acceptodds
Under review as a conference paper at ICLR 2027

View-AID: Multi-View Spatial Reasoning via Auxiliary-Image Distillation

Abstract

Multi-view spatial reasoning requires multimodal large language models to reconcile fragmented observations into a coherent spatial state. Direct textual chain-of-thought (CoT) construction serializes this global state into local relational statements, leaving cross-step spatial consistency implicit. Conditioning CoT generation on a reference answer further specifies the endpoint without constraining the intermediate spatial process. We introduce View-AID, which improves spatial supervision at its source by externalizing task-relevant spatial hypotheses as auxiliary images before translating them into textual CoT. This visual-first construction enriches reasoning supervision while preserving a standard inference interface that uses only the original views and question. Confidence-Guided Policy Optimization (CGPO) then complements binary outcome rewards with trace-conditioned answer confidence. On Qwen3-VL-4B, View-AID improves MindCube-Tiny and MMSI-Bench by 48.4 and 9.4 percentage points, respectively, and exceeds the strongest listed spatial baseline by 4.2 points in Overall accuracy. The gains transfer to disjoint spatial reasoning tasks and scale effectively to a larger backbone, supporting visual organization before textualization as an effective approach to multi-view spatial reasoning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
spatial reasoning, multimodal language models

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_pewCOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. This article proposes View-AID, which first generates auxiliary images based on multi-view images, and then converts the spatial relationships into text reasoning for training multi-modal models.
  2. After supervised fine-tuning, add reinforcement learning based on answer confidence. The final model only requires the original image and question, and does not need to generate auxiliary images.
Critical strengths
  1. Use auxiliary diagrams to organize cross-view spatial relationships, and then construct text supervision. The idea is intuitive and consistent with the characteristics of the task.
  2. The auxiliary graph is only used for offline data construction, does not increase the image generation overhead during inference, and the deployment method is relatively simple.
  3. Improvements are achieved on multiple spatial reasoning test sets, which also complement the 8B base and out-of-distribution task experiments, and are not limited to a single model or test set.
Critical weaknesses
  1. The independent contributions of auxiliary graphs have not been sufficiently separated. Compared with the original base, the improvement is great, but it lacks a strict comparison of directly generating text CoT under the same teachers, samples and training budget.
  2. The spatial accuracy of the auxiliary plots lacks validation. The generated diagrams may contain incorrect layouts, and it is difficult to rule out incorrect spatial relationships or to write explanations based on the answers based on the format and answer support checks alone.
  3. Confidence is not equal to the quality of reasoning. CGPO uses the probability of the final answer, but it has not been shown to improve the correctness of intermediate inferences, and direct evaluation of confidence calibration is lacking.
  4. Data construction details are incomplete. The specific version of the teacher model, training sample size, screening criteria and retention ratio are insufficiently explained, which affects reproduction and judgment of revenue sources.
Questions for authors
  1. Using the same teacher and training samples to directly generate answer-guided text CoT, how far is there compared with the auxiliary diagram solution?
  2. How accurate are the object correspondences, positions, and perspective relationships in the auxiliary images? Will error auxiliary maps be retained and affect training?
  3. Does CGPO actually improve confidence calibration and inference correctness? Can calibration metrics and inference process evaluation be provided?
  4. Can you add data size, teacher version, filter retention rate, and scenario-level deduplication description of training and test data?
3 comments

W1 — The independent contribution of auxiliary images is not sufficiently isolated. We respectfully clarify that a visual generator alone cannot synthesize textual CoT rationales; in our pipeline, rationales are generated by the lightweight Qwen3-VL-8B conditioned on synthesized auxiliary views. To rigorously evaluate the Reviewer's suggestion and highlight our advantages in parameter efficiency and reasoning performance, we distilled a Text-CoT baseline directly from a 4× larger teacher—Qwen3-VL-32B—without auxiliary views:

ModelAro.Rot.Amo.MindCube AllPos.Att.Mot.MSRMMSI AllOverall
Qwen3-VL-4B-Instruct41.634.120.035.934.232.326.023.727.731.8
Text-CoT (32B)76.038.057.358.137.835.330.018.833.045.6
View-AID-SFT82.877.586.083.641.836.028.722.235.559.6

Despite using a much smaller 8B VLM teacher, View-AID-SFT substantially outperforms 32B-distilled Text-CoT, boosting Overall accuracy from 45.6% to 59.6% (+14.0 pts). Notably, scaling the text-only teacher to 32B fails to resolve complex mental transformations—Text-CoT improves marginally on Rot. (34.1% to 38.0%) and degrades on MSR (23.7% to 18.8%). In contrast, View-AID-SFT (8B teacher) more than doubles Rot. accuracy (77.5%, +39.5 pts over Text-CoT) and surges to 86.0% on Amo. (+28.7 pts), driving MindCube All to 83.6% and MMSI All to 35.5%. This proves explicit view synthesis overcomes spatial bottlenecks that scaling a text-only teacher cannot solve.

