acceptodds
Under review as a conference paper at ICLR 2027

Self-Evolving Spatial Reasoning in Vision Language Models via Geometric Logic Consistency

12 views

Abstract

Vision-Language Models (VLMs) have made striking progress, yet their spatial reasoning remains fragile. Models that answer an original input correctly can still fail under valid transformations with predictable answer mappings, revealing a gap between instance-level correctness and robust spatial reasoning. To address this, we propose **S**patial **A**lignment via **G**eometric **E**volution (SAGE), a self-evolving framework that improves robust spatial reasoning through geometric and linguistic duality operations. SAGE incorporates duality consistency into GRPO training, encouraging models to produce coherent answers across original and transformed inputs. SAGE co-evolves duality generation and solution, allowing the model to continually expose and address its own reasoning weaknesses. A dynamic operation pool identifies challenging operations and retires mastered ones, keeping training focused on informative duality signals. SAGE is model-agnostic, data-efficient compared to prior post-training methods, and can be applied as a lightweight adaptation stage to any existing VLM. Experiments on video and spatial reasoning benchmarks demonstrate consistent improvements over strong baselines and enhanced generalization to unseen data.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Models, Recursive Self-Improvement, Self-Evolving Learning, Spatial Reasoning, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_osu4OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SAGE, a reinforcement learning framework to improve spatial reasoning in vision-language models.

Critical strengths

The idea of self-evolving post-training is interesting.

Critical weaknesses
  1. The paper is quite hard to read and follow. The method section is especially hard to follow, with terminologies such as "original–dual pairs", "canonical answer space", "extraction failure", "bijective answer transformation", "context–proposal examples (h,y)" are introduce but not clearly defined or explained. Due to the low readability of the method section, it's hard to provide constructive feedback on the methodology. The method figure (figure 2) is especially hard to read and to understand the method. There is too much information and mathematical equations in very small fonts. It will be more readable if it (1) presents the overall workflow without every mathematical detail, and (2) present the most novel and important designs of the method compare to prior methods.

  2. It is unclear whether the experiments sufficiently support the main claims. The authors also combine literature-reported results with their own evaluations. Some closely related methods, such as SVQA-R1 and SpatialEvo, are discussed but not directly compared with SAGE. Without adapting existing methods to the same data and training conditions, it is difficult to assess SAGE's improvements over prior approaches. Additionally, the evaluation metrics are not clearly defined in the main paper. For example, it is hard to interpret what "VSI-Bench (+1.6–6.7)" means without consulting the appendix. Overall, the low readability of the experimental section makes it difficult to provide feedback.

Overall, in its current form, the paper has substantial issues with readability and experimental validation that need to be addressed before it is ready for publication.

Questions for authors

Additional questions related to W1:

  • Can the authors provide clear definitions and intuitive examples for the above notions, rather than relying primarily on mathematical equations?
  • What is the motivation behind each major method component? How does each component address limitations of prior methods?

Additional questions related to W2:

  • Table 3 compares SAGE with fixed transformations and without online discovery. Can the authors provide a comparison between fixed and self-evolving transformations? Is adaptation the only difference between these settings?
  • How does the choice of VLM for proposing duality operations affect SAGE's performance? What happens when using different VLMs as proposers?
9 comments
AuthorsOpenReview ↗

Response to Reviewer osu4 Part 1

We sincerely thank the reviewer for the careful reading and detailed suggestions. Your feedback has guided our revisions to the method presentation and experimental comparisons. Below, we clarify the key definitions and design choices, present the added comparison, and address each question.

W1 Q1. Definitions, examples, and the method figure

The revised Figure 2 highlights the workflow, evolving pool, and correctness gate. We enlarged key labels and simplified the connections between probing, operation evolution, and consistency training. Detailed optimization equations now appear in the text and appendix rather than the figure.

Consider a cup to the left of a plate in the camera view. Horizontal reflection places it on the right. This example illustrates the terms below.

TermMeaning
Original–dual pairAn input and its transformed counterpart, whose answers follow a specified mapping.
Canonical answer spaceTask answers extracted from completions, such as directions, option labels, or numbers.
Extraction failureA completion with no extractable task answer is counted as incorrect.
Bijective answer transformationA reversible, one-to-one mapping between answers. Horizontal reflection exchanges left and right.
Context–proposal example (h,y)(h,y)hh contains operation history and solver performance on original–dual pairs. yy specifies transformation code, an answer mapping, applicability conditions, and a rationale.

