acceptodds
Withdrawn from ICLR 2027

Binding What to Where: Configurational Binding via On-Policy Self-Distillation for Multi-Object Spatial Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-Language Models (VLMs) excel at visual understanding, but still struggle to reason reliably about spatial relationships. Models may capture relevant object cues while failing to bind them into a globally consistent spatial configuration, limiting their ability to reason about relations across multiple objects. Models may capture relevant visual cues while failing to organize them into a coherent spatial representation, limiting their ability to reason reliably about spatial structure. To address this, we propose BindOPSD, a post-training framework that strengthens spatial reasoning through configurational binding. BindOPSD first constructs a privileged teacher context with a joint spatial view and correspondence-aware attention, which organizes local object evidence within a shared spatial frame. It then transfers this capability to the student through on-policy self-distillation, enabling the model to internalize the enhanced spatial reasoning while requiring only the original image at inference. To preserve general visual capabilities during spatial adaptation, we further introduce gradient decomposition to reduce interference between spatial learning and general capability retention. We also introduce FineBind, a fine-grained spatial reasoning benchmark designed to evaluate global configuration and object-to-position binding. Experiments across diverse spatial reasoning benchmarks demonstrate consistent improvements over strong baselines, with effective transfer beyond the training distribution and no additional visual inputs at inference.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-Language Models, Multi-Object Spatial Reasoning, On-Policy Self-Distillation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FQorOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes BindOPSD to improve object-to-location binding in vision-language models. It equips a teacher with a joint scene view and correspondence-aware attention, then transfers this spatial signal to a student that uses only the original image at inference through on-policy self-distillation. The authors also introduce a gradient correction and the FineBind benchmark. Experiments with 4B and 9B models report gains on FineBind and most of nine external spatial benchmarks; ablations examine the contributions of the joint view and correspondence-aware attention.

Critical strengths

(1)With the joint view held fixed, adding correspondence-aware attention improves the overview-only student on all four ablation benchmarks at both model sizes. This is the clearest evidence for the paper’s central mechanism. (2)Breadth of reported gains: The 9B student improves on eight of nine external spatial benchmarks while using only the original image at inference. If the inconsistent baseline figures are reconciled, this would be a meaningful practical result.

Critical weaknesses

(1)Inconsistent baselines: The 9B FineBind baseline is 47.0 in Table 1 but 53.2 in Table 4; other baseline scores also differ. These discrepancies change the claimed gains and prevent a reliable assessment of the main result. (2)Insufficient validation: FineBind’s construction and split controls are poorly documented, and there is no matched ablation for the proposed gradient correction despite declines in the 9B model’s general reasoning scores. These gaps weaken two central claims.

Questions for authors

(1)Why do the same base models have different scores in Tables 1–2 and Table 4? Please identify the evaluation protocols and provide corrected results. (2)How were FineBind examples sourced, checked, and split to prevent image or scene overlap between training and test sets? (3)What matched comparison without gradient correction supports its claimed retention effect, given the 9B model’s declines on the reported general reasoning tests?

Reviewer_v5NuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BindOPSD addresses object-to-position binding in multi-object spatial reasoning. JOINT constructs a shared scene view, while correspondence-prior attention (CPA) associates object crops with locations in that view. On-policy self-distillation transfers the privileged teacher's predictions to a student that uses only the original image at inference. The paper also introduces FineBind and evaluates the method on additional spatial reasoning benchmarks.

Critical strengths
  1. Preserving both object details and relative layout is a relevant problem. The teacher-side design has a clear motivation, and the deployed student does not require the additional visual modules.
  2. Evaluation covers two model sizes and multiple benchmarks, with substantial improvements on FineBind and SpatialQA.
Critical weaknesses
  1. The necessity of the proposed design is not established by the comparisons. A simpler teacher input could mark target objects with numbered boxes in the original image and provide correspondingly numbered crops. This explicitly associates local details with global positions without JOINT reconstruction or attention modification. The paper does not establish an advantage over this alternative. Moreover, comparisons mainly involve spatial reasoning models with different backbones, many of which underperform the paper's starting model. A conventional visual OPD baseline, such as the cited Vision-OPD, with the same backbone, training data, and comparable training and visual-input budgets is needed to isolate the contribution beyond ordinary crop-based distillation.
  2. Result inconsistencies and insufficient FineBind documentation weaken the empirical evidence. Table 1 reports FineBind base-model scores of 31.00/47.00 for 4B/9B, whereas Table 4 reports 45.80/53.20 without clearly explaining the differing conditions. Section 4.1 and Appendix C give dataset sizes and three broad task categories, but insufficient detail about image sources, annotation and validation procedures, category distributions, or image-level train/test separation. These details matter because FineBind supplies both training and evaluation data. Relative to the strongest competing result in each column, the largest margins are on FineBind (+12.80 points) and SpatialQA (+4.82); the other seven wins are only 0.48–1.60 points, with a tie on SpatialMQA. Thus, the evidence for broad generalization is less compelling than the headline results suggest.
Questions for authors
  1. What evidence establishes an advantage over a teacher using numbered boxes on the original image plus corresponding crops, under equivalent object information and comparable visual budgets? Please report any existing comparison.
  2. Why do the base-model results differ between Tables 1 and 4? Please clarify the checkpoints, splits, prompts, and evaluation protocols, and identify the directly comparable results.
  3. Are matched-backbone results available for conventional visual OPD methods such as Vision-OPD? How do the existing experiments isolate JOINT/CPA from the benefits of privileged crop-based distillation?
  4. Please specify FineBind's image sources, question/answer construction and validation, task-category counts, and whether training and test sets are separated at the image or scene level.
Reviewer_UvHnOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors enhance the visual input and cross-view binding capabilities of a homologous teacher model via JOINT and CPA, and then transfer these capabilities to a student model, which does not receive the augmented inputs through on-policy self-distillation. This paper demonstrates that both global and cropped views play a vital role in visual reasoning tasks, and that such capabilities can be internalized into the model to a certain extent via self-distillation. Finally, the authors validate the effectiveness of their approach across a wide range of benchmarks.

Critical strengths

The paper proposes a coherent and low deployment cost solution for multi-object spatial binding. Specifically, JOINT preserves the joint spatial layout of relevant objects, while CPA explicitly associates local object evidence with scene locations; subsequently, on-policy self-distillation transfers the teacher's advantages to a student model that only receives the original image. Experiments demonstrate consistent improvements across downstream tasks using this approach. This provides direct evidence for the practical value of correspondence modeling, while requiring neither additional cropped views nor auxiliary modules during inference, making the contribution practically significant for real world deployment.

Critical weaknesses

The paper does not report the results of equivalent distillation training using the 'original global image + individual object crops' (i.e., the Dense input defined in the paper) as the teacher's input without incorporating JOINT and CPA. If a Dense teacher can achieve comparable student performance, the independent contribution of the proposed spatial configuration binding mechanism would be significantly weakened. Additionally, the documentation and details regarding the self-constructed dataset are insufficient.

Questions for authors

What are the sources of the images used to construct FineBind, and how were the questions and ground-truth answers generated and validated?

Could the authors explain why the distilled student outperforms the teacher on several benchmarks, despite having access only to the original image rather than the teacher’s privileged visual inputs?

Have the authors evaluated the method using instruction-tuned models? What motivated the choice of Base models over their instruction-tuned counterparts?

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

Related papers

Discussion (0)

Sign in to comment.