknesses
  1. The efficiency and scale comparison is incomplete for the deployed reward system. The headline emphasizes an 8B backbone that is approximately 40 times smaller than a 320B comparator, but the appendix lists approximately 6.0B additional tool parameters, giving a total agent-and-tool footprint of about 14.8B parameters. More importantly, the normalized latency illustration is explicitly based on a single example rather than aggregate runtime. Without end-to-end measurements over the benchmark, including tool invocation counts and cost, the practical efficiency advantage remains difficult to assess.
Questions for authors
  1. Could you report macro-averaged or otherwise category-balanced SpaRW-Eval accuracy, with per-category confidence intervals, in addition to the pair-weighted overall score? How does SpatialCritic perform on an external human-preference dataset or naturally generated images outside the proposed synthetic construction pipeline?

  2. Could you provide aggregate end-to-end reward evaluation cost for the complete agent-and-tool stack, including latency, memory, number and type of tool calls, and total active parameters? How does this cost compare with tool-free reward models and the large-model baselines at matched evaluation quality?

4 comments
AuthorsOpenReview ↗

Response to Q1 (1/2)

[Q1] [Q1-1] Could you report macro-averaged or otherwise category-balanced SpaRW-Eval accuracy, with per-category confidence intervals, in addition to the pair-weighted overall score? [Q1-2] How does SpatialCritic perform on an external human-preference dataset or naturally generated images outside the proposed synthetic construction pipeline?

Response. We thank the reviewer for these constructive questions. We answer [Q1-1] with category-balanced accuracy and confidence intervals, and [Q1-2] with an evaluation beyond our synthetic pipeline.

Revision. In the main manuscript, we added equal-category Macro accuracy and 95% confidence intervals in Tab. 1. In the App., we added the five-run repeated evaluation in App. A.4 and Tab. S4, and the external real-photograph evaluation on What'sUp in App. A.5 and Tab. S5.


[Q1-1] Could you report macro-averaged or otherwise category-balanced SpaRW-Eval accuracy, with per-category confidence intervals, in addition to the pair-weighted overall score?

[A1-1] We report equal-category Macro accuracy alongside pair-weighted Overall accuracy with 95% confidence intervals in Tab. 1 of the revised manuscript, and provide five repeated runs with per-category confidence intervals in App. A.4.

First, Tab. R1 shows the updated Tab. 1. Overall weights categories by their number of pairs, Macro averages the 13 categories equally, and the values after ± give 95% confidence intervals.

Table R1. Accuracy on SpaRW-Eval with 95% confidence intervals, corresponding to Tab. 1 in the revised manuscript.