Applicability conditions specify when a mapping is valid, here for left/right in the camera view. Answering “left” on both images is correct on the original but inconsistent across the pair. The proposer outputs an operation y, while the solver answers the task.

W1 Q2. Component motivation and distinction from prior work

SAGE adapts its transformation pool to the solver and uses paired predictions to guide learning. Probing finds inconsistently handled relations, guiding operation selection. Discovery adds operations, while retirement and rechecking keep the pool responsive as the solver changes. The correctness gate prevents rewarding consistency between incorrect answers.

The submitted paper already compares methods across 13 benchmarks, including published work such as Video-R1, VideoRFT, and SpatialLadder. Although ICLR's guidelines do not require comparisons to work available solely on arXiv [3], we appreciate the relevance of the suggested works and have expanded our comparison.

In their arXiv versions, SVQA-R1 [1] uses view-consistent rewards with prescribed spatial perturbations, while SpatialEvo [2] evolves question generation and solving using geometric environments for supervision. SAGE proposes executable transformations of existing labeled examples, with explicit answer mappings and applicability conditions, and uses solver feedback to manage them. The revised Stage 2 explains each component's role before its optimization details.

References

[1] Wang P, Ling H. SVQA-R1: Reinforcing spatial reasoning in MLLMs via view-consistent reward optimization[J]. arXiv preprint arXiv:2506.01371, 2025.

[2] Li D, Zhao Y, Cheng X, et al. SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments[J]. arXiv preprint arXiv:2604.14144, 2026.

[3] ICLR 2027. Reviewer Guidelines. FAQ on comparisons with recent and non-peer-reviewed work.

AuthorsOpenReview ↗

Response to Reviewer osu4 Part 2

W2. Direct comparison and evaluation conditions

We appreciate the reviewer’s concern about combining literature-reported scores with our own evaluations. We initially relied on some reported results because reproducing the full evaluation required substantial implementation and computational effort. Following this suggestion, we have rerun those evaluations under our protocol and replaced the reported scores with our own results.

We also added direct comparisons with the two suggested methods. We reproduced SVQA-R1 on Qwen2.5-VL-7B and evaluated the released SpatialEvo-7B checkpoint using the same evaluation protocol as SAGE.

MethodVSI-BenchViewSpatialVideoMMMUMMVU (MC)MVBenchTempCompassVideoMMECV-BenchSPARMMSIMindCubeOSTOmniSpatialAvg.
Qwen2.5-VL-7B (base)32.537.446.159.259.170.450.674.221.826.429.538.123.743.8
SVQA-R1-7B32.037.148.959.060.470.852.474.022.626.230.738.924.244.4
SpatialEvo-7B41.142.848.258.259.368.252.475.331.729.233.338.930.746.9
SAGE-7B37.944.153.465.868.178.466.577.234.432.344.546.338.352.9

SAGE achieves an average of 52.9, compared with 44.4 for SVQA-R1 and 46.9 for SpatialEvo. It outperforms SVQA-R1 on all 13 benchmarks and SpatialEvo on 12, with SpatialEvo scoring higher on VSI-Bench. The revised related work explains how SAGE's evolving operation pool and correctness-gated consistency reward differ from these approaches. We clarify the source and evaluation settings of each result in the appendix.

W2. Meaning of the reported metrics

The reported VSI-Bench range summarizes absolute gains across three backbones. These are +5.4 for Qwen2.5-VL-7B, +1.6 for Qwen3-VL-4B, and +6.7 for Qwen3-VL-8B.

VSI-Bench equally averages multiple-choice accuracy and numerical mean relative accuracy. The latter averages correctness across relative-error thresholds. Our ablation average weights all 13 benchmarks equally. We clarified the gains in the main results discussion and VSI/MMVU scoring in the table caption. The appendix gives full metric definitions.

W2 Q1. Fixed versus self-evolving transformations

Table 3 tests operation selection in Panel C and online pool evolution in Panel D. Panel C changes which available operations are used, while Panel D changes whether the pool can evolve.

SettingOperation useAverage score
Fixed-allApply every applicable operation49.3
Random-KKSample up to KK operations randomly47.0
Adaptive-KKBalance usage across operations49.8
Fixed initial poolDisable discovery and retirement/replacement44.6
SAGESelect by probe feedback and evolve the pool52.9

The 44.6 setting fixes the initial operation pool by disabling discovery and retirement/replacement. It retains the same backbone, initialization, training data, and reward. It tests pool evolution as a whole, rather than discovery alone. We clarified its label and protocol. Removing only retirement gives 48.6, and removing only reactivation gives 50.2.

