Understanding the Limits of Vision Language Models Through the Lens of the Binding Problem

Recent work has documented striking heterogeneity in the performance of state-of-the-art vision language models (VLMs), including both multimodal language models and text-to-image models. These models are able to describe and generate a diverse array of complex, naturalistic images, yet they exhibit surprising failures on basic multi-object reasoning tasks -- such as counting, localization, and simple forms of visual analogy -- that humans perform with near perfect accuracy. To better understand this puzzling pattern of successes and failures, we turn to theoretical accounts of the binding problem in cognitive science and neuroscience, a fundamental problem that arises when a shared set of representational resources must be used to represent distinct entities (e.g., to represent multiple objects in an image), necessitating the use of serial processing to avoid interference. We find that many of the puzzling failures of state-of-the-art VLMs can be explained as arising due to the binding problem, and that these failure modes are strikingly similar to the limitations exhibited by rapid, feedforward processing in the human brain.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Peer review

Reviewer Ga5j8/10 · confidence 4/52024-07-04

Summary

Vision Language Models (VLMs), including models that are able to describe images, such as GPT-4v, and models that generate images given a text description, such as DALL-E, display several failures in tasks that humans are able to perform effortlessly. Such tasks include counting, localizing a given object in an image, and visual analogy. This paper interprets these failures as resulting from a trade-off between representational flexibility (the ability to generalize compositionally to novel combinations of features) and channel capacity (the number of entities that can be represented simultaneously). This trade-off corresponds to a well-established phenomenon in cognitive science, the binding problem. The binding problem identifies a series of failures in human visual perception, that arise in many situations in which human subjects need to process multi-object scenes under time constraints. Such constraints prevent them from processing those scenes in a sequential way, and force them to rely on parallel processing, which incurs into capacity limits. This leads to phenomena such as illusory conjunctions, in which a feature is bound to the wrong object: for example, a green triangle might be falsely detected in a scene comprising red triangles, green circles and red circles. The failures shown by VLMs are remarkably similar, suggesting that they might result from the lack of a serial processing mechanism like the one humans can rely on in the absence of time constraints. To test this hypothesis, in this work VLMs are tested on two classic cognitive tasks (counting and visual search), showing error patterns similar to time-constrained human subjects. In counting, a sharp drop in performance is observed with scenes comprising more than 4 to 6 objects, a similar range to that beyond which humans are unable to “subitize” (reliably count objects in parallel, without serial processing). In visual search, conjunctive search, in which a unique combination of features needs to be detected, led to a sharp drop in performance with increasing numbers of objects, while disjunctive search, in which targets can be detected based on a single feature, did not. These results are remarkably similar to the symptoms of the binding problem in humans. The authors further posit that these failures result from interference between target and distractor objects, and not by the number of objects alone. They test this hypothesis by parametrically generating scenes that vary in their feature entropy: the amount of variation in the two features, shape and color, across all objects. Using a scene description task, they find that errors were most frequent when the feature entropy was intermediate, leading to a combination of diversity in features and sharing of features between objects. Finally, the VLMs are tested on a visual analogy task. The known failures of these models in solving these tasks is hypothesized to also depend on limitations in the number of objects that can be processed in parallel. To alleviate this problem, the authors propose a simple “decomposed condition”: feeding the target stimuli (pairs of objects) to the models in sequence, rather than as part of a single image (”unified condition”). The decomposed condition mimicks the serial processing that humans adopt in similar tasks. In both the analogy task and extraction of individual object features, VLMs perform better in the decomposed condition than the unified condition, providing further evidence that VLMs’ failures are due to capacity limits in binding features to multiple objects in parallel.

Strengths

- The paper presents a simple framework to explain failure cases of VLMs, directly inspired by a long tradition of research in cognitive science. It is a great example of how cognitive science research can inform the study of widespread deep learning models, with many potential applications in evaluating and improving such models. - The research methods are sound: the link between the theoretical question and the experiments is clear, and models are tested across multiple tasks and conditions (e.g. 2D and 3D scenes, different object categories) to ensure that results are robust. - Creative, but at the same time simple and elegant, ways to measure the relation between the binding problem and VLM failures (feature entropy) and to alleviate the binding problem in VLMs (decomposed condition) are proposed. - The interpretation of the results is fair and clear.

