Grounding Before Generalizing: How AI Differs from Humans in Causal Transfer

Abstract: Extracting abstract causal structures and applying them to novel situations is a hallmark of human intelligence. While LLMs and Vision LLMs (VLMs) have shown strong performance on a wide range of reasoning tasks, their capacity for interactive causal learning -- inducing latent structures through sequential exploration and transferring them across contexts -- remains uncharacterized. Human learners accomplish such transfer after minimal exposure, whereas classical Reinforcement Learning (RL) agents fail catastrophically. Whether state-of-the-art AI models possess human-like mechanisms for abstract causal structure transfer is an open question. Using the OpenLock paradigm requiring sequential discovery of Common Cause (CC) and Common Effect (CE) structures, here we show that models exhibit fundamentally delayed or absent transfer: even successful models require initial environmental-specific mapping -- what we term environmental grounding -- before efficiency gains emerge, whereas humans leverage prior structural knowledge from the very first solution attempt. In the text-only condition, models matched or exceeded human discovery efficiency. In contrast, visual information -- in both the image-only and text-and-image conditions -- overall degraded rather than enhanced performance, revealing a broad reliance on symbolic processing rather than integrated multimodal reasoning. Models further exhibited systematic CC/CE asymmetries absent in humans, suggesting heuristic biases rather than direction-neutral causal abstraction. These findings reveal that large-scale statistical learning does not produce the decontextualized causal schemas underpinning human analogical reasoning, establishing grounding-dependent transfer as a fundamental limitation of current LLMs and VLMs.

• The paper demonstrates that AI models require environmental grounding before transferring causal structures while humans excel at immediate structural transfer.
• It shows that LLMs perform well in text conditions but suffer performance degradation with multimodal inputs due to modality interference.
• The study highlights that models exhibit asymmetric causal reasoning between common cause and common effect, contrasting with human balanced abstraction.

Grounding Before Generalizing: Divergence of AI and Human Causal Transfer Mechanisms

Abstract and Motivation

The paper "Grounding Before Generalizing: How AI Differs from Humans in Causal Transfer" (2604.24062) systematically interrogates the causal abstraction and transfer capabilities of LLMs and VLMs using the OpenLock paradigm, contrasting their behavior with human performance in structurally invariant causal reasoning tasks. The study isolates the capacity for active, interactive causal learning and analogical transfer—hallmarks of human cognition—and delineates the operational gap between human learners and contemporary AI systems. The investigation is grounded in two core causal structures: common cause (CC) and common effect (CE), probed across multiple modalities (text-only, image-only, text-and-image).

Methodological Framework

The OpenLock paradigm, originally introduced to benchmark human and RL agent causal induction, is adapted here for LLMs and VLMs. Each environment consists of seven levers and a door, with variable superficial properties but invariant underlying causal topology (CC or CE). Models and humans must discover three unique solutions within 30 attempts, with each solution representing a successful trajectory through the latent causal graph.

Four leading models (GPT-5.2, Claude-4.5-Sonnet, Gemini-3-Flash, DeepSeek-V3.2) are evaluated under three conditions: text-only (T), image-only (I), and text-and-image (TI). Human baseline data is sourced from prior empirical studies for rigorous comparative analysis.

Results: Discovery Efficiency and Structural Asymmetry

In single-environment causal discovery:

• Text-Only Superiority: LLMs perform at or beyond human efficiency under the T condition. Gemini-3-Flash and GPT-5.2 showed marked reduction in attempts compared to humans, with Gemini-3-Flash achieving 100% success in both structures.
• Modality Degradation: Introducing visual data consistently impaired performance (notably GPT-5.2 and Claude-4.5-Sonnet). Image-only input led to the lowest success rates, indicating heavy reliance on symbolic rather than integrated multimodal reasoning.
• Causal Structure Bias: Models display pronounced asymmetries between CC and CE configurations. GPT-5.2 and Gemini-3-Flash favored CC; Claude-4.5-Sonnet favored CE; humans showed no significant directional preference, evidencing abstract causal schema formation.

Results: Transfer Dynamics and Grounding Dependency

Structure transfer is probed via provision of explicit solutions from prior environments with identical causal topology but different superficial attributes:

• Human Immediate Transfer: Humans demonstrate robust transfer, with a significant reduction in attempts for the first solution in novel environments, leveraging abstract causal structure from minimal exposure.
• Model Delayed Transfer: AI models uniformly require an initial environmental grounding (i.e., contextual mapping between abstract structure and situational tokens) before transfer efficiency emerges. No model showed decreased first-solution discovery cost in new environments, in stark contrast with humans.
• Retrospective vs Prospective Transfer: For models, structural abstraction operates retrospectively—activated only after direct interaction with a novel environment. Gemini-3-Flash exhibits statistically significant transfer, but gains are realized only after the first solution attempt, not before.

Theoretical Implications

The findings reveal three key qualitative distinctions:

1. Grounding-Dependent Transfer: Models lack decontextualized causal schemas necessary for prospective transfer; human analogical reasoning is guided by abstract relational patterns independent of surface features.
2. Absence of Insight-like Restructuring: Human learners exhibit non-linear acceleration post-initial discovery, consistent with representational insight and hypothesis space restructuring. Most LLMs optimize incrementally, suggestive of statistical refinement rather than logical schema revision.
3. Multimodal Interference: VLMs cannot suppress irrelevant perceptual features during symbolic abstraction, leading to performance degradation. Architectural heterogeneity (e.g., Claude-4.5-Sonnet’s resilience to modality changes) underscores critical differences in cross-modal attention control.

These results undermine the assumption that large-scale corpus-trained models inherently develop flexible, decontextualized causal representations. Reliance on surface-level statistical regularities and heuristic biases persists, as evidenced by CC/CE asymmetries and modal interference.

Practical Implications and Future Directions

The structural abstraction gap highlighted by the OpenLock paradigm has direct implications for AI systems targeting transfer learning in real-world, perceptually varied domains. Addressing this gap demands:

• Structure-Mapping Curricula: Explicit training on relational schema transfer across contexts to foster prospective abstraction.
• Architectural Cross-Modal Control: Enhanced mechanisms for cross-modal attention filtering, with symbolic reasoning prioritized and perceptual input regulated.
• Benchmarking Interactive Reasoning: Continued use of OpenLock and analogous paradigms for dissociating abstraction from context-bound pattern matching.

Further investigation into architectural factors underlying Claude-4.5-Sonnet’s insight-like learning dynamics is warranted, along with the exploration of prospective schema instantiation strategies for AI.

Conclusion

The work conclusively demonstrates that current LLM and VLM architectures fail to replicate the flexible and immediate causal transfer mechanisms of human cognition. Despite achieving local causal search efficiency, models operate in a grounding-dependent fashion and do not prospectively instantiate abstract relational schemas. The abstraction gap is qualitative, not resolved by scale or multimodal fusion, and must be addressed via targeted curriculum design and architectural innovation to approach human-like analogical reasoning and causal transfer.