ModelParamsObj-AttrObj-ExistCountBgComp3D-PoseDepthOrientPos-2DTopoLayoutOcclDistOverallMacro
Aesthetic predictor0.4B54.351.446.846.745.047.848.142.451.947.546.755.451.351.6±1.348.9±2.8
CLIP0.4B72.785.884.452.066.776.552.862.050.663.366.763.159.075.3±1.165.8±2.7
PickScore1.0B82.386.485.349.365.079.372.278.350.672.573.361.571.881.1±1.171.4±2.5
HPSv21.0B84.394.090.840.055.075.467.672.846.878.376.052.379.583.1±1.070.2±2.5
UnifiedReward7B85.791.489.961.373.387.163.971.757.084.281.366.289.785.7±0.977.1±2.3
SpatialScore7B89.198.492.798.778.361.776.984.886.195.090.780.092.387.5±0.986.5±1.8
Qwen3-VL-8B8B89.582.385.358.770.091.058.367.448.182.581.367.779.584.9±1.074.0±2.5
DeepSeek-V4.1-Flash†552B92.294.095.470.788.391.070.479.374.791.786.772.376.990.9±0.883.3±2.2
Qwen3-VL-235B235B95.394.693.688.081.795.768.576.167.195.090.770.892.393.3±0.785.3±2.0
Qwen3-Omni-30B35B95.395.595.469.388.395.174.181.568.491.789.380.089.793.5±0.785.7±2.0
Qwen3-VL-30B30B95.395.992.773.383.394.484.378.369.695.885.380.094.993.7±0.786.4±2.0
GLM-5.3-Flash†320B93.996.799.177.393.397.087.092.474.795.894.786.294.994.5±0.691.0±1.6
SpatialCritic8B94.293.690.890.788.392.688.082.679.793.393.381.589.792.9±0.789.1±1.8
SpatialCritic w/ AHA8B96.096.298.284.088.395.690.785.981.095.094.781.589.795.1±0.690.5±1.8
SpatialCritic30B96.696.095.489.395.095.988.090.291.197.590.778.592.395.5±0.692.0±1.6
Doubao-Seed-2.0-Lite-86.092.393.668.080.090.665.776.168.491.788.061.576.987.2±0.979.9±2.3
Gemini 3.5 Flash-90.393.192.770.765.088.356.576.163.390.081.352.376.988.4±0.976.7±2.4
Qwen3.8-Flash-Next†-97.097.999.184.081.796.676.994.679.796.797.376.992.395.9±0.690.1±1.8
Grok 4.6†-96.698.398.288.091.795.493.591.392.495.894.784.697.496.3±0.593.7±1.4
Doubao-Seed-2.0-Pro-96.998.597.293.396.796.684.395.783.595.897.383.1100.096.6±0.593.8±1.3
Gemini 3.8 Flash†-96.498.998.286.795.097.491.787.097.595.094.787.794.996.7±0.593.9±1.4
AuthorsOpenReview ↗

Response to Q1 (2/2)

Second, we rerun GLM-5.3-Flash and 8B SpatialCritic five times with the same prompts and settings in App. A.4, which yields per-category confidence intervals alongside the single-run sampling intervals in Tab. R1. Tabs. R2 and R3 list the results, where Reported is the run in Tab. R1. μ\mu and σ\sigma are the mean and standard deviation over Repeats 1–5, and CI95\text{CI}_{95} is the half-width of the 95% confidence interval of the mean.

Table R2. Five repeated runs of GLM-5.3-Flash on SpaRW-Eval, corresponding to the upper block of Tab. S4 in the revised manuscript.

SettingObj-AttrObj-ExistCountBgComp3D-PoseDepthOrientPos-2DTopoLayoutOcclDistOverallMacro
Reported93.996.799.177.393.397.087.092.474.795.894.786.294.994.591.0
Repeat 194.096.899.176.091.797.486.192.476.095.894.781.594.994.590.5
Repeat 294.097.299.173.393.397.686.192.474.795.894.783.194.994.790.5
Repeat 394.296.899.176.090.097.285.291.376.095.893.383.194.994.690.2
Repeat 494.196.999.174.795.097.486.192.478.594.293.381.594.994.690.6
Repeat 594.596.899.177.391.797.486.192.474.795.894.784.694.994.890.8
μ±σ\mu^{\pm\sigma}94.2±0.296.9±0.299.1±0.075.5±1.592.3±1.997.4±0.185.9±0.492.2±0.575.9±1.695.5±0.794.1±0.782.8±1.394.9±0.094.6±0.190.5±0.2
μ±CI95\mu^{\pm\text{CI}_{95}}94.2±0.296.9±0.299.1±0.075.5±1.992.3±2.497.4±0.285.9±0.592.2±0.675.9±1.995.5±0.994.1±0.982.8±1.694.9±0.094.6±0.190.5±0.2

Table R3. Five repeated runs of 8B SpatialCritic w/ AHA on SpaRW-Eval, corresponding to the lower block of Tab. S4 in the revised manuscript.

