At the core of this research is a one-shot learning task designed to challenge participants and LLMs, such as ChatGPT, to discern and manipulate sentence structure. A total of 372 individuals, comprising native speakers of English and Chinese along with bilingual participants, engaged in the task alongside high-performing language models. The objective was to test their ability to predict which words from sentences could be deleted without compromising the overall meaning, a task that requires an intrinsic grasp of syntactic relationships and hierarchical sentence organization.

The findings reveal an interesting trend where both groups exhibit a tendency to delete certain constituents—grammatical units that contribute to the overall structure and meaning of sentences—while avoiding the removal of non-constituent strings of words. This behavior highlights a fundamental aspect of linguistic understanding that transcends mere reliance on word frequency or positional information within a sentence. Instead, both humans and LLMs seem to leverage an innate grasp of tree-like structures that underpin language.

Through careful analysis of which words were discarded in the task, researchers were able to reconstruct the underlying constituency tree structures that both groups relied upon. This not only speaks to the sophistication of human cognitive processes but also suggests that large language models are capable of encoding and utilizing similar tree-structured representations in their sentence understanding and generation. This convergence raises critical questions about the nature of language processing in humans versus artificial intelligence and challenges pre-existing models that fail to account for such complexities.

The implications of these findings extend beyond mere academic curiosity. By establishing that constituency tree structures are actively employed by both humans and LLMs, this research paves the way for more nuanced approaches in the development of AI systems that align more closely with human-like language processing. It also invites further exploration into how language is fundamentally structured in the brain, potentially offering insights that could enhance educational methodologies and AI training protocols.

Moreover, the research raises ethical considerations surrounding the use and understanding of AI in linguistic tasks. As LLMs become increasingly integral to our daily lives, an understanding of how they mimic human-like processing can inform safeguards against miscommunication and ensure that AI systems serve to complement human capabilities rather than replace them. This research thus acts as a critical touchstone, illuminating the intricate connections between human cognition and machine learning.

As researchers continue to probe the depths of language processing, the parallels drawn between human cognition and the abilities of LLMs present exciting possibilities for future AI developments. There remains much to explore in the realms of syntax, semantics, and the neural underpinnings of language that could guide both cognitive science research and advancements in AI technology. Understanding these relationships within structural frameworks may redefine our approach to language acquisition and the deployment of AI in communication settings.

Ultimately, this pioneering study opens a host of new avenues for inquiry within linguistics, cognitive science, and artificial intelligence. By demonstrating that both humans and LLMs generate and interpret language rooted in complex tree structures, researchers have taken an essential step toward unraveling the mysteries of linguistic representation. As such insights continue to build and evolve, they promise to deepen our comprehension of not only how language operates but also how emerging technologies can effectively bridge the gap between human communication and machine understanding.

Continued exploration in this domain could yield important innovations in educational technologies, language processing applications, and AI-enhanced communication platforms. The potential to develop systems that rival human linguistic abilities is within reach, suggesting a future where seamless interaction between humans and machines becomes a natural occurrence.

The convergence of human and AI capabilities indicates not only a fascinating scientific endeavor but also a significant leap toward integrating AI into everyday tasks, enriching communication, and fostering deeper understanding across diverse languages and cultures. This study exemplifies how interdisciplinary research can create synergies between cognitive science and artificial intelligence, leading to groundbreaking revelations about language representation and processing.

In conclusion, the research by Liu, Xiang, and Ding forms a compelling narrative about the shared linguistic capabilities of humans and LLMs, urging further examination of tree-structured representations. This compelling evidence establishes a robust framework that enhances both theoretical and applied linguistics, revolutionizing how we comprehend and harness the power of language across varying modalities. The journey to fully understand the complexities of human and machine language processes is far from over, but this foundational work sets a powerful precedent for future discoveries.

Subject of Research: Tree-structured sentence representations in humans and large language models.

Article Title: Active use of latent tree-structured sentence representation in humans and large language models.

Article References:

Liu, W., Xiang, M. & Ding, N. Active use of latent tree-structured sentence representation in humans and large language models. Nat Hum Behav (2025). https://doi.org/10.1038/s41562-025-02297-0

Image Credits: AI Generated

DOI: 10.1038/s41562-025-02297-0

Keywords: language processing, large language models, cognitive science, sentence structure, constituency trees, artificial intelligence.