Weaknesses

The weaknesses are mostly related to the lack of detail concerning some of the methods, or the lack of explanations about certain choices. No major shortcoming was found by this reviewer. - In some of the prompts used to run experiments on the models (e.g. for counting), the model is first prompted to describe every object in the image, and then to perform the task. In others (e.g. disjunctive visual search) the model is directly asked to answer the question of interest (e.g. whether a target object is present in the image or not). Moreover, prompts from some tasks (e.g. 2D conjunctive search) include a form of role-playing "You have perfect vision and pay great attention to detail...". Since it is well known that similar techniques can have a great impact on the quality of language models' responses, the authors should provide an explanation of why these techniques were used in certain conditions and not others. Otherwise, performance comparisons between conditions (e.g. conjunctive vs. disjunctive search) can become harder to interpret. - In the scene description task, a language model is used to parse the descriptions. These are, I suppose, harder to parse than the simple answers provided in other tasks, such as the number of objects in the counting task or True or False in the visual search task. The authors should show that the language model was accurate in parsing these descriptions. Manually validating all of them would be infeasible, but at least a few examples to show the language models' accuracy should be provided in the supplementary materials. - Also in the scene description task, the generation of scenes with different feature entropies should be explained in more detail. Did the feature entropy manipulation simply consist in setting a particular number of shapes and features, and assigning them to a fixed number of objects? Or was the feature entropy computed in a different way? Making this clear would also help interpreting the values on the feature entropy that are shown in Figure 3. - In Figure 3, it is a bit hard to tell how many levels of feature entropy each number of objects has. For example, the 10 objects condition has two small bends that make it look like there are only three levels. For visualization purposes, I would recommend adding markers to the plot. Additionally, the values on the x-axis don't seem to align with the actual levels of feature entropy that were used. I understand that different number of objects had different levels, making it impossible to accomodate all possible levels, but if at least some of them could be aligned that would make the plot easier to read. - Also related to the scene description task, an explanation should be added of how the "edit distance between the true description of the scene and the model's description", the measure of errors, was computed. Does this mean that, for each (ground-truth) object in the scene, the most similar object (in terms of shape and color) in the description was chosen, and differences between that object and the ground truth were counted (e.g. color different = 1 error, shape different = 1 error)? How was, for example, a missing object in the description counted? As 2 errors, as both the color and shape were missing? A detailed explanation should be added. - The visual analogy task is similarly missing a detailed explanation of the pair generation process. In particular, what were the relations included in the task? Figure 4 apparently shows a "sameness" relation, in which target pair 2 is more similar to the source pair as they both feature objects that differ in color. Was this the only relation included? Also, on page 8, line 272 you write that the incorrect target "shared only one of the relations" with the source pair. I assume that this means that in figure 4, for example, TP2 shares both the "same shape" and the "different color" relation with the source pair, while TP1 only shares "same shape". Making this explicit and clarifying with an example would make this part easier for the reader. - The decomposed condition was only applied to the analogy task: was there a reason for this? As it's proposed as a general recipe for alleviating feature binding problems in VLMs, it would be more convincing if it was shown to work on a variety of tasks. The other tasks in the paper, in particular, seem to me like they would be amenable to be presented sequentially to the models. For example, a sequence of images containing only a few items at a time could be presented in visual search, simulating the way in which human participants serially scan such displays with eye movements in visual search experiments. If I am missing some practical reasons why conducting these experiments with a decomposed condition would be infeasible, or uninformative, I would be grateful to the authors if they could explain them.

Questions

No particular questions. I have mentioned a few points that I believe should be clarified in the "weaknesses" section, as I believe the paper would benefit from the addition of their answers.

Rating

8

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

One limitation which has not been acknowledged is the use of closed-source models. This could compromise the reproducibility of this work, and will limit the ability of follow-up works to investigate the mechanisms underlying these findings.

Reviewer Ga5j2024-08-13

I thank the authors for their rebuttal. My concerns were quite minor in the first place, and mostly related to presentation. Beyond responding to my own concerns, given that the authors have substantially expanded the paper with more thorough analyses, I have increased my score by one point.

Reviewer MYwe7/10 · confidence 4/52024-07-08

Summary

This work investigates the performance of the vision language model (VLM) GPT-4v and text-to-image model DALL-E-3 on multi-object reasoning tasks. According to the literature, humans perform well on multi-object tasks because they learn a compositional representation of the world and use serial processing to dynamically combine object features, overcoming the binding problem. However, in time-constrained tasks, they fail to use serial processing and rely on parallel processing, leading to illusory conjunction. In this work, the authors hypothesize, without testing, that VLMs learn a compositional representation for the sake of generalization but are not compelled to develop a serial-processing mechanism. According to the authors, this leads to failure in multi-object reasoning tasks, thus exhibiting the same limitations as humans in time-constrained tasks. The authors test these limitations on four multi-object reasoning tasks: counting, visual search, scene description, and visual analogy. In these tasks, the authors show that the performance of VLMs decreases as the number of objects increases, supporting the conclusion that VLMs are not able to solve the binding problem because they lack the mechanism to perform serial processing.

