Core Knowledge Deficits in Multi-Modal Language Models

>>> Q1 The tasks are much harder than children cannot easily solve compared to developmental CogSci
>>> A1 Thanks for bringing up this nuance. We address the concern in two aspects. First, while all our tasks are derived from standard developmental cognitive science prototypes, e.g., "Three-Mountain", we extend the questions beyond the few typical examples from the textbooks to comprehensively evaluate the capabilities of MLLMs. This leads to a significantly larger and more systematic set of tasks and evaluations, rather than relying on the limited examples typically found in conventional literature. Second, it’s important to note that the subjects being evaluated are not infants or young children with limited acquired knowledge, but rather MLLMs adapted from large language models (LLMs), which are reported to possess human PhD-level knowledge and reasoning abilities [1].

[1] OpenAI (2024). Learning to reason with LLMs. https://openai.com/index/learning-to-reason-with-llms/

---

>>> Q2 One thing that this paper needs further discussion on is how the authors performed the human study, for example, how many human participants were recruited and how were they paid?
>>> A2 We recruited college students as participants, each receiving the same questions as the MLLMS. To ensure attentiveness, we included 1% mismatched question-answer or question-image pairs, asking participants to flag unclear or complex items within 90 seconds. Participants were paid only if they passed the attention check. 22 participants' answers were accepted for final statistics.

---

>>> Q3 Providing more task examples in the supplementary would be more helpful.
>>> A3 Thanks for the advice. We have provided test cases in both the main paper and the supplementary. We will add more examples following your suggestion.

---

>>> Q4 Suggested references.
>>> A4 Thanks for the advice. We will include all suggested papers in our citations and provide an extended discussion about them.

---

>>> Q5 Language is treated as a separate aspect of core knowledge, and it might account for explaining why this paper found low-level tasks are harder for models than high-level tasks. The VQA format, especially the fact that almost every task is entangled with language understanding, makes the core knowledge evaluated here not as clean as controlled experiments performed in the original developmental psych literature. This may lead to not directly comparable results (e.g., why low-level tasks are harder than high-level tasks for machines).
>>> A5 Thanks for raising this concern. We will address your concern in the following two aspects. First, indeed, we acknowledge that the VQA format is entangled with language understanding, as demonstrated by the impressive capabilities of large language models (LLMs) in both language comprehension and their use as a tool for reasoning [1, 2, 3], we argue that the influence of language bias is relatively minimal when evaluating core knowledge.

Second, we contend that the core challenge lies in developing a scientifically robust evaluation methodology that can specifically probe the abilities of LLMs and MLLMs. Regardless of the evaluation method—whether VQA, retrieval tasks, or using output logits—assessing the model's specific capabilities requires auxiliary task demands [4]. In this context, the interaction between core knowledge and other factors (whether linguistic or otherwise) is inevitable, as LLMs inherently rely on language as a vehicle for expressing and processing information.

We hope this clarifies our rationale for using the VQA format. We will discuss this shortcoming in the limitations Section.

[1] Millière, Raphaël. "Language Models as Models of Language." arXiv, 2024, arXiv:2408.07144.
[2] Brown, Tom, et al. "Language models are few-shot learners." Advances in neural information processing systems 33 (2020): 1877-1901.
[3] Wei, Jason, et al. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.” Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, 2022. [4] Hu, Jennifer, and Michael C. Frank. "Auxiliary task demands mask the capabilities of smaller language models." arXiv, 2024, arXiv:2404.02418.

---

>>> Q6 How will the proposed dataset/benchmark be released, and how will the benchmark be maintained? I recommend doing something like a dataset card as the common practice in neurips benchmark tracks.
>>> A6 All data will be released upon acceptance and after internal review. Following standard procedures, we will open-source the dataset on GitHub and Hugging Face, including dataset cards, descriptions, and a viewer. Additionally, all model predictions and results will be made publicly available to ensure reproducibility.

>>> Q1 typo: double dots.
>>> A1 Thanks. We will revise and remove all typos.