The study, spearheaded by postdoctoral research associate Jake Russin, aims to articulate the nuanced interplay between two critical learning modes: flexible, context-driven learning and slower, incremental learning. These two approaches mirror the functions of working memory and long-term memory found within the human brain. This discovery not only enhances our understanding of human cognitive strategies but also informs the design of more intuitive artificial intelligence systems.

A deeper exploration reveals that humans often employ two distinct strategies when acquiring new knowledge. For instance, when learning the simple rules of games such as tic-tac-toe, individuals typically engage in “in-context” learning. This method allows for rapid understanding of rule structures through minimal examples. Conversely, incremental learning requires time and repetition, akin to the extensive practice involved in mastering an instrument, such as playing a song on the piano. Researchers had long acknowledged the coexistence of these two learning styles in both humans and AI, yet the interaction between them had remained somewhat elusive.

The research team’s innovative approach revealed that the relationship between these methodologies can be likened to the dynamics of working and long-term memory in humans. To put this theory to the test, Russin employed a cutting-edge technique known as meta-learning. This type of training is designed to enable AI systems to learn about their own learning processes. Through various experiments, the team discerned critical attributes of both learning styles, significantly advancing the understanding of how they operate in tandem.

One particularly illuminating experiment involved assessing the AI’s ability to engage in in-context learning. By tasking the AI with recombining familiar concepts—such as colors and animals—in novel scenarios, researchers examined whether it could successfully identify new combinations, like a green giraffe, that had not been previously encountered. Through rigorous testing involving around 12,000 similar tasks, the AI demonstrated a remarkable capacity for recognizing and identifying new pairings, showcasing an important facet of its learning capability.

The findings from this study indicate that both humans and AI experience enhanced flexibility in in-context learning as a direct result of previous incremental learning experiences. This phenomenon mirrors human learning processes, where repeated exposure to a variety of situations cultivates a more agile cognitive response. Consequently, just as players may grasp new board game rules more quickly after mastering numerous past games, both AI and humans can learn more effectively through cumulative experiences.

Moreover, the researchers uncovered intriguing trade-offs in this learning dynamic. They noted parallels between AI’s learning retention capabilities and those observed in humans. Interestingly, the more difficult the task was for the AI, the stronger its retention of the learned information. This observation aligns with the notion that errors during learning prompt cognitive systems, be they human or AI, to update their long-term memory structures more effectively. In contrast, actions performed error-free—while contributing to flexibility—do not engage long-term memory in a similar manner.

As Michael Frank, a leading scholar in computational neuroscience and one of the research team members, explains, these revelations illustrate how insights gained from analyzing learning strategies in artificial neural networks can enrich our understanding of the complexities of human cognition. The synthesis of these insights presents a more unified perspective on human learning, unveiling connections that had previously been overlooked within the field.

The implications of this research extend far beyond academic curiosity; they provide critical considerations for the ongoing development of AI technologies. With the acceleration of AI integration into sensitive fields such as mental health, ensuring that these systems are intuitive and trustworthy is of paramount importance. The researchers emphasize the necessity for both human and AI systems to possess a mutual awareness of their respective cognitive processes and to recognize the similarities and differences that exist in those processes.

This research effort was made possible through support from the Office of Naval Research and the National Institute of General Medical Sciences, highlighting the collaborative nature of scientific advancement. By intertwining the expertise of cognitive science and artificial intelligence, this study represents a pivotal step toward forging connections between human and AI learning paradigms.

In conclusion, the remarkable findings from Brown University not only shed light on the intricacies of cognition but also hint at a future where AI can interact with human systems in a more understanding and context-sensitive manner. The potential to harness these insights for developing advanced AI tools opens up new realms of possibility, underscoring the importance of continued exploration at the intersection of human cognition and artificial intelligence.