Random-KK and Adaptive-KK share SAGE's active-capacity limit and solver rollout budget. Fixed-all uses more operations, costing 42.8 GPU-hours versus SAGE's 26.8.

W2 Q2. Different VLMs as proposers

We evaluated Qwen2.5-VL-7B, Qwen3-VL-4B, and Qwen3-VL-8B as the shared proposer and solver. The following table summarizes the main results across all 13 benchmarks.

Shared proposer and solverBase avg.SAGE avg.Gain
Qwen2.5-VL-7B43.852.9+9.1
Qwen3-VL-4B46.055.7+9.7
Qwen3-VL-8B48.458.9+10.5

SAGE improves every benchmark for each of the three VLMs. These results show that the shared proposer–solver design works across the tested model families and sizes.

Using separate proposer and solver models achieves the same average score as sharing one model, but requires more computation and memory. The relevant results from Table 3 in the main paper are shown below.

ConfigurationAverage scoreGPU-hoursPeak memory (GiB/GPU)
Separate proposer and solver52.939.079.8
Shared proposer and solver52.926.876.2

Sharing provides a simpler architecture with one trainable model, reducing training cost by 31.3% in GPU-hours and peak memory by 3.6 GiB per GPU, while retaining the same average performance.

We especially appreciate the suggestions that helped clarify the manuscript and its comparisons, and welcome further discussion.

Dear Reviewer osu4,

Thank you for your insightful comments and suggestions. We have carefully addressed the raised concerns in our rebuttal and provided additional clarifications and experimental results where applicable. We would greatly appreciate it if you could take some time to review our responses.

Thank you again for your time and consideration.

Best regards,
Authors

Reviewer_7a5mOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Current VLMs are still bad at spatial reasoning. This paper finds that existing training objectives do not directly enforce consistency across paired formulations of the same spatial relation. To address this, the paper first formalizes paired transformations as duality operations and then proposes SAGE, a self-evolving post-training framework that integrates duality consistency into GRPO. Experimental results show that SAGE effectively improves the spatial reasoning performance of VLMs.

Critical strengths

The idea of incorporating duality consistency into the training objective to improve spatial reasoning ability is good and makes sense.

They perform extensive experiments to validate the effectiveness of the proposed method.

Critical weaknesses

The descriptions of some parts are not clear enough.

Questions for authors
  1. The design of duality probing is based on multiple-choice QA tasks. I wonder whether the trained model can also improve spatial reasoning performance for generation tasks.

  2. I wonder whether incorporating duality consistency into the training objective would affect/decrease the performance on other visual reasoning tasks that do not involve spatial reasoning.

  3. L178 states that “Flips exclude text-reasoning questions and samples for which …”. I wonder how you identify these samples in implementation.

  4. The writing of Section 2.2 can be improved. For example, it would be appreciated if, in the heading, the authors could briefly describe the overall pipeline of Stage 2 and the goal of each step. Also, I wonder what (h,y) specifically refers to here.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7a5m

We thank the reviewer for recognizing the motivation behind duality consistency and the breadth of our experiments. Your questions helped us clarify the evaluation, implementation, and presentation of the method.

Q1. Generation tasks beyond multiple choice

Both our training data and evaluation include tasks requiring generated answers without predefined options. These include numerical answers such as counts and spatial measurements.

During training, an applicable operation transforms the input and specifies the corresponding answer relation. The model answers both inputs, and correctness gated agreement contributes to the reward for the original response.

For example, horizontal reflection preserves object counts but reverses signed horizontal displacement in fixed image coordinates.

TaskOriginal targetTarget after reflectionAnswer relation
Number of objects33TA(a)=aT_{\mathcal A}(a)=a
Horizontal displacement from red to blue12 pixels−12 pixelsTA(a)=−aT_{\mathcal A}(a)=-a

Here aa is the original target answer. Neither question provides answer choices. In the counting example, answers of 3 on both views receive the consistency bonus, whereas answers of 4 on both views receive none. The numerical targets and answer relations allow correctness and consistency to be checked directly.

Our evaluation includes VSI-Bench regression, SPAR-Bench depth and distance estimation, and OST-Bench estimation and enumeration. The main video and spatial benchmark tables report the following overall scores for Qwen2.5-VL-7B.