Strengths

- The paper is straightforward and easy to understand. The authors did a good job making it accessible, and the presentation and sectioning help in grasping the content. - This work seeks to understand the limitations of VLMs through the lens of human cognition, which could provide valuable insights. - Originality in the problem statement: In terms of performance, VLMs in their normal regime are equivalent to humans in time-limited tasks. - The investigation utilizes various types of experiments, each guided by studies conducted on humans, and they look easy to redo and the empirical results seem relevant.

Weaknesses

My score would increase if the following weaknesses were addressed: - The hypothesis that VLMs learn a compositional representation for generalization purposes could benefit from a reference if possible, since the discussion depends on that hypothesis. - Despite conducting various types of experiments (four in total), each individual experiment lacks sufficient variability. For instance, in the counting experiment, although they vary the colors of the circles, they do not vary their sizes or shapes. Another example, for the poppout task it would have been nice to also alternate the color of target and distractors. - The authors carried out the experiment on a very small group of language models (two). Some typos:
 - Figure 1, top right: the x-axis should be integers in the plot of GPT-4v counting performance - Line 156, 188: should be ‘Figure 2’ and not ‘Figure 1’

Questions

If we carefully engineer prompts for VLMs to sequentially focus and analyze each object in an image, considering they can learn a compositional representation as claimed, thus they should theoretically possess all the elements necessary to solve the binding problem? Can you elaborate on that?

Rating

7

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

See “Weaknesses”

Reviewer gAnB7/10 · confidence 3/52024-07-12

Summary

**Summary:** The paper investigates the counting abilities of VLMs. It puts forward a cognitive science inspired perspective, where VLMs are fundamentally constrained in their processing of multi-object scenes due to a lack of serial processing mechanisms. **Decision:** The paper is well structured and investigates an interesting (albeit a bit constrained) question concerning VLMs object reasoning abilities. While some of the conclusions seem overstated to me, I nonetheless recommend accept, since the authors present novel findings that improve the research community's understanding of VLMs visual reasoning abilities.

Strengths

*Originality* The underlying question is timely and well motivated. The cognitive science derived approach is principled and the authors perform a number of interesting experiments to further delineate the issues underlying counting in VLMs. *Quality* The submission is technically sound. I don't agree with all resulting claims, but the methods are appropriate. *Clarity* The submission is clearly written and follows a good outline. *Significance* The paper yields new data which allows for unique conclusions, however the scope of the investigation is somewhat limited.

Weaknesses

I have brought it up in the questions sections.

Questions

- Figure 1: There is a spike at 9 objects for both models. Do you have an idea on why this might be the case? [1] (which you already cite) also find the same spike and offer some potential explanations for this (see Figure 5). - Answer parsing is again conducted by a large language model. I am wondering how accurate this parsing is. Did you compute any metrics to check the accuracy of the final parsed output or run some qualitative checks? - In line 189 you write "GPT-4v demonstrates human-like capacity constraints in its ability to perform visual search in multi-object settings." I was wondering what human level performance is in this task? It would be nice to have some more information to compare the model performance to. The same is true for line 127, where I'm not sure this is a fair comparison. Sure, the number of objects the models can count well are somewhat similar to the subitizing limit of humans, but this task should not require rapid estimation of the number of objects and therefore we would expect humans to be able to count well above 6 objects here anyways. - The sentence in line 303 reads a bit odd, it should probably read "facilitate generalization" instead of facilitation and there might be a comma missing before? - If I understand correctly you describe a different measure for the model performance for the last task in Section 4.2, would it not be possible to again report the proportion of correct answers here? I find it hard to compare this experiment with the previous experiments. - To me the take aways seem overstated. What does lack of serial processing for VLMs mean? I don't think VLMs are fundamentally unable to sequentially process an image and therefore they should in theory be able to "selectively attend to individual objects one at a time". This is also why I don't completely follow the conclusion that the VLMs show "human-like capacity constraints". I would argue they should be able to do sequential processing and should therefore not be measured against the human subitizing limit anyways. Maybe you can offer some more details on this argument. [1] Rane, S., Ku, A., Baldridge, J., Tenney, I., Griffiths, T., & Kim, B. (2024). Can Generative Multimodal Models Count to Ten? Proceedings of the Annual Meeting of the Cognitive Science Society, 46. Retrieved from https://escholarship.org/uc/item/8kz5787g

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The paper is well structured and investigates an interesting (albeit a bit constrained) question concerning VLMs object reasoning abilities. While some of the conclusions seem overstated to me, I nonetheless recommend accept, since the authors present novel findings that improve the research community's understanding of VLMs visual reasoning abilities.

