PaperlayerView exact revision on arXiv

Source-linked AI summary

Core Knowledge Deficits in Multi-Modal Language Models

Yijiang Li, Qingying Gao, Tianwei Zhao, Bingyang Wang, Haoran Sun, Haiyun Lyu, Robert D. Hawkins, Nuno Vasconcelos, Tal Golan, Dezhi Luo, Hokin Deng

arXiv:2410.10855v4cs.CLcs.AIcs.CV

TL;DR

MLLMs show strengths in high-level perception and reasoning yet remain weak on human-intuitive abilities, motivating investigation of core knowledge. The paper introduces CoreCognition and Concept Hacking; its findings indicate persistent low-level deficits and shortcut reliance rather than genuine progress with scale.

  • Problem

    MLLMs can underperform humans on intuitive tasks such as counting, spatial reasoning, and compositional reasoning despite high-level strengths.

    Source: §1 · pp. 1–2

  • Method

    The authors introduce CoreCognition, a 12-concept benchmark, evaluate models across prompts, and use Concept Hacking to test for shortcut reliance.

    Source: §1 · p. 2, §7 · p. 11

  • Results

    MLLMs show worse performance on low-level than high-level abilities, misaligned ability dependencies, and reduced or absent low-level scalability, while scaling models rely on shortcuts rather than progressing toward conceptual understanding.

    Source: §1 · p. 3, §6 · p. 10

  • Takeaways & Limitations

    The findings support the authors’ conclusion that core abilities grounding high-level perception and reasoning cannot be acquired through scaling alone.

    Source: §6 · p. 10

    CoreCognition’s VQA format retains auxiliary demands and limits evaluation to models capable of linguistic processing, while shortcuts may inflate performance.

    Source: §7 · p. 11

Abstract

from arXiv · show

While Multi-modal Large Language Models (MLLMs) demonstrate impressive abilities over high-level perception and reasoning, their robustness in the wild remains limited, often falling short on tasks that are intuitive and effortless for humans. We examine the hypothesis that these deficiencies stem from the absence of core knowledge--rudimentary cognitive abilities innate to humans from early childhood. To explore the core knowledge representation in MLLMs, we introduce CoreCognition, a large-scale benchmark encompassing 12 core knowledge concepts grounded in developmental cognitive science. We evaluate 230 models with 11 different prompts, leading to a total of 2,530 data points for analysis. Our experiments uncover four key findings, collectively demonstrating core knowledge deficits in MLLMs: they consistently underperform and show reduced, or even absent, scalability on low-level abilities relative to high-level ones. Finally, we propose Concept Hacking, a novel controlled evaluation method that reveals MLLMs fail to progress toward genuine core knowledge understanding, but instead rely on shortcut learning as they scale.

1. Introduction

The paper asks whether MLLM weaknesses on human-intuitive tasks reflect missing core knowledge, and introduces CoreCognition to examine that hypothesis. Its evaluations report persistent deficits in low-level abilities relative to high-level abilities.

1. Introduction

  • MLLMs underperform humans on simple tasks such as counting, perspective-taking, spatial, temporal, and compositional reasoning, despite strengths in related high-level reasoning.The paper also notes that small changes in task conditions can cause substantial performance drops in real-world or out-of-distribution settings.Source: §1 · pp. 1–2
  • CoreCognition operationalizes 12 abilities grounded in developmental cognitive science and contains 1,503 samples, with over 95 examples for each concept.The benchmark is intended to examine the existence, representation, and use of core knowledge in MLLMs.Source: §1 · p. 2
  • The evaluation covers 230 models and 11 prompting techniques, yielding 2,530 data points for analyses of ability levels, dependencies, predictive power, and scaling.The authors also manipulate causal image features to test whether models use the targeted knowledge or approximate it through shortcuts and spurious correlations.Source: §1 · pp. 2–3
  • The reported findings include worse low-level performance, misaligned dependencies between high- and low-level abilities, and weaker or absent scaling on low-level abilities.The authors further report that larger models show core deficits and shortcut-taking rather than progress toward conceptual understanding.Source: §1 · p. 3

2. Related Works

Related work spans MLLM development and benchmarks, shortcut learning, and developmental accounts of human core knowledge. CoreCognition is positioned as a multimodal benchmark for early-emerging abilities that support higher-level reasoning.