---

>>> Q2 Assume that replicating human-like core knowledge is essential for the effective functioning of AI systems is controversial and may not necessarily hold?
>>> A2 Thank you for the question! We approach this from 2 complementary perspectives: 1) The question presumes that core knowledge is unique to humans or human-aligned AGI. While this is resonable, it's also possible that core knowledge reflects more general cognitive principles or serves as mechanistic prerequisite that may be useful—or even necessary—for any form of general intelligence; such as chimpanzees showing behaviors aligned with core knowledge (Regolin and Vallortigara,1995; Santos, 2004; Lake 2017). This suggests that certain cognitive abilities emerge across intelligent systems, regardless of the path. Then even non-human pathways to AGI, e.g. scaling, might still lead to the emergence of these core abilities. In this light, a benchmark on core knowledge, such as ours, could offer a useful lens for evaluating the progress toward AGI, regardless of the path taken. 2) Even if core knowledge is a human-specific trait, the human trajectory remains a valuable source of insight given the current absence of clearly defined path to AGI. Despite great advancements, MLLMs continue to struggle with hallucination, lack of OOD generalization and robustness, indicating that important cognitive capacities may still be missing. Humans as the only model of high-level intelligence can serve as a useful reference for evaluating artificial systems. Moreover, if AGI were to eventually follow a human-aligned path, our benchmark would become vital in assessing high-level reasoning and perception grounded in foundational cognitive structures rather than driven by spurious shortcuts.

---

>>> Q3 Assumption that core knowledge can be clearly defined and operationalized within the context of AI systems, but from pure evaluation based on QA answers, this could lead to benchmarks that do not accurately reflect the underlying theories or mechanisms of core knowledge.
>>> A3 While isolating core knowledge dimensions is notoriously difficult, particularly in AI, there are certain aspects that makes this effort feasible. Extensive literature suggests that core knowledge exists and can be robustly probed and evaluated in humans, including infants who don't yet possess language or the ability to speak, as in Appendix B. The taxonomy of 12 core-abilities addresses this challenge by operationalizing core knowledge through a set of lower-level abilities. These abilities serve as tractable proxies for otherwise abstract cognitive dimensions and also connect naturally to stage-based developmental theories (Piaget, 1976; Rochat, 2024). This makes it possible to evaluate how models handle fundamental cognitive operations and how such abilities may scaffold more complex reasoning.

---

>>> Q4 From pure evaluation based on QA answers, this could lead to benchmarks that do not accurately reflect the underlying theories or mechanisms of core knowledge.
>>> A4 We acknowledge that the VQA format is intertwined with confounding factors, such as language understanding (see Q5 of Reviewer nvFw). However, evaluating specific abilities in AI models inherently requires auxiliary task demands (Hu and Frank, 2024)—regardless VQA, retrieval, or output logits—and every evaluations and benchmarks face more or less a similar challenge. To mitigate biases introduced by the VQA format, our effort includes

• Instruct annotators to minimize the interplay of different abilities and exclude questions that require multiple competencies to answer.
• Manually filter out ambiguous or confusing questions and use LLMs to enhance clarity and precision.
• Extensive experiments with various phrasing and prompting techniques to alleviate biases arising from specific wordings or prompt formulations.

---

>>> Q5 Can this benchmark provide valuable insights into the training or designing of MLLMs? How the benchmark results can help researchers improve their models?
>>> A5 Our benchmark reveals that current training methods fail to effectively emerge these great abilities, suggesting future research into large-scale pretraining methods that can better scale core knowledge. More concretely, if core knowledge cannot be directly scaled, it may be beneficial to first teach or distill core knowledge into MLLMs prior to large-scale pretraining, enabling more data-efficient generalization akin to human learning. From an evaluation and design perspective, our benchmark uncovers distinct failure modes, including the lack of permanence, spatiality, boundary, and continuity. These limitations further hinder abilities such as visual perspective-taking and contribute to an over-reliance on shortcuts, which is a fundamental cause of poor out-of-distribution generalization.