SettingObj-AttrObj-ExistCountBgComp3D-PoseDepthOrientPos-2DTopoLayoutOcclDistOverallMacro
Reported96.096.298.284.088.395.690.785.981.095.094.781.589.795.190.5
Repeat 196.096.698.286.788.395.191.785.982.393.396.089.289.795.291.5
Repeat 296.395.996.389.390.094.991.784.881.095.093.389.292.395.291.5
Repeat 396.196.596.389.390.094.690.785.983.594.292.087.789.795.291.3
Repeat 495.996.097.289.390.095.493.587.083.594.296.087.789.795.292.0
Repeat 596.296.596.392.088.394.293.585.979.894.293.384.689.795.191.1
μ±σ\mu^{\pm\sigma}96.1±0.296.3±0.396.9±0.889.3±1.989.3±0.994.8±0.492.2±1.285.9±0.882.0±1.694.2±0.694.1±1.887.7±1.990.3±1.195.2±0.091.5±0.3
μ±CI95\mu^{\pm\text{CI}_{95}}96.1±0.296.3±0.496.9±1.089.3±2.389.3±1.194.8±0.592.2±1.585.9±1.082.0±2.094.2±0.794.1±2.287.7±2.390.3±1.495.2±0.191.5±0.4

Across the five repeats in Tabs. R2 and R3, SpatialCritic obtains 95.2% Overall and 91.5% Macro accuracy, higher than 94.6% and 90.5% for GLM-5.3-Flash, and exceeds GLM-5.3-Flash on both metrics in every repeat. The per-category standard deviation ranges from 0.2 to 1.9 for SpatialCritic and from 0.0 to 1.9 for GLM-5.3-Flash, indicating that the results are stable overall.

[Q1-2] How does SpatialCritic perform on an external human-preference dataset or naturally generated images outside the proposed synthetic construction pipeline?

[A1-2] To evaluate SpatialCritic outside our synthetic pipeline, we construct an external benchmark of 820 real-photograph pairs from What'sUp [1] in App. A.5, where paired photos show the same objects in the same scene and differ only in spatial relation.

Table R4. Accuracy on 820 real-photograph pairs from What'sUp, corresponding to Tab. S5 in the revised manuscript.

ModelParamsWhat'sUp
Qwen3-VL-8B8B64.4
GLM-5.3-Flash320B98.9
SpatialCritic8B98.3

In Tab. R4, SpatialCritic raises accuracy from 64.4% for its tool-free policy model Qwen3-VL-8B to 98.3%, on par with 98.9% for the 320B GLM-5.3-Flash. The small remaining gap arises because SpatialCritic uses only generated images in its post-training, whereas GLM-5.3-Flash is pretrained on large-scale real photographs. Reaching accuracy on par with GLM-5.3-Flash thus shows that SpatialCritic transfers to real photographs, while using 5 GPUs versus 16 in Tab. R8.


We thank the reviewer for these suggestions.

References

[1] What's "up" with vision-language models? Investigating their struggle with spatial reasoning. EMNLP 2023.

AuthorsOpenReview ↗

Response to W1 & Q2 (1/2)

[W1] The efficiency and scale comparison is incomplete for the deployed reward system. The headline emphasizes an 8B backbone that is approximately 40 times smaller than a 320B comparator, but the appendix lists approximately 6.0B additional tool parameters, giving a total agent-and-tool footprint of about 14.8B parameters. More importantly, the normalized latency illustration is explicitly based on a single example rather than aggregate runtime. Without end-to-end measurements over the benchmark, including tool invocation counts and cost, the practical efficiency advantage remains difficult to assess.

[Q2] [Q2-1] Could you provide aggregate end-to-end reward evaluation cost for the complete agent-and-tool stack, including latency, memory, number and type of tool calls, and total active parameters? [Q2-2] How does this cost compare with tool-free reward models and the large-model baselines at matched evaluation quality?

Response. We thank the reviewer for this helpful suggestion on efficiency. We address [W1] and [Q2-1] with benchmark-wide measurements of the complete agent-and-tool stack, and [Q2-2] with comparisons at the reported evaluation quality.

Revision. In the abstract, we now state that SpatialCritic achieves accuracy comparable to 320B GLM-5.3-Flash. In App. A.6, we added the benchmark-wide runtime breakdown in Fig. S3, the tool invocation and memory analysis in Fig. S4, and the end-to-end cost and active-parameter comparison in Tab. S6.


[W1] […] Without end-to-end measurements over the benchmark, including tool invocation counts and cost, the practical efficiency advantage remains difficult to assess.