W2 — The spatial accuracy of the auxiliary images is unvalidated. Ensuring visual validity motivates our Valid(z, y) filtering module (Algorithm 1), where Qwen3-VL-8B automatically prunes low-quality or irrelevant auxiliary images to ensure reliable AID-CoT data while preserving scalability. To quantify the trade-off between automation and per-sample accuracy, we manually audited 100 randomly sampled training instances against their original and auxiliary views across two binary criteria:

CriterionPASSFAILPass Rate
Logical Consistency99199%
Factual Grounding891189%

Logical Consistency checks for self-contradiction, while Factual Grounding verifies whether spatial claims (object presence, position, heading, motion) align with original views. AID-CoT attains 99% logical consistency and 89% factual grounding. Inspecting the 11 ungrounded cases reveals two error sources:

  1. Cross-view perception bottlenecks (9 cases): Original views contain sufficient evidence, but the model errs during multi-view integration (e.g., misbinding visual markers, forcing directional claims without clear correspondence, or miscalculating egocentric rotations).
  2. Auxiliary image defects (2 cases): The generated view hallucinates or crops out objects, causing the CoT to faithfully describe flawed inputs.

Thus, residual errors rarely stem from auxiliary image defects (only 2%), but rather from the base VLM's (Qwen3-VL-8B) inherent limits in multi-frame correspondence and coordinate transformation. Overall, our pipeline achieves a favorable trade-off between scalability and supervision quality. We will include this analysis in the revision.

W3 — Confidence is not equal to reasoning quality. To quantify probability calibration, we measure True Confidence—the mean probability assigned to the ground-truth token on correct predictions. We compared Qwen3-VL and View-AID on 100 paired samples randomly selected from MMSI-Bench and MindCube:

ModelCorrect AnswersLogical ConsistencyFactual GroundingTrue Confidence
Qwen3-VL24/100 (24%)57/100 (57%)18/100 (18%)25.0%
View-AID63/100 (63%)83/100 (83%)55/100 (55%)40.3%
View-AID Gain+39.0 pts+26.0 pts+37.0 pts+15.3 pts

Alongside substantial gains in accuracy (+39.0 pts) and CoT fidelity (+26.0 pts in Logical Consistency; +37.0 pts in Factual Grounding), CGPO training boosts True Confidence by +15.3 pts (25.0% to 40.3%), confirming improvements in both reasoning quality and confidence calibration.

W4 — Data construction details are incomplete. Our pipeline's "teacher" comprises two components: Flux2-klein-4B for auxiliary view generation and Qwen3-VL-8B for conditional verification and rationale curation. Matching ViewFusion's scale, our training set contains ~18k SFT and ~15k RL samples. In the curation workflow (Figure 1), Qwen3-VL-8B acts as the filtering and routing module: out of 18,570 SFT questions, 9,135 pass verification to construct AID-CoT rationales. We will fully open-source our code, data, and models.

Thank you for the author's reply; it resolved my doubts, and I will improve my score.

AuthorsOpenReview ↗

Thanks for your paticipating in our work!

We are delighted that our additional experiments have addressed your concerns and deeply appreciate your willingness to raise the score. Thank you once again for your valuable time and constructive feedback, which have significantly strengthened our paper.

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

This paper proposes View-AID, a training framework for improving multi-view spatial reasoning in VLMs. View-AID first uses an image synthesis model to generate an auxiliary image that serves as a visual cue for a given question. An MLLM then generates textual chains of thought (CoTs) conditioned on the synthesized image, and the resulting CoTs are used to train a model that does not require auxiliary images at inference time. The paper also introduces Confidence-Guided Policy Optimization (CGPO), which uses answer likelihood as a measure of confidence to prioritize CoTs that lead to correct answers with higher confidence over those with lower confidence. Experimental results demonstrate that View-AID and CGPO improve the performance of Qwen3-VL-4B on MindCube-Tiny and MMSI-Bench.

Critical strengths
  1. The paper is clearly presented, with writing, figures, and tables that are generally easy to follow.

  2. The idea behind CGPO, using answer likelihood to distinguish between CoTs, is intuitive, and Table 5 provides empirical evidence of its effectiveness compared with standard GRPO.