>>> Q1 "high-level abilities do not correlate with the corresponding low-level abilities" is too strong
>>> A1 Thanks! In Sec 4.3, the correlations between lower- and higher-level abilities are generally below 0.4. This is considerably lower than what's commonly observed in humans, though the phrasing may be strong. We will revise it to better describe the observation.

>>> Q2 Scaling analysis: add numbers to make the plot more digestible. "low-level cognitive abilities is minimal...even detrimental" seems not perfectly supported.
>>> A2 Thanks for the advice. Our current efforts include 1) comparison of scalability (slope) in the bottom left of Fig. 6, and 2) $R^2$ values represented by the size of the blobs to the left of the ability names in the upper-left sub-fig of Fig. 6. We will also add specific numbers in the text to enhance clarity as suggested. We acknowledge that most abilities show some scalability; however, the scaling effect on lower-level abilities is significantly smaller (half the value) than that of higher-level abilities, which supports our conclusion. We will revise the text to better reflect the observation. Please see A4 below for a discussion on the perspective-taking exception.

>>> Q3 Suggested citation. A similar technique...for visual illusions in VLMs in [3].
>>> A3 Thanks for bringing this up. While [3] concurrently introduces a related idea, there are significant differences between the two. We offer a comprehensive controlled methodology grounded in the evaluation of core knowledge. Specifically, we formally propose a systematic procedure to manipulate input samples (${X}$) by flipping concept-level labels ($Y$) through targeted changes to causal features (${S}$) while holding non-causal features (${B}$) constant (Formally, a true predictive dist factorize as $p(Y|X) = \int p(Y| S, B) p(S, B| X)$). Whereas [3] presents only 10 samples of visual illusion (valuable but largely anecdotal), not situated within a broader diagnostic or theoretical framework with clear motivation and research questions. In addition, we present Figure 8, showing the development trend of current MLLMs (with respect to scaling) is not in the ideal direction but biased towards illusion or shortcut. This adds an interpretable, diagnostic dimension to our evaluation that is absent in [3].

>>> Q4 "Worse on lower-level abilities...there exist core knowledge deficits". Aren't all 12 abilities core abilities?
>>> A4 All 12 abilities in our benchmark are grounded in core knowledge dimensions. However, lower-level abilities are operational approximations of basic cognitive systems and are thus more directly aligned with the notion of "core knowledge", while higher-level abilities are more abstract or compositional cognitive tasks. The observed upward trend in performance does not imply that only lower-level abilities reflect core knowledge. Rather, it suggests that while models may perform better on higher-level tasks—potentially by pattern matching or spurious correlation—they often struggle with the more fundamental reasoning required for lower-level tasks. This gap is what we refer to as a core knowledge deficit: the failure to demonstrate a robust understanding of the foundational abilities that higher-level tasks presuppose.

>>> Q5 Why do perspective tasks show such poor external validity and do not scale?
>>> A5 Thanks for the question, which relates to a key finding here. The perspective-taking task in our benchmark is based on the "Three Mountains" experiment, a type of level-2 perspective-taking (Moll, 2010), which requires mental simulation—the ability to build an internal model of the world and reason from it (Johnson-Laird, 1982). Whether MLLMs possess such internal world models is debated (Mitchell, 2023; Goddu et al., 2024) and the lack of scalability in perspective suggests current models may not. Unlike other tasks, perspective-taking additionally demands constructing a spatial model, the absence of which could explain the unexpected downward trend we observe in performance as model size increases.

>>> Q6 A large number of models learn wrong visual intuitions?
>>> A6 Core-illusion refers to models' response driven by a natural perception of the world, i.e., a lack of core knowledge. Models that learn from statistical correlations in the data may fall short in acquiring core abilities, as MLLMs trained on vast, multimodal datasets are often biased by shortcut signals, producing answers that resemble advanced reasoning but lack the conceptual grounding that allows humans to apply knowledge flexibly and consistently across contexts (Mitchell, 2023). Due to limited space, kindly refer to Reviewer ZrNC's A2 for a technical explanation of why such a distinction can be rooted in the pretraining process.