Reviewer N3JA7/10 · confidence 4/52024-07-14

Summary

The paper puts forth the hypothesis that VLMs suffer from similar constraints as humans in multi-object visual processing, especially with respect to "binding". Results from cognitive science show human constraints in numerical estimation, visual search, and relation identification when not able to do serial processing, which is required for "binding" different object features together. The paper hypothesizes that lack of serial processing and binding in VLMs is responsible for failures in these tasks. The paper gives experimental results that support this hypothesis.

Strengths

This is a good paper -- it gives a cognitive-science-based approach to understanding VLM behavior. It is well-written and has a mostly clear discussion of hypotheses, experiments, and results.

Weaknesses

The paper leaves out some important information. For each task, the authors prompted GPT-4v to describe the input images, but as far as I can tell, they did not give any evaluation of how well GPT-4v was able to describe them. Other work (e.g., http://www.arxiv.org/abs/2407.06581) has shown that GPT-4v has trouble describing such synthetic images and often will hallucinate. Here, the authors should include the results on how well GPT-4v can even parse what is in the images (including shapes, colors, etc.) and the authors should assess whether that failure to describe simple images is related to its failure on these tasks. To what extent are the failures a matter of lack of binding features, or of missing or hallucinating features? A couple of typos/corrections: In Section 3, "Figure 1" should be "Figure 2". In the caption for Figure 2, "blue column" should be "red column" and "red column" should be "blue column".

Questions

Figure 2 caption: "Performance for 2D and 3D task variants" -- performance of what? GPT-4v? I also suggest in that caption that you should say what the targets are for each search (red circle, red "L", red sphere, red sphere). In Section 3.1., explain what the task is for GPT-4V: "Is there a red circle?" "Is there a red L? This is explained in the appendix (A.2) but should be explained here in the main paper. Looking at the visual search task as described in A.2, it seems that the task wasn't actually "search" in the sense that, unlike humans, GPT-4V doesn't have to localize the target object. In these kinds of tasks, aren't humans asked to localize the target? How does this affect the comparability of the results between GPT-4V and humans on this task? In Section 3.1, explain why in half the images, targets were left out. This doesn't make sense unless readers know what the specific task is. In Section 3.1 you say "VLMs demonstrate similar capacity constraints to humans." What do you mean by "similar"? Do humans show a similar success/failure curve as those in Figure 2? Or is the similarity just that humans perform well on disjunctive search and badly on conjunctive search (if not enough time for serial processing)? In Section 3.2, explain how the "accuracy" of the model is calculated. In Section 4.2 you refer to "the true description of the scene". Where did this "true description" come from? Also, define "edit distance". I was also confused because each scene could have different correct descriptions with different levels of detail. Was it possible that the VLM gave one version of a correct description but it was counted as wrong because it included details that the "true description" didn't contain? I was also confused by the visual analogy task in Figure 4. It seems to me that both Target Pair 1 and Target Pair 2 could be considered the correct answer. The source pair has two obvious relations: "same star shape" and "different color". The former would map to Target 1 and the latter would map to Target 2. Which Target did you consider correct, and why?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors do a good job of discussing limitations.

Authorsrebuttal2024-08-07

## Accuracy measure for scene task When submitting our rebuttal, we forgot to include our response to the question about including an accuracy measure for the scene description task. In this task, we chose to focus on 'edit distance' because it is a more sensitive measure than accuracy for this task. The models very often make at least one mistake for each image, meaning that accuracy (proportion of trials for which the entire scene description is correct) is very low, and thus not as sensitive for testing our specific hypothesis about binding errors. However, we agree that it is useful to include an accuracy measure for the purpose of comparing across tasks. We have now included these plots in the appendix. Thank you for the suggestion.

Reviewer gAnB2024-08-12

I thank the authors for their detailed explanation on serial and parallel processing. I now have a better understanding of the assumptions and objectives behind the submission. Since, as far as I can see, all of my major concerns were addressed, I will increase my score.

Reviewer N3JA2024-08-09

Thanks for the rebuttal

The authors' rebuttal and paper revisions respond to all the concerns I had. Thank you to the authors for the thorough rebuttal.

Reviewer MYwe2024-08-10

Thank you for thoroughly addressing all my questions and concerns in your rebuttal and revised paper. I appreciate the effort and clarity you’ve provided, and as a result, I have decided to increase my score by one point.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC