Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas

We thank Reviewer 3qjz for the comments. However, we are concerned that some comments suggest a disconnect or oversight of our paper. We respond to each point below, and respectfully encourage the reviewer to revisit our work.

Essential References Not Discussed

Among the four references mentioned, [1] and [3] are precisely the benchmarks we work on, and our experiments are exactly built on these two papers with the same dataset names WhatsUp and VSR cited and mentioned throughout the paper. Both are benchmark papers that present datasets and baselines like prompting, without proposing novel methods or addressing mechanistic interpretability (MI) .

[2] and [4] are: (1) about LLMs rather than VLMs, while we mentioned in the title that we work on VLMs; (2) neither about MI, while we highlighted in abstract (Lines 18-20), introduction (Lines 43-45), conclusion (Lines 422-423) that our paper is to open up VLMs and observe their internal behavior to understand their failures in spatial reasoning.

We believe the reviewer has a misunderstanding of the paper that we are NOT proposing a new model to better leverage object grounding to beat SOTA models, but rather conducting MI for VLMs, ​​by analyzing attention patterns and applying targeted interventions to understand their behavior. So the suggestions to compare with other object-grounding methods reflect the misunderstanding, detailed in W2.

(MI is a well-established area to reverse-engineer complex systems, offering a more granular and causal understanding of how they process information. [5,6,7])

[5] A Survey on Mechanistic Interpretability for Multi-Modal Foundation Models
[6] Mechanistic Interpretability for AI Safety: A Review
[7] awesome-mechanistic-interpretability-lm-papers in Github

W1:

datasets being small and primarily synthetic

• Benchmarks are not primarily synthetic, with 85% natural images and all human annotated captions (reformatted to generative QA settings better suited for SOTA decoding-only VLM architectures, Lines 143-149). | |Cont_A|Cont_B|Coco_one|Coco_two|VG_one|VG_two| |-|:-:|:-:|:-:|:-:|:-:|:-:| |Synthetic image|yes|yes|no|no|no|no| |Synthetic caption|no|no|no|no|no|no| |samples|406|408|1160|291|2240|440|
• MI research requires clean datasets to observe model mechanisms. Existing MI studies use smaller, clean datasets rather than larger, noisy ones [5, 6, 7]. We have made every effort to include all available datasets with clean, high-quality spatial relationship annotations.

W2:

do not compare to any prior work in their paper.

• We respectfully disagree. With a focus on MI of VLMs, our experiments on attention intervention aims to give conclusions about how VLMs work. Thus, experiments are expected to be between w/ and w/o intervention, rather than comparing with other object-grounding methods.
• We did compare our methods with two prior decoding methods: Dola [8] and VCD [9] (Line 327). Like DoLa, we leverage internal signals but tailor them to VLMs using visual features. While sharing VCD’s goal of improving visual grounding through decoding, we go further by introducing fine-grained attention interventions for more precise calibration.
• Again, we believe it reflects a misunderstanding of the paper, especially given the context that other three reviewers do not have such misunderstandings.

[8] DoLa: Decoding by Contrasting Layers Improves Factuality in LLMs.
[9] VCD: Mitigating Object Hallucinations in Large VLMs through Visual Contrastive Decoding.

W3:

Does not use established benchmarks.

• We use established spatial-reasoning benchmarks WhatsUp [1] and VSR [3], and as detailed in Lines 143-149, we have to reformat the dataset from classification settings to QA settings to observe the attention behavior and information flow across layers of Transformers when VLMs generate correct or wrong answers, which is also a standard setting of MI studies [5,6,7] for observing the attention behavior.
• We also include general VQA benchmarks in Table 6 at Appendix.

W4:

Unclear visualizations in Figure 9 & 10

Thanks for the helpful suggestions! We will add more annotations to these figures.

W5:

backing "attention modification" claims with some numbers

Thanks for the helpful feedback. We extend our experiments to test YOLO overlap w/ and w/o intervention in middle layers, where our observations (line 196) identify as key for processing spatial information. We evaluate two representative subsets: Controlled_A (synthetic images) and VG_two_obj (real images). After intervention, overlap rates improve on both. Attention–YOLO similarity increases by 2.9% on Controlled_A and 8.2% on VG_two_obj at layer 16, with similar trends across other middle layers.

While it can support our claims, it is not a golden metric since attention map is influenced by the prompt, and YOLO may fail to detect objects in complex images (Figure 24). We will include additional discussions in our paper.