Critical weaknesses
  1. Comparison with a textual CoT baseline without auxiliary images: A comparison against a pipeline that generates textual CoTs without auxiliary images is needed to isolate the contribution of auxiliary images to the quality of the resulting training data.

  2. Comparison with the use of auxiliary images at inference time: Evaluating a setting in which the model also receives auxiliary images at inference time would provide a useful reference point. In particular, it would help clarify how much of the benefit of auxiliary visual information View-AID retains after distillation.

  3. Discussion of methods that generate images at inference time: View-AID can be viewed as distilling a reasoning pipeline that uses a world model or image generation model to produce visual cues for spatial reasoning. The paper would benefit from a more explicit discussion of its relationship to methods that use such cues at inference time, such as [1]. I do not believe these prior works undermine the novelty or contributions of this paper, but clarifying the connections and differences would help position the proposed approach.

[1] Yang et al., MindJourney: Test-Time Scaling with World Models for Spatial Reasoning, NeurIPS 2025.

Questions for authors

Have the authors evaluated image generation models other than FLUX? I am interested in how the choice and quality of the image generation model affect the quality of the resulting textual CoTs and the downstream performance of the trained model.

3 comments

W1 — Comparison against textual CoT without auxiliary images. To isolate the contribution of auxiliary views, we used a much larger Qwen3-VL-32B teacher to synthesize purely textual rationales (Text-CoT, without auxiliary views) for SFT:

ModelAro.Rot.Amo.MindCube AllPos.Att.Mot.MSRMMSI AllOverall
Qwen3-VL-4B-Instruct41.634.120.035.934.232.326.023.727.731.8
Text-CoT (32B)76.038.057.358.137.835.330.018.833.045.6
View-AID-SFT82.877.586.083.641.836.028.722.235.559.6

While 32B-distilled Text-CoT improves Overall accuracy over Qwen3-VL-4B-Instruct (31.8% to 45.6%), it hits a ceiling on complex mental transformations—notably Rot. (+3.9 pts to 38.0%) and MSR (dropping to 18.8%). In contrast, View-AID-SFT—using a much lighter pipeline (4B generator + 8B VLM)—surges to 77.5% on Rot. (+39.5 pts over Text-CoT) and 86.0% on Amo. (+28.7 pts), pushing MindCube All to 83.6%, MMSI All to 35.5%, and Overall to 59.6% (+14.0 pts). Achieving higher accuracy with far lower curation compute validates both the necessity of auxiliary views and our framework's effectiveness.

W2 — Auxiliary images at inference time. We compare View-AID against a test-time image generation baseline (based on View-AID-CGPO) on MindCube and MMSI:

ModelAro.Rot.Amo.MindCube AllPos.Att.Mot.MSRMMSI AllOverall
Qwen3-VL-4B-Instruct41.634.120.035.934.232.326.023.727.731.8
View-AID-SFT82.877.586.083.641.836.028.722.235.559.6
View-AID-CGPO85.274.087.384.343.636.031.423.737.160.7
Generation at Inference83.274.585.783.042.036.228.727.037.760.3

While View-AID-SFT slightly trails test-time generation (59.6 vs. 60.3 Overall), View-AID-CGPO closes this gap and edges out explicit test-time synthesis (60.7 vs. 60.3 Overall). This confirms our training paradigm effectively internalizes the generative teacher's spatial priors without test-time synthesis overhead.

W3 — Relation to inference-time image generation. We agree View-AID can be viewed as distilling a generative visual-reasoning pipeline, and clarify the connection and distinctions below. Connection. While methods like [1] invoke a generator at inference time to synthesize visual cues, View-AID uses the generator as a training-time teacher: the spatial reasoning it enables is distilled into the student model so these cues are leveraged implicitly at test time. Difference 1 — Inference efficiency. Test-time generation synthesizes NN views over TT rounds per query while serving a world model. View-AID entails no generative step in its inference graph—processing only original views in a single forward pass, regardless of the teacher's training compute. This efficiency is structural rather than empirical: the generator is absent at test time. Difference 2 — Locus of learning. View-AID enhances the VLM's intrinsic understanding by distilling cross-view reasoning into its chain of thought. Conversely, tool-augmented pipelines [1] query the generator as an external tool while the VLM remains largely fixed, relying on tool orchestration—deciding where to look and parsing returned views—while core spatial competence stays externalized in the world model. Briefly, View-AID enriches the model's internal representation, whereas tool pipelines augment its external interface. Practical implications. In latency-critical embodied settings (e.g., VLA policies, autonomous driving), multi-round image generation per query is prohibitive. Internalizing spatial competence removes external service dependencies for real-world deployment. Thus, the two directions are complementary: View-AID acts as an amortized form of test-time visual reasoning whose ceiling scales with stronger teachers at a constant, single-pass inference cost.