BenchmarkBaseSAGE
VSI-Bench32.537.9
SPAR-Bench21.834.4
OST-Bench38.146.3

The benchmark metrics appendix specifies the question types and scoring rules. We will clarify that both training and evaluation include generated answers and include the counting example in the revised manuscript.

Q2. Effects on other visual reasoning tasks

The spatial gains are accompanied by improvements on broader video understanding benchmarks. The main video benchmark table reports the following results for the same backbone.

BenchmarkBaseSAGE
VideoMMMU46.153.4
MMVU (MC)59.265.8
VideoMME50.666.5

Training retains the original task accuracy and format rewards alongside consistency feedback. The revised evaluation description specifies the benchmark scope, including the MC subset used for MMVU.

Q3. How samples are excluded from a transformation

Each operation's applicability conditions determine which samples it can transform. For image operations, these conditions are implemented as a predicate receiving the question, answer type, and answer choices. When supplied, a predicate that returns false or fails to compile or execute excludes the sample from that operation.

For example, a horizontal mirror operation checks for “from the camera” or “camera perspective” in the question, and “left”, “right”, and either “front” or “back” in the options. A camera relative direction question can pass these checks; “What word is written on the sign?” fails the wording check and is excluded.

For operations with no predicate, the default is to accept image inputs. We have clarified the applicability conditions in the method section and documented the predicate inputs and default behavior in the appendix.

Q4 and W1. Stage 2 workflow and the meaning of (h,y)(h, y)

Stage 2 adapts the operation pool to the solver's changing weaknesses. After initialization teaches the proposal format, the online loop has four steps.

  1. Propose and validate. The shared VLM proposes a transformation, answer mapping, and applicability conditions. Executable checks reject malformed rules.
  2. Probe and admit. Predictions on original and transformed inputs reveal unresolved inconsistencies. Useful operations enter the pool once sufficient probe evidence is available.
  3. Select and train. Informative operations receive active slots and supply consistency feedback for solver training.
  4. Retire and revisit. Mastered operations release slots. Periodic probes allow reactivation when their relations become difficult again.

In the proposer initialization objective, hh is the input context and yy is the target operation proposal. The context contains operation history and solver pair profiles. The proposal contains transformation code, an answer mapping, an applicability predicate, and a rationale. For example, hh may describe errors on camera relative left/right questions, while yy specifies horizontal reflection and the left/right exchange on applicable samples.

We have added the workflow and component roles at the start of the Stage 2 description and defined (h,y)(h, y) before the objective. The method overview figure has also been revised for clearer layout and labels. We appreciate these suggestions and welcome further discussion.

The authors' responses address my concerns. Therefore, I would keep my score.

Esteemed Reviewer,

Thank you for engaging with our work helping to make it better. We will incorporate the suggested improvements in the revised manuscript. If there is anything else we can improve or answer, kindly let us know.

Best regards,
Authors

Reviewer_a22yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a training method to improve VLMs' spatial reasoning by imposing a consistency reward during RL training. It adds a consistency reward for answering both the original and a transformed question (e.g., a flipped image, where left becomes right) and adaptively selects which transformations to train on, including new ones proposed by the model itself. Trained on 10k examples, it shows gains across several video and spatial reasoning benchmarks.

Critical strengths
  • The correctness gate on the consistency bonus is well motivated and validated.
  • The method is well validated: trained with only 10K samples and evaluated on 13 benchmarks.
Critical weaknesses
  • SAGE has multiple components: training on augmented inputs with mapped answers, a correctness-gated consistency reward, an adaptive operation pool, and a self-proposing loop. The paper should show how much of each of these actually contributes to the improvements. The transformations themselves are standard data augmentation in computer vision, and Table 3 suggests augmentation alone accounts for a large share of the gain (Base 43.8 vs. SAGE 52.9 vs. augmentation only 51.4).
  • In Table 3, "without online discovery" scores only 44.6, close to Base 43.8 and below original-only GRPO 47.7, which uses no transformation. Do the authors have an explanation for why removing the proposer, while keeping the operations, the pool, and the reward, performs below plain GRPO?
Questions for authors

N/A

6 comments
AuthorsOpenReview ↗

Response to Reviewer a22y

We thank the reviewer for recognizing the correctness gate and the breadth of our evaluation. We agree that the contribution of generated pairs should be distinguished more clearly from that of the reward and operation management.