We thank Reviewer p97N for the encouraging comments and thoughtful feedback. Below, we address the concerns raised in detail.

I'm wondering how well the proposed method works for other VLMs.

• Experiments on Qwen2-VL, a SOTA VLM with different architecture. We intervene in the image attention distribution using our temperature-scaling method, showing consistent improvements below, particularly in challenging cases. For example, on VG_two_object, where the baseline performance is the lowest among all benchmarks, our method yields a significant improvement of 10+ absolute points. The gains observed on Qwen2-VL further demonstrate the generalizability of our approach.

• Experiments on more benchmarks, including POPE, GQA, and TextVQA, which proves that attention intervention can maintain the performance on more general tasks without hurting the performance. Compared with spatial reasoning tasks, general QA tasks achieve relatively smaller improvement. A possible reason is that such tasks are less sensitive to the geometric structured distribution of image attention. For example, given a question like “Is there a dog in this picture?”, the model only needs to detect the presence of an object, and is therefore less likely to suffer from misallocated attention across spatial regions.

provide more results (analytical and experimental) using other VLMs

We present the additional experimental results above. For the analytical results, we also extend our analysis to the Qwen2-VL model and find that our insights generally hold across VLMs.

• The first claim image receives much less attention than the text tokens is generally valid for VLMs: we calculate the Qwen2-VL’s attention scores below (average attention logits for single text/image tokens respectively), which matches our previous claims. | Benchmark | Text attention scores | Image attention scores | |:-----:|:----:|:----:| | Controlled_A | 1.57e-02 | 7.59e-05 | | Controlled_B | 1.58e-02 | 7.69e-05 | | Coco_one_obj | 1.77e-02 | 4.48e-04 | | Coco_two_obj | 1.65e-02 | 3.50e-04 | | VG_one_obj | 1.54e-02| 4.75e-04 | | VG_two_obj | 1.42e-02| 4.19e-04|
• The second important claim is still valid about confidence could serve as a proxy for the model performance as below:
  • Spatial relationships with lower confidence, such as "Left" and "Under" tend to exhibit lower accuracy when compared to those with higher confidence, like "On" and "Right".
  • After the intervention with AdaptVis, certain spatial relationships like “Left: show improved performance, accompanied by an increase in confidence. This demonstrates a pattern consistent with our observations for LLaVA, as depicted in Figure 25 in the Appendix. | Benchmark | Qwen2-VL Uncertainty | Qwen2-VL+ScalingVis Uncertainty |Qwen2-VL Acc | Qwen2-VL+ScalingVis Acc | |:---:|:----:|:----:|:----:|:-----:| | Left | 0.4408 | 0.4474 |70.11 | 74.71 | | On | 0.5758 | 0.5668 |100 | 100| | Right | 0.5982 | 0.5911 | 100 | 100 | | Under | 0.5554 | 0.5418 |98.82 | 98.82 |

Q1:

What is the major difference between a spatial-relation understanding task and a standard QA task that makes the proposed method less effective on standard QA task?

• That’s a great question. In fact, our initial and primary motivation was to gain deeper insights into the spatial reasoning capabilities of VLMs. Through our analysis, we observed that the geometric distribution of image attention plays a critical role in this reasoning process. Based on this observation, we proposed two temperature scaling techniques to adjust the attention distribution and address the problem.

• As for the relatively smaller improvement on general QA tasks, we believe this may be because such tasks are less sensitive to the geometric structure of image attention. For example, given a question like “Is there a dog in this picture?”, the model only needs to detect the presence of an object, and is therefore less likely to suffer from misallocated attention across spatial regions.

We appreciate your thorough review and detailed comments! Your suggestions will be helpful in improving the paper. We address your concerns below.

Q1:

whether the claims of this paper are generally valid for VLMs or is it specifically valid for LLAVA models that this paper tests on.

• The claims are generally valid for VLMs.
• The first claim image receives much less attention than the text tokens is generally valid for VLMs: we calculate the Qwen2-VL’s attention scores below (average attention logits for single text/image tokens respectively), which matches our previous claims.

• The second important claim is still valid about confidence could serve as a proxy for the model performance as below:
  • Spatial relationships with lower confidence, such as "Left" and "Under" tend to exhibit lower accuracy when compared to those with higher confidence, like "On" and "Right".
  • After the intervention with AdaptVis, certain spatial relationships like “Left" show improved performance, accompanied by an increase in confidence. This demonstrates a pattern consistent with our observations for LLaVA, as depicted in Figure 25 in the Appendix.

Q2:

more experiments on other VLM and other benchmarks.

• Experiments on Qwen2-VL, a SOTA VLM with different architecture. We intervene in the image attention distribution using our temperature-scaling method, showing consistent improvements below, particularly in challenging cases. For example, on VG_two_object, where the baseline performance is the lowest among all benchmarks, our method yields a significant improvement of 10+ absolute points. The gains observed on Qwen2-VL further demonstrate the generalizability of our approach.

• Experiments on more benchmarks, including POPE, GQA, and TextVQA, which proves that attention intervention can maintain the performance on more general tasks without hurting the performance. Compared with spatial reasoning tasks, general QA tasks achieve relatively smaller improvement. A possible reason is that such tasks are less sensitive to the geometric structured distribution of image attention. For example, given a question like Is there a dog in this picture?, the model only needs to detect the presence of an object, and is therefore less likely to suffer from misallocated attention across spatial regions.

Thanks for your comments and advice, We address your concerns below.

it is unclear whether the findings would generalize to other VLMs.

• We agree generalizability is important so we have especially included different variances of LLaVA-series models, with LLaVA-1.5 (224×224 visual encoder, MLP projection layer) and LLaVA-1.6 (336×336 visual encoder, resampler projection layer), since they are the most widely-used open-source architecture but with different model features like confidence, as mentioned in line 742.
• To assess the generalizability of our findings, we extend our analysis to Qwen2-VL, a SOTA VLM with different architecture. We show our findings for VLMs’ attention patterns are still consistent [see response for Reviewer 3nwP Q1].
• To further assess the generalizability of our intervention approach, we extend experiments to Qwen2-VL. We intervene in the image attention distribution using our temperature-scaling method, showing consistent improvements below, particularly in challenging cases. For example, on VG_two_object, where the baseline performance is the lowest among all benchmarks, our method yields a significant improvement of 10+ absolute points. The gains observed on Qwen2-VL further demonstrate the generalizability of our approach.

• We also evaluate more generic benchmarks for generalizability, including POPE, GQA, and TextVQA, which proves that attention intervention can maintain the performance on more general tasks without hurting the performance. Compared with spatial reasoning tasks, general QA tasks achieve relatively smaller improvement. This may be because such tasks are less sensitive to the geometric structure of image attention. For example, when asked Is there a dog in this picture?, the model only needs to detect the object's presence, and is therefore less likely to suffer from misallocated attention across spatial regions.

• We believe this assumption is theoretically generalizable. The only assumption we have is the different information density between image and text. In VLMs, attention is distributed across both textual and visual tokens. Since key information in images tends to be more sparse compared to text, the attention distribution over visual tokens can be more difficult to allocate accurately.

The paper claims that the model's confidence is correlated with attention correctness…More robust, quantitative analysis to support the correlation between attention distribution and model confidence.

• As detailed in Lines 184-251 of Section 4.1, we use YOLO to annotate the relevant entities in the images, and conduct quantitative analysis on AUROC of the overlap between YOLO annotations and the attention patterns , as well as performance are shown in Figure 7 in the main paper and, Figure 18, Figure 19, Figure 20, Figure 21, Figure 22, Figure 23, Figure 24 in the Appendix.
• In Figure 25 in the Appendix, we perform quantitative analysis on confidence and accuracy and their correlation with the coefficient, where we can see the confidence and accuracy follow similar trends as the temperature coefficient varies, supporting the use of confidence as a reliable proxy for performance.
• From the first and third subfigures: for low-confidence relationships, applying a small coefficient (<1) improves performance; for high-confidence relationships, a coefficient (>1) improves performance.
• We will move the additional quantitative evidence (from Appendix) into Section 4.2 to make this reasoning clearer in the paper.

A more adaptive approach that adjusts thresholds dynamically based on the specific task or dataset...

• We believe there may be a misunderstanding. As described in lines 323, we adaptively adjust the coefficient using a small held-out subset (20% of the whole subset as the validation set) from the specific target task or dataset, rather than pre-defining hyperparameters.
• Additionally, we propose a more adaptive variant ScalingVis with only a single hyperparameter $\alpha$ to modulate the attention distribution based on model confidence. This design maintains both simplicity and adaptability, contributing to the method’s robustness and generalizability across tasks.