We thank the reviewer for the valuable feedback. However, due to limited space, we could only answer the most significant questions here. We look forward to discussing the rest (i.e., illusion vs. shortcut and 4 quadrants of Fig 8, suggested citations, etc) in the discussion phase!

>>> Q1 Why this taxonomy was chosen, and if there are any specific inspirations from the ML literature e.g. OOD robustness, robotics?

>>> A1 We chose core knowledge as the basis for our taxonomy because it offers a theoretically grounded and developmentally validated account of the foundational systems underlying human cognition (Spelke & Kinzler, 2007). These dimensions are widely seen as essential to general intelligence, making them a meaningful lens for evaluating AI capabilities (Carey & Gelman, 1993; Carey, 2009; Spelke, 2022). While not directly inspired by robotics or OOD robustness, core knowledge is highly relevant to both, as generalization, transfer, and embodied reasoning rely on intuitive world understanding—precisely what core knowledge aims to capture. Thus, we provide a complementary perspective that can inform evaluations of generalization, transfer, and robustness in AI.

>>> Q2 The claim on "No observable scaling on low-level abilities..." Did the authors expect otherwise and why?...Do we a) expect and b) desire that our models follow the same curriculum as humans do? Will models follow the same path to AGI as we humans learn?...thinking in meta-terms about the overall capabilities, and how we scale these capabilities.

>>> A2 First, we do not expect otherwise as we anticipate challenges for the emergence of core-abilities simply through large-scale pretraining of statistical occurrence. Another potential reason is that, compared to high-level details required for complex tasks, core knowledge used in simpler tasks is spread across diverse contexts and the vast parameter space of the network (Shani et al., 2023), thus harder to isolate and apply systematically, leading to inconsistent or surface-level reasoning.

It's a great question whether AI should follow the path of humans. We elaborate on a discussion in A2 of Reviewer Kcuh (due to limited space), that our paper neither argues for nor against the necessity of mirroring humans in pursuit of AGI. Rather, core knowledge is introduced as an evaluation for MLLMs, and the core knowledge deficit hypothesis is proposed as an explanation for the observed brittleness of MLLMs. More broadly, we hypothesize that core knowledge may represent a general principle of intelligence--human or otherwise. The critical research question is not whether models should mimic human learning, but how core abilities can emerge through scaling with cognitive or human-inspired adaptation, or other measurements.

>>> Q3 'Cognitive instruction' seems to perform superiorly. This implies that certain cognitive priming to the models can help boost accuracy on the benchmarks. Any comments?

>>> A3 For now, we don't have a definitive explanation, but we find this effect aligned with early insights from the connectionist literature, which suggests that distributed representations pose a challenge for structured knowledge retrieval. As networks scale, retrieving specific conceptual structures becomes increasingly difficult (Hinton et al., 1986; Chalmers, 1990), especially for core knowledge. Unlike high-level knowledge, e.g., historical events, which are likely encoded in clustered patterns, core knowledge is highly distributed across the model’s parameters, as they are in multiple instances in training data, making them hard to isolate and deploy systematically for reasoning tasks (Garrigan, 2008; Green, 2024). We hypothesize that cognitive instruction may act as a retrieval cue that helps direct the model's internal attention toward latent but relevant knowledge. However, we do not believe this constitutes a permanent or scalable solution. In real-world environments, models are unlikely to receive such explicit guidance, limiting the practical utility of this strategy. Nevertheless, this result may point toward a promising research direction for improving reasoning via targeted scaffolding or memory-based mechanisms.

>>> Q4 Is the space of LLMs for spurious correlations any different from the space of spurious correlations for ResNets/ViTs, etc.? Can this benchmark be used to evaluate existing vision models? If not, what makes this eval benchmark specific to LLMs?
>>> A4 The space is different for LLM and vision models as the former is in an abstracted and discrete token space while the latter is high-dim pixel space. Our benchmark cannot be directly applied to vision models, since the questions are in VQA format (i.e., question answering). However, adaptation can be made to convert it into retrieval / binary classification for vision models to test. However, this falls out of the scope of this paper, and we leave it to future efforts.