Benchmarks for Multi-modal Large Language Model

  • Prior MLLM benchmarks cover tasks including VQA, captioning, OCR, and higher-level reasoning; M3GIA and Marvel address cognitive complexity but emphasize task coverage or general intelligence.CoreCognition instead targets early-emerging core abilities that support higher-level perception and reasoning.Source: §2 · p. 3
  • DevBench uses a developmentally inspired framework to study language-learning trajectories, whereas CoreCognition targets multimodal core knowledge.Source: §2 · p. 3

Shortcut Learning

  • Prior shortcut-learning methods remove biased features, train bias predictors, minimize feature-bias mutual information, or generate bias-conflicting samples.Source: §2 · p. 3

Core Knowledge in Humans

  • Developmental research describes rudimentary infant knowledge of objects, actions, numbers, space, and social relations as supporting early learning and later complex reasoning.The passage frames core knowledge as a foundation for reasoning in variable environments.Source: §2 · p. 3

3. Benchmarking Core Knowledge in Multi-modal Large Language Models

CoreCognition translates a developmental account of core abilities into a curated multimodal benchmark and evaluates MLLMs using scalable inference and position-robust scoring.

3. Benchmarking Core Knowledge in Multi-modal Large Language Models

  • CoreCognition comprises 12 core abilities and 1,503 questions with diverse input types and formats, organized around a developmental cognitive taxonomy.The section outlines taxonomy, curation, and model inference and evaluation.Source: §3 · p. 3

3.1. Cognitive Framework

  • The taxonomy follows developmental stages from sensorimotor concepts through symbolic representation and systematic reasoning to abstract reasoning.Abilities described include object permanence, perceptual constancy, perspective-taking, conservation, intuitive physics, hierarchical relations, intentionality, and mechanical reasoning.Source: §3.1 · p. 4

3.2. Dataset Curation

  • Curation operationalizes theoretical constructs as explicit examples, aiming for discriminative items with minimal confounding and minimal text-only shortcuts.The process prototypes each ability, instantiates prototypes as diverse image or video assets paired with questions, options, and ground-truth answers, then cross-validates QA pairs and checks them with human annotators.Source: §3.2 · p. 4

3.3. Inference and Evaluation Strategy

  • Circular evaluation rotates each question’s answer options through all positions and averages correctness to mitigate option-position bias.Free-form responses are mapped to choices using templates and, when needed, an LLM judge; FAIL responses count as incorrect, and models with high FAIL rates are re-evaluated and may be excluded.Source: §3.3 · p. 5

4. Experiments

Across 230 diverse MLLMs, the experiments examine core abilities, their relationships to higher-level performance, scaling, reasoning models, and prompting. Results reveal persistent weaknesses in low-level abilities despite stronger high-level performance.

4.1. Core knowledge Deficits

  • MLLMs perform significantly better on higher-level abilities but struggle with lower-level abilities associated with early developmental stages.The comparison is statistically significant; selected models also substantially trail humans on lower-level stages.Source: Figure 4 · p. 5, §4.1 · pp. 5–6, Table 2 · p. 6

4.2. Dependencies Between Core Abilities

  • Many ability correlations are modest (ρ < 0.4), while some clusters are strong (ρ > 0.65), with weak cross-stage links among foundational abilities.High-level abilities correlate robustly within their stage; Hierarchy clusters more closely with Formal Operational abilities.Source: Figure 5 · p. 6, §4.2 · p. 7

4.3. Core Abilities are Predictive of Higher-level Abilities

  • Core abilities strongly predict most public-benchmark and SEED-Bench performance, except for Perspective, Intuitive Physics, and ChartQA.The authors hypothesize that ChartQA's textual understanding is largely orthogonal to the core abilities examined.Source: Figure 6 · p. 7, §4.3 · p. 7

4.4. Scaling Effect on Core Knowledge?

  • Seven of nine low-level abilities improve significantly less with model size than Formal Operational abilities, while Perspective-taking declines with scale.Hierarchical relation and perceptual constancy are the two low-level exceptions; the authors link the decline in perspective-taking to intensifying egocentric bias.Source: §4.4 · pp. 7–8, Figure 7 · p. 8

4.5. Does Reasoning Help?

  • Reasoning and test-time scaling show no significant difference in 10 of 12 core abilities, with inconsistent differences in the other two.The two exceptions are perceptual constancy (reasoning models perform better, P = 0.0669) and perspective-taking (worse, P = 0.0037).Source: §4.5 · p. 8

4.6. Performance with Different Prompts

  • Most prompt variants fail to improve performance, while cognitive instruction increases accuracy by over 6%.The authors caution that explicit conceptual scaffolding is unlikely in real-world scenarios and is not a permanent or scalable solution.Source: §4.6 · pp. 8–9, Table 3 · p. 9

5. Concept Hacking: A Controlled Experiment

The paper tests whether MLLMs demonstrate core knowledge or rely on shortcuts by pairing original tasks with versions whose correct answers are inverted. The results distinguish core deficits from shortcut-taking and show that larger models do not necessarily progress toward conceptual understanding.