W1. Contributions beyond augmentation

The augmentation baseline already benefits from SAGE's proposer, using its generated pairs and corresponding answer mappings. Its improvement from 43.8 to 51.4 in Table 3 of the main paper therefore also reflects the value of SAGE's pair generation. SAGE builds on these pairs through consistency guided learning, further improving the score to 52.9.

Both approaches retain accuracy and format rewards. Dual augmentation updates on the original input and then its transformed counterpart, using both rewards in each update. SAGE uses predictions on both inputs to compute consistency and trains on the original completions with

r(oi)=racc(oi)+rfmt(oi)+rcons(oi).r(o_i) = r_{\mathrm{acc}}(o_i) + r_{\mathrm{fmt}}(o_i) + r_{\mathrm{cons}}(o_i).

Here, rcons(oi)r_{\mathrm{cons}}(o_i) is the correctness-gated consistency reward defined in the paper. Table 3 reports the following comparison, with the average score taken over all 13 benchmarks.

MetricDual augmentationSAGE
Average benchmark score51.452.9
Joint correctness (%)35.345.7
Wrong consistency (%)33.630.2
Training cost (GPU hours)37.626.8

SAGE gains 1.5 points in average score and 10.4 points in joint correctness, with 10.8 fewer reported GPU hours. Its benefit beyond augmentation concerns both prediction quality and training cost. This comparison evaluates the combined reward and update design, rather than the bonus alone.

To illustrate how the bonus affects individual predictions, we further analyze 2,048 pairs of original and mirrored inputs. The following percentages compare predictions with and without the bonus on these pairs, whereas Table 3 compares SAGE with dual augmentation on the paper's evaluation sets.

MetricWithout bonusWith bonus
Original accuracy (%)51.5666.99
Mirrored accuracy (%)49.9065.04
Joint correctness (%)24.8544.29

On the original view, 446 predictions improve and 130 regress. At the pair level, 217 cases improve from correct only on the original view to correct on both views, while 117 previously jointly correct pairs lose correctness on at least one view.

Panels C and D of Table 3 report the contributions of operation selection and maintenance.

ConfigurationAverage score
Fixed all49.3
Random K47.0
Adaptive K with usage balancing49.8
Without retirement48.6
Without reactivation50.2
Full SAGE52.9

Removing retirement and reactivation reduces the score by 4.3 and 2.7 points, respectively. These results support the corresponding choices within SAGE, although interactions among components preclude an additive attribution. The revised protocol distinguishes these ablations from dual augmentation and clarifies the role of transformed responses. We will also clarify how the augmentation pairs are generated.

W2. Why the fixed initial pool underperforms GRPO

The fixed pool repeatedly applies initial transformations that can produce low quality pairs, as observed in our inspection of the transformed examples. The 44.6 configuration disables both online discovery and retirement or replacement, leaving these operations in use throughout training.

Consider original examples AA, BB, and CC, with transformed counterparts A∗A^{\ast}, B∗B^{\ast}, and C∗C^{\ast}. Their learning signals depend on pair quality. For example, the response to A∗A^{\ast} enters the consistency bonus used to compute the advantage for AA. Thus, even when AA remains intact, a poorly matched transformation can affect its training. The correctness gate checks answers against the specified targets; it does not establish that the transformation preserves the intended semantics.

Repeated exposure to a narrow set of such transformations may bias learning toward their particular relations or artifacts. Through the consistency bonus, these pairs can alter the relative advantages used to train the original answers. This provides a plausible explanation for the fixed pool's 44.6 score falling below original only GRPO's 47.7.

Table 3 in the main paper supports the distinction between agreement and useful supervision. The fixed pool has higher consistency than original only GRPO, 74.0% versus 62.3%, but also more wrong consistent predictions, 38.3% versus 33.5%. Online discovery renews the available transformations, while retirement limits repeated use of operations that have become uninformative. Removing retirement alone also reduces the score from 52.9 to 48.6. The revised protocol explicitly identifies the mechanisms disabled in this comparison.

AuthorsOpenReview ↗

Response to Reviewer a22y

We especially appreciate your comments on the presentation, which helped us clarify the roles of the individual components and improve the exposition in the revised manuscript. We welcome further discussion and would be happy to address any remaining questions.

Thanks for the response. Two remaining questions:

  1. As we understand it, the dual-augmentation baseline replays pairs generated during a SAGE run, so it already includes the pool's selection and the proposer's operations. Could you report standard augmentation with only the built-in operations and the accuracy reward? That would separate SAGE's contribution from conventional augmentation.