[Q2-1] Could you provide aggregate end-to-end reward evaluation cost for the complete agent-and-tool stack, including latency, memory, number and type of tool calls, and total active parameters?

[A2-1] We report benchmark-wide runtime breakdowns, tool-call counts, loaded weight memory, and active parameters for the complete agent-and-tool stack in App. A.6, and list them in Tabs. R5 to R8.

(1) We profile the runtime of SpatialCritic over three full runs on SpaRW-Eval and split it by stage and by server in Tab. R5.

Table R5. Runtime per pair of SpatialCritic by stage and by server, corresponding to Fig. S3 in the revised manuscript.

BreakdownComponentTime per Pair, sShare, %
By stagePlanning14.648.5
Verification6.722.4
Final judgment6.220.8
Tools2.48.0
I/O & scheduling0.10.3
By serverPolicy LLM over all 3 stages27.591.6
Tools2.48.0
I/O & scheduling0.10.3

The policy LLM takes 91.6% of the 30.1 s per pair, and tools take 8.0%.

(2) We count the tool calls of one full run and measure the share of calls and tool time for each tool group in Tab. R6.

Table R6. Share of tool calls and tool time by tool group in one full run, corresponding to Fig. S4a and S4b in the revised manuscript.

Tool GroupShare of Calls, %Share of Tool Time, %
Detection39.772.6
Spatial metrics on CPU28.40.2
Attributes13.013.1
Orientation10.80.8
VLM checks5.00.1
Counting2.512.5
Depth0.40.5
3D & scene graph0.10.2
Pose & segmentation0.00.0

The agent calls tools on 67.4% of pairs at 1.41 planned calls per pair, which the tool server expands to 3.16 executed calls, and detection accounts for 72.6% of the tool time while CPU spatial metrics take only 0.2%.

(3) We measure the weight memory of every loaded tool model and the number of calls it serves in Tab. R7.

Table R7. Weight memory and calls of each tool model loaded for SpaRW-Eval, corresponding to Fig. S4c in the revised manuscript.

Loaded Tool ModelWeight Memory, GBShare of Tool Weights, %Calls in One Run
Metric3D v25.5852.830
Florence-2-L1.5614.70
CountGD++0.948.92
Grounding DINO-B0.948.97,006
OWLv2-B0.625.93
Orient Anything V20.353.31,910
ViTPose-B0.343.30
SAM 2.1-T0.131.20
Depth Anything V2-S0.100.933
Total loaded weights10.56100.0-

Grounding DINO-B handles 7,006 calls with 0.94 GB of weights. Each call activates only the selected tool model, and the stack thus averages 8.1B active parameters per token and runs on 5 GPUs end to end.

AuthorsOpenReview ↗

Response to W1 & Q2 (2/2)

[Q2-2] How does this cost compare with tool-free reward models and the large-model baselines at matched evaluation quality?

[A2-2] SpatialCritic achieves accuracy comparable to GLM-5.3-Flash at 95.2% versus 94.6% and costs less to deploy, and no single-GPU tool-free model reaches this accuracy.

(1) Tab. R8 compares the deployment cost per pair at a concurrency of 1, as reported in App. A.6.

Table R8. End-to-end cost of SpatialCritic and GLM-5.3-Flash at a concurrency of 1, corresponding to Tab. S6 in the revised manuscript.

SpatialCriticGLM-5.3-Flash
Architecture8B dense + 16 toolsMoE
Total parameters8.8B + 6.0B tools320B
Active parameters per token8.1B†18B‡
Weight precisionBF16FP8
Policy weight memory18.2 GB326.7 GB
Tool server memory4 × 13.5 GB-
GPUs for the policy116
GPUs end to end516
Latency per pair30.2 s153.5 s
Policy LLM latency27.5 s-
Tool latency2.6 s-
GPU-seconds per pair1512,456
Accuracy, mean ± std95.2 ± 0.094.6 ± 0.1

† denotes the 8.0B language model of the policy plus 0.1B tool parameters per call, averaged over call frequencies. ‡ denotes 8 routed experts and 1 shared expert out of 288.