Question — Generators other than FLUX. Incorporating the recently released Mage-Flow (Mage) into View-AID-SFT further elevates our upper bound:

ModelAro.Rot.Amo.MindCube AllPos.Att.Mot.MSRMMSI AllOverall
Qwen3-VL-4B-Instruct41.634.120.035.934.232.326.023.727.731.8
View-AID-SFT (FLUX)82.877.586.083.641.836.028.722.235.559.6
View-AID-SFT (Mage)88.082.588.787.342.134.733.422.436.361.8

Atop our View-AID-SFT baseline (59.6%), Mage pushes Overall accuracy to 61.8% (+30.0% over base VLM), lifting MindCube All to 87.3% (+5.2% on Aro., +5.0% on Rot.) and MMSI All to 36.3% (+4.7% on Mot.), confirming seamless scalability with stronger generators.

Reviewer_VirbOpenReview ↗

Thank you for the rebuttal

Thank you for the clarifications in your rebuttal. Your response addressed my concerns. I have therefore raised my overall score.

I believe incorporating these clarifications into the revised manuscript will strengthen the paper.

AuthorsOpenReview ↗

Big thanks! :)

We sincerely thank Reviewer Virb for acknowledging our rebuttal and raising the rating! As promised, we will incorporate all newly conducted experiments and discussions from the rebuttal phase into the revised manuscript.

Reviewer_y2YBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces View-AID (multi-View Auxiliary-Image Distillation), built on the principle of organizing spatial relations visually before expressing them as textual supervision. Its core contributions lie in the AID and CGPO modules. When built upon Qwen3-VL-4B, View-AID achieves gains of 48.4 and 9.4 percentage points on MindCube-Tiny and MMSI-Bench, respectively, and outperforms the reported spatial baseline by 4.2 points in overall accuracy.

Critical strengths

The core contributions lie in the AID and CGPO components. The proposed framework achieves remarkable performance gains against the baseline and outperformes SOTA.

Critical weaknesses

Are there any failure cases? I encourage the authors to evaluate the method on more challenging examples to examine its robustness. For instance, test samples with varying illumination or substantial scene changes between paired images would be valuable.

Questions for authors

The authors only conduct experiments with Qwen3-VL-4B. It would be beneficial to see retraining results on other state-of-the-art open-source models.

Additionally, I would like the authors to discuss failure cases. Does View-AID consistently bring performance improvements across all test scenarios, or are there cases where it fails?

1 comment

We thank the reviewer for these constructive comments. Beyond Qwen3-VL-4B, we have also scaled our framework to Qwen3-VL-8B (see Table 2 in the main paper). We selected the Qwen3-VL series as our primary testbed because it represents one of the most widely adopted, state-of-the-art open-source VLM families. Furthermore, following the suggestion of Reviewer Virb, we evaluated our pipeline with an alternative visual generator (Mage-Flow) and observed substantial performance gains, which strongly validates the cross-model scalability and plug-and-play nature of our approach.

In addition, we conducted a qualitative failure case analysis to better understand the remaining bottlenecks. Take Sample #472 in MMSI-Bench[1] as a representative example:

  • Question: "The bicycle is moving forward. What is the order in which the pictures were taken? Options: (A) third, second, first; (B) second, first, third; (C) first, third, second; (D) second, third, first."
  • Prediction vs. Ground Truth: The model predicts A, whereas the ground truth is D.

This failure indicates that when temporal ordering hinges on subtle inter-frame displacements (e.g., slight forward motion of the bicycle across consecutive views), the model still struggles with fine-grained spatial perception. In future work, we plan to further enhance the model's sensitivity to subtle visual variations and fine-grained spatial-temporal dynamics.