Subject of Research: The interplay between in-context and incremental learning in humans and AI.
Article Title: Parallel trade-offs in human cognition and neural networks: The dynamic interplay between in-context and in-weight learning.
News Publication Date: 28-Aug-2025
Web References: Proceedings of the National Academy of Sciences
References: 10.1073/pnas.2510270122
Image Credits: Brown University

Keywords

Artificial intelligence, Cognition, Neuroscience

The study, spearheaded by cognitive science expert Leyla Isik and doctoral candidate Kathy Garcia, highlights the limitations of current deep learning models in decoding social dynamics essential to real-world interactions. As autonomous systems increasingly intertwine with everyday life, the capacity to discern intentions, goals, and interpersonal context becomes paramount. Conventional AI architectures, primarily modeled on brain regions adept at processing static images, appear ill-equipped to capture the fluid and multifaceted nature of social scenes.

Central to the investigation were brief, three-second video clips depicting varied social scenarios: individuals engaged in direct interaction, participants performing parallel but independent tasks, and solitary actors disconnected from social exchanges. Human observers consistently rated these clips with high inter-rater agreement across features crucial for social comprehension. In stark contrast, a diverse collection of over 350 AI models, encompassing language, video, and image processing systems, failed to emulate human judgment or brain activity patterns when tasked with interpreting these scenes.

Intriguingly, large language models demonstrated relatively better alignment with human evaluations when analyzing concise, human-generated captions describing the video content. This contrasts with video-based AI systems, which struggled not only to accurately describe actions but also to predict the corresponding neural responses in human observers. Image models, supplied with static frames extracted from the videos, proved inadequate in reliably identifying communicative exchanges or the intent behind the observed behaviors.

This discrepancy between static and dynamic scene processing underscores a foundational challenge in AI development. Although recognizing objects and faces in still images has long been achievable with increasing precision, the temporal complexity of social interactions demands sophisticated integration of spatial and contextual information over time. Humans effortlessly parse subtle cues such as gaze direction, body language, and proxemics to infer underlying intentions—a level of cognitive acuity absent in current AI paradigms.

Isik and Garcia suggest that this shortfall stems from the architectural inheritance embedded within AI neural networks, which largely emulate the ventral visual stream responsible for static image analysis in the human brain. By contrast, dynamic social vision recruits distinct neural circuits, including those involved in social cognition and real-time scene interpretation. The evident “blind spot” in AI implies that future models must extend beyond traditional frameworks to incorporate mechanisms for representing and reasoning about ongoing social processes.

The ramifications of this research extend deeply into the domain of autonomous systems. For example, self-driving cars navigating urban environments must anticipate the trajectories and behaviors of pedestrians and other drivers, discerning whether individuals are about to cross the street or engaged in social interaction. Failure to accurately interpret these cues could compromise safety and efficiency. Similarly, assistive robots designed to aid elderly or disabled individuals rely on nuanced social understanding to respond appropriately and empathetically.

Moreover, the study’s findings call into question the prevailing reliance on static datasets and benchmarks in AI training. Dynamic social scenarios present challenges in variability, ambiguity, and the necessity for contextual reasoning that static images cannot capture. Advancing AI to human-comparable levels of social comprehension will likely require novel training paradigms, hybrid model architectures, and integration of multi-modal sensory data reflective of real-world complexity.

This research also invites broader reflections on the relationship between biological and artificial intelligence. The human brain seamlessly integrates perceptual input with memory, emotion, and learned social norms to construct rich, dynamic interpretations of the environment. Replicating even a fraction of this capacity demands interdisciplinary collaboration, drawing insights from neuroscience, cognitive science, computer vision, and machine learning.

In conclusion, while AI has excelled in recognizing and categorizing static visual information, the frontier of dynamic social vision remains elusive. Bridging this gap is critical not only for enhancing machine perception but also for ensuring that emerging technologies harmonize safely and intuitively with human social environments. Johns Hopkins University’s pioneering work lays bare the current limitations and charts a course toward more socially intelligent AI systems, emphasizing that understanding human behavior in motion is a challenge yet to be fully met by deep learning.

Subject of Research:
Understanding the limitations of current AI models in interpreting dynamic social interactions and the gaps between human and AI social vision.

Article Title:
MODELING DYNAMIC SOCIAL VISION HIGHLIGHTS GAPS BETWEEN DEEP LEARNING AND HUMANS

News Publication Date:
24-Apr-2025

Web References:
https://cogsci.jhu.edu/directory/leyla-isik/
https://iclr.cc/

Keywords:
Artificial intelligence, Neural networks, Image processing