As shown in Tab. R8, SpatialCritic achieves accuracy comparable to GLM-5.3-Flash at 95.2% versus 94.6% while reducing latency from 153.5 s to 30.2 s, end-to-end GPUs from 16 to 5, policy weight memory from 326.7 GB to 18.2 GB, and active parameters from 18B to 8.1B.

(2) Tool-free 7B and 8B models run on a single GPU. Their highest accuracy is 87.5% on SpaRW-Eval in Tab. R1 and 88.3% after RL in Tab. S2, below 95.2% for SpatialCritic, and Qwen3-VL-8B obtains 64.4% on What'sUp versus 98.3% for SpatialCritic in Tab. R4.


We thank the reviewer for this suggestion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank the AC and all reviewers for their constructive feedback. Reviewers bb6P and GJgY recommend weak acceptance, and all reviewers recognize that SpatialCritic improves spatial verification from 84.9% to 95.1% across held-out generators and improves downstream generation on SD3.5-Medium and BAGEL.

To address the main concerns raised by the reviewers, we made the following key changes.

  • We strengthened the category-balanced and external evaluation (raised by bb6P [Q1], GJgY [W1, Q1], and 1wPn [W4, Q4]) by adding equal-category Macro accuracy with 95% confidence intervals in Tab. 1, five repeated runs in App. A.4 and Tab. S4, and 820 real-photograph pairs from What'sUp in App. A.5 and Tab. S5. Across five repeats, 8B SpatialCritic reaches 95.2% Overall and 91.5% Macro accuracy, higher than 94.6% and 90.5% for 320B GLM-5.3-Flash in every run, and per-category standard deviations range from 0.0 to 1.9. On What'sUp, SpatialCritic raises accuracy from 64.4% for its tool-free policy to 98.3%, on par with 98.9% for GLM-5.3-Flash, while using only generated images in its post-training. We also revised the abstract and Sec. 5.1 to state that SpatialCritic is comparable to GLM-5.3-Flash.

  • We expanded the end-to-end cost analysis (raised by bb6P [W1, Q2] and 1wPn [W3, Q3]) by profiling runtime, tool calls, memory, and active parameters over full SpaRW-Eval passes in App. A.6, Figs. S3 and S4, and Tab. S6. The policy LLM takes 91.6% of the 30.1 s per pair and tools take 8.0%. Compared with GLM-5.3-Flash, SpatialCritic reduces latency per pair from 153.5 s to 30.2 s, GPUs from 16 to 5, and active parameters from 18B to 8.1B, and completes 16.3 times as many evaluations per GPU-hour.

  • We specified the curriculum and its overhead (raised by GJgY [W2, Q2] and 1wPn [W3, Q3]) by documenting the 15,222-prompt curriculum pool in App. D.5 and Tabs. S16 and S17, the mastery probe and rollout accounting in Apps. D.3 and D.4, and the curriculum overhead in App. A.7 and Tabs. S7 and S8. The curriculum adds no rollouts, since review and remediation replace forward steps within the shared budget, and it adds 29.3 s, or 0.79%, to a 3,704 s training step.

  • We isolated the contributions of the reward and of agent–harness alignment (raised by 1wPn [W1, W2, Q1, Q2, Q4]) by clarifying the differences from SpatialReward, SpatialScore, and GenEval-2 in Secs. 2 and 5.2, adding matched-curriculum and alignment ablations in App. A.3 and Tab. S3, and adding a blinded user study in App. A.8 and Fig. S5. Under the identical curriculum, SpatialCritic reaches 83.9, compared with 82.0 without agent–harness alignment, 79.8 for SpatialReward, and 76.8 for SpatialScore. Under the same RL method Flow-GRPO, SpatialCritic without the curriculum exceeds GenEval-2 at 81.9 versus 81.2, and annotators prefer SpatialCritic in 83% and 88% of spatial judgments over the base model and the GenEval baseline.

We are grateful to reviewers bb6P, GJgY, and 1wPn for their thoughtful comments, which helped us improve the clarity and rigor of the paper, and we thank the Area Chair for their consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.