[1]MMSI-Bench[https://huggingface.co/datasets/RunsenXu/MMSI-Bench]

Comments on the paper

2 comments

We thank Reviewers y2YB, Virb, and pewC for their constructive feedback. Notably, none raised fundamental theoretical flaws; their comments focused on empirical ablations, data quality verification, and positioning against test-time image generation. We conducted five new experiments and audits to respond to every point:

1. Key New Empirical Results & Findings

  • Auxiliary Views vs. Text-Only CoT Scaling (Virb W1, pewC W1): To isolate auxiliary views, we distilled a textual CoT baseline (Text-CoT) from a 4× larger teacher (Qwen3-VL-32B) without auxiliary images. Using a much lighter pipeline (4B generator + 8B VLM), View-AID-SFT (59.6% Overall) outperforms 32B Text-CoT (45.6%) by +14.0 pts across MindCube and MMSI. Scaling the text-only teacher to 32B fails on complex mental rotations (Rot.: 38.0% vs. 77.5% for View-AID-SFT, +39.5 pts), proving explicit view synthesis overcomes spatial bottlenecks that text-only scaling cannot resolve.
  • Comparison & Positioning vs. Test-Time Generation (Virb W2, W3): Across MindCube and MMSI, View-AID-CGPO (60.7% Overall) edges out explicit Generation at Inference (60.3%) in a single forward pass without test-time synthesis overhead (NN views ×\times TT rounds). Conceptually, View-AID acts as amortized test-time visual reasoning: rather than augmenting an external tool interface at test time, it distills cross-view competence into the VLM's internal representation for latency-critical embodied tasks (e.g., VLA, autonomous driving).
  • Scalability with Stronger Generators (y2YB, Virb Q1): Replacing FLUX with Mage-Flow elevates View-AID-SFT from 59.6% to 61.8% Overall (+30.0 pts over Qwen3-VL-4B-Instruct at 31.8%, reaching 87.3% on MindCube All). Alongside our Qwen3-VL-8B results (Table 2), this confirms seamless scaling as generators and VLMs advance.
  • Manual Audit of Data Fidelity (pewC W2): Auditing 100 random AID-CoT instances post-Valid(z, y) filtering shows 99% Logical Consistency and 89% Factual Grounding. Among the 11 ungrounded cases, only 2% stem from auxiliary image defects (9% arise from the 8B VLM's cross-view integration limits), confirming effective hallucination suppression.
  • Confidence Calibration & CoT Quality (pewC W3): On 100 paired test samples from MindCube and MMSI, View-AID (CGPO) boosts accuracy (+39.0 pts, 24%$\to$63%), Logical Consistency (+26.0 pts, 57%$\to$83%), Factual Grounding (+37.0 pts, 18%$\to$55%), and True Confidence (+15.3 pts, 25.0%$\to$40.3%), confirming genuine reasoning gains over superficial overconfidence.
  • Pipeline Clarity & Failure Analysis (y2YB, pewC W4): We clarified exact curation statistics (9,135/18,570 SFT queries pass Qwen3-VL-8B filtering to form AID-CoT, plus ~15k RL samples) and analyzed MMSI-Bench Sample #472 on subtle inter-frame motion challenges.

2. Summary of Reviewer-by-Reviewer Responses

ReviewerPoint RaisedOur Response & New EvidenceRevision Action
y2YBModel scalabilityHighlighted Qwen3-VL-8B (Table 2) + new Mage-Flow generator results (61.8% Overall).Add Mage-Flow ablation
y2YBFailure casesAnalyzed MMSI-Bench #472 on subtle inter-frame temporal displacements.Add failure analysis
VirbText-CoT ablation (W1)Added Qwen3-VL-32B Text-CoT; View-AID-SFT leads by +14.0% Overall (+39.5% on Rot.).Add 32B Text-CoT table
VirbTest-time gen. (W2)Added test-time generation baseline; View-AID-CGPO edges it out (60.7% vs. 60.3%).Add empirical comparison
VirbConceptual relation (W3)Clarified amortized representation learning vs. test-time tool interface.Expand related work
VirbOther generators (Q1)Added Mage-Flow results (+2.2% Overall over FLUX-based View-AID-SFT).Add generator table
pewCIsolating views (W1)Added 4× larger Qwen3-VL-32B Text-CoT ablation (59.6% vs. 45.6%).Add 32B Text-CoT table
pewCImage accuracy (W2)Audited 100 AID-CoT samples: 99% consistency, 89% grounding, only 2% image defects.Add data quality audit
pewCConfidence vs. CoT (W3)Audited 100 test pairs: +37% grounding, +26% consistency, +15.3% true confidence.Add calibration analysis
pewCCuration details (W4)Detailed teachers (Flux2-klein-4B + Qwen3-VL-8B) and pass rate (9,135/18,570).Clarify curation stats

We will incorporate all new results and discussions into the revision and open-source our code, data, and checkpoints.

AuthorsOpenReview ↗

Revised manuscript upload !

Thank you for all the reviewers' prompt feedback and for acknowledging the additional experiments conducted during the rebuttal phase. We have uploaded the revised manuscript for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.