5. Concept Hacking: A Controlled Experiment

  • The experiment is designed to distinguish apparent task proficiency from genuine conceptual understanding.The authors frame shortcut learning as a risk because models may exploit spurious features.Source: §5 · p. 9, §5.1 · p. 9

5.1. Methodology

  • Concept hacking manipulates task-relevant features while preserving irrelevant conditions, creating paired tasks with opposite correct answers.Correct responses on both indicate core knowledge; success only on the original task followed by failure on its inverted version indicates shortcut-taking.Source: §5.1 · pp. 9–10

5.2. Results: Core Deficits v.s. Shortcut Taking

  • Many models show core deficits or shortcut-taking, and almost none achieve roughly equal accuracy on both original and manipulated tasks.Some models score highly on manipulation but below chance on control tasks; others show the reverse, and both patterns occur across model sizes and performance levels.Source: §5.2 · p. 10

6. Discussion

The findings support the hypothesis that MLLMs lack core knowledge grounding higher-level abilities and that scaling alone does not supply it. The discussion connects these deficits to shortcut reliance, robustness, and how progress toward AGI might be evaluated.

6. Discussion

  • The findings support the hypothesis that MLLMs lack core knowledge grounding higher-level perception and reasoning, and that scaling alone cannot acquire these abilities.The authors connect this to the persistent difficulty machines have with tasks intuitive to humans.Source: §6 · p. 10
  • As model size increases, core deficits and shortcut reliance intensify rather than approaching the human-like core-knowledge region.Figure 11 encodes performance with circles, parameter size with circle size, and average CoreCognition accuracy with color; the green triangle marks human performance and the red dot marks chance.Source: Figure 11 · p. 11
  • The authors argue that explicit foundational-capacity training may be needed because current training paradigms do not instill core cognitive abilities.They suggest teaching or distilling core knowledge before large-scale pretraining as a possible approach.Source: §6 · p. 11
  • Core knowledge could offer a lens for evaluating AGI across different developmental paths, including non-human routes such as scaling.The authors frame this as conditional on the possibility that core knowledge reflects learning principles shared across intelligent agents.Source: §6 · p. 11
  • The authors present core-knowledge deficits as a plausible account of current MLLM weaknesses in hallucination, generalization, and robustness.They connect cognitively grounded evaluation to exposing gaps in systems that still show these weaknesses.Source: §6 · p. 11

7. Conclusion and Limitations

The paper introduces CoreCognition with Concept Hacking to evaluate MLLMs’ understanding of core knowledge. It reports persistent deficits and notes limits from VQA’s auxiliary demands and Concept Hacking’s scalability.

7. Conclusion and Limitations

  • CoreCognition and Concept Hacking are introduced to evaluate MLLMs’ conceptual understanding of core knowledge, and the authors report four findings indicating a consistent lack of it.The core concepts include basic world knowledge about objects, actions, numbers, space, and social relations.Source: §7 · p. 11
  • VQA adds linguistic, counting, and object-recognition demands, restricts evaluation to models with linguistic processing, and Concept Hacking’s meticulousness limits large-scale use.Filtering and validation across 11 prompts mitigate some auxiliary factors but do not eliminate them.Source: §7 · p. 11

Impact Statement

The impact statement frames CoreCognition as a cognitively grounded evaluation framework. The accompanying sections describe human cognitive development, define benchmark abilities, explain input formats, and justify the benchmark’s difficulty for MLLMs.

Impact Statement

  • CoreCognition is presented as a framework for evaluating MLLMs and informing the design of more robust, interpretable models with grounded reasoning and perception.The authors also note that adapting Concept Hacking’s adversary idea for military purposes may pose misuse risks.Source: Impact Statement · p. 11

A.1. Core Knowledge in Human

  • Human cognition develops from rudimentary abilities in object, number, space, action, and social domains toward more complex reasoning, with early abilities supporting later ones.The section describes sensorimotor capacities such as permanence and spatiality, later mental operations, and formal reasoning and social cognition.Source: §A.1 · pp. 22–23

A.2. Definition of the 12 Core abilities in CoreCognition

  • The benchmark’s ability definitions include object boundaries, spatial relations, perceptual constancy, permanence, continuity, and conservation across transformations.For example, conservation concerns properties such as quantity remaining unchanged when an object’s arrangement or shape changes.Source: §A.2 · p. 23
  • Other defined abilities cover perspective-taking, hierarchical relations, intention understanding, mechanical reasoning, and tool use toward goals.The definitions describe both social interpretation and reasoning about or acting on physical systems and objects.Source: §A.2 · pp. 23–24