  2. What fraction of training pairs came from discovered operations versus built-in ones? Which operations did the proposer discover, and how do they differ from standard image and question transformations?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair,

We sincerely thank you and the reviewers for evaluating our work. We are encouraged by the strengths identified in the reviews:

  • Reviewer 7a5m recognized the value of incorporating duality consistency into the training objective to improve spatial reasoning.
  • Reviewer a22y highlighted the correctness gate on the consistency bonus as well motivated and validated.
  • Reviewer a22y emphasized the evaluation across 13 benchmarks using only 10K training examples, while Reviewer 7a5m recognized the extensive experiments validating the method’s effectiveness.
  • Reviewer osu4 found the idea of self-evolving post-training interesting.

SAGE improves Qwen2.5-VL-7B from 43.8 to 52.9 across 13 benchmarks. The discussion added controlled comparisons to clarify the sources of this gain and strengthen the evidence for the method.

Following the discussion, Reviewer a22y raised the rating from 2 to 3 (weak acceptance), while Reviewer 7a5m maintained 3 (weak acceptance). Both confirmed that our responses addressed their concerns. Reviewer osu4 acknowledged improvement and that several concerns had been addressed, but maintained 1 (clear rejection). The latest reservation concerned the extent of revisions made in response to the reviewer’s requests. The core method and main SAGE results remain unchanged. We find it difficult to reconcile a score of 1, without even moving to 2, with the added evidence and acknowledged improvements. A score of 1 should rest on clearly identified substantive flaws, yet the latest reply did not identify any that remain unresolved.

We made these focused revisions in the constructive spirit of ICLR’s rebuttal process, where authors and reviewers work together towards a clearer and stronger paper. We would be grateful if they could be assessed through the concerns they address.

Summary of Rebuttal and Discussion

Key issue (reviewer)How we addressed itStatus
Contribution beyond augmentation (Reviewer a22y)Original-only GRPO scores 47.7, conventional augmentation 48.2, and dual augmentation using SAGE-generated pairs 51.4, versus 52.9 for SAGE. Relative to dual augmentation, SAGE’s reward and update design raises joint correctness from 35.3% to 45.7% and reduces training cost from 37.6 to 26.8 GPU-hours.Resolved. Reviewer a22y confirmed that the main concerns were addressed.
Operation discovery and evolution (Reviewer a22y, Reviewer osu4)Discovered operations supply 7,644 of 9,595 training pairs (79.7%). We observed 130 operations across nine families during training. Fixing the initial pool, which disables discovery and retirement/replacement, scores 44.6. Removing retirement or reactivation separately gives 48.6 and 50.2.Resolved. Reviewer a22y accepted the operation statistics. Reviewer osu4 acknowledged the added ablations.
Training comparability and related methods (Reviewer osu4)SVQA-R1 was retrained with the same Qwen2.5-VL-7B starting checkpoint, 10K examples, and number of epochs, scoring 44.4 versus SAGE’s 52.9. SFT controls using the same checkpoint and data score 48.1 with assistant-response loss and 47.6 with full-text loss. The released SpatialEvo checkpoint scores 46.9 under our evaluation protocol, also used to reevaluate additional baselines.Resolved. Reviewer osu4 acknowledged that the additional comparisons addressed several concerns.
Generated answers and broader reasoning (Reviewer 7a5m)Training and evaluation include generated numerical answers. VideoMMMU improves from 46.1 to 53.4 and VideoMME from 50.6 to 66.5.Resolved. Reviewer 7a5m confirmed that our responses addressed the concerns.
Different VLMs and model sharing (Reviewer osu4)Average gains across three backbones are 9.1, 9.7, and 10.5 points. Shared and separate proposer and solver models both score 52.9. Sharing reduces training cost from 39.0 to 26.8 GPU-hours.Resolved. Reviewer osu4 thanked us for the detailed response and acknowledged our efforts.
Implementation and presentation (Reviewer 7a5m, Reviewer osu4)Clarified applicability predicates, terminology, Stage 2, and metrics. Revised the workflow in Figure 2 and added six examples using real images in Figure 3.Resolved. Reviewer 7a5m confirmed resolution. Reviewer osu4 acknowledged improvement.

Revisions addressing reviewer comments are marked in blue for inspection. We appreciate your consideration of the revised manuscript and the evidence developed through the discussion.

Best regards,
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.