B. Input Types and Formats

  • CoreCognition uses true/false and multiple-choice questions with single-image, video, and multiple-image inputs to accommodate differing ability tasks and model support.The benchmark includes varied visual formats because some abilities require images while others require videos; not all MLLMs accept every format.Source: B. Input Types and Formats · p. 24, Figure 12 · p. 24

C. Justification for the Difficulty in CoreCognition

  • Tasks expand established developmental-science prototypes into systematically varied question suites and are made challenging to probe MLLMs rather than reproduce child-level diagnostics.The authors justify this difficulty by noting that evaluated MLLMs may have knowledge and reasoning abilities comparable to or beyond educated adults.Source: C. Justification for the Difficulty in… · p. 25

D. Inference and Evaluation

The evaluation covers models with diverse sizes and input capabilities, then uses hybrid answer matching, filters high-failure models, and scores multiple-choice answers across rotated option orders.

D.1. Model Inference

  • The evaluation includes 231 commercial and open-source models ranging from 1B to 110B parameters, grouped by single-image, multi-image, or video input support.The groups contain 85 single-image, 105 multi-image, and 41 video models.Source: §D.1 · p. 25

D.2.1. MATCHING ANSWERS TO CHOICES

  • Hybrid Matching first applies templates, then accepts an LLM-ensemble extraction only when at least three of four models agree.Exact and “in” matching had high fail rates; template matching required iterative adaptation, while LLM matching handled text-rich outputs but could hallucinate in long contexts.Source: §D.2.1 · pp. 25–26

D.2.2. FILTERING

  • After manual review of models with matching failure rates of at least 17%, the authors excluded models at or above a 20% fail rate, removing 12 and retaining 219.The filtering distinguishes systematic or detrimental failures from intrinsic model limitations.Source: §D.2.2 · p. 26, Figure 13 · p. 26

D.2.3. SCORING

  • Circular evaluation rotates the answer order across all k options and scores the proportion of rotations answered correctly, reducing option-position bias without requiring perfect consistency.Responses that fail answer matching are classified as incorrect.Source: §D.2.3 · pp. 26–27

E. Normalization in Concept-wise Comparison

The section reports raw and chance-normalized accuracies, which show a similar upward trend supporting the core knowledge deficit finding. Paired t-tests also find statistically significant performance differences between the Formal Operational stage and the other two stages.

E. Normalization in Concept-wise Comparison

F. Details on Statistical Testing

  • p ≪ 0.001 indicates statistically significant performance differences between the Formal Operational stage and the other two stages.The paired t-test statistic uses the mean difference between paired observations, its standard deviation, and the number of pairs.Source: F. Details on Statistical Testing · p. 27, Table 6 · p. 27

G. Core Knowledge Deficits Under Different Conditions

The core knowledge deficit finding persists across input formats, prompt types, and model selections. Lower-level tasks continue to show lower accuracy, although widening the model selection increases variability and slightly lowers median accuracy, particularly on higher-level tasks.

G. Core Knowledge Deficits Under Different Conditions

H. Dependencies Between Core Abilities Under Different Conditions

Inter-concept dependency patterns generally persist across input settings, prompt types, and model subgroups. However, several concepts remain weakly correlated with other abilities, and the single-image Permanence assessment is unrepresentative because it includes fewer than 10 questions.

H. Dependencies Between Core Abilities Under Different Conditions

I. Scaling Effect on Core Knowledge Under Different Conditions

The scaling effect is similar across model subgroups defined by input constraints, while further partitions are not analyzed because of the study design and regression data requirements. The additional Concept Hacking example and qualitative examination describe shortcut reliance, including GPT-4o reasoning that is not grounded in manipulated image content.

I. Scaling Effect on Core Knowledge Under Different Conditions

  • All three input-constraint subgroups show a scaling effect aligned with the study’s earlier observations.Further scaling partitions by competence and prompt type were not analyzed: the analysis studies model groups, and single-prompt data would not meet stable regression-fit requirements.Source: §I · p. 30, Figure 21 · p. 30

K. Explanation of Concept Hacking with an Example

  • In the perceptual-constancy example, the bridge is manipulated to taper while task-irrelevant details remain identical to the standard task and control.A model with perceptual constancy should answer both correctly; reliance on spurious correlations may produce success on the original task but failure on the manipulation.Source: K. Explanation of Concept Hacking with… · pp. 31–32

L. Qualitative examination of Concept Hacking

  • GPT-4o’s low performance on manipulation samples primarily stems from shortcut reliance, possibly learned during training, in the qualitative examination.For the manipulated bridge, it repeats an explanation of the visual illusion despite the bridge actually decreasing in width, indicating reasoning not grounded in the image.Source: Figure 22 · p. 31, L. Qualitative examination of Concept… · p. 32

Related papers

Loading 2410.10855v4…