# Linguistic relativity in large language models

## Theoretical Foundations of Linguistic Relativity

The inquiry into whether the structure of a language determines the boundaries of human thought has been a central preoccupation of cognitive science, linguistics, and philosophy for over a century. Formally articulated as the Sapir-Whorf hypothesis, the theory of linguistic relativity posits that the semantic and structural parameters of a specific language shape its speakers' perception, memory, and cognitive categorization [cite: 1, 2, 3]. Early interpretations of this hypothesis proposed a strong form known as linguistic determinism, which argued that cognitive categories are entirely dictated by linguistic constraints; if a language lacks a specific word or grammatical structure, its speakers are theoretically unable to conceptualize the corresponding phenomenon [cite: 3]. 

Modern cognitive science has largely rejected strict linguistic determinism in favor of a weaker, yet empirically robust, formulation of linguistic relativity. This framework suggests that language operates as a cognitive guide, selectively directing attention, shaping habitual thought patterns, and structuring abstract reasoning through conventionalized metaphors [cite: 2, 3, 4]. Empirical evidence for this phenomenon spans numerous cognitive domains. For instance, color perception is measurably influenced by linguistic categorization. Speakers of the Himba language in Namibia, which groups blue and green under a single lexical term while differentiating multiple shades of green that English conflates, exhibit distinct reaction times in color discrimination tasks. Himba speakers process distinctions encoded in their language significantly faster than English speakers, demonstrating that linguistic categories alter perception prior to conscious judgment [cite: 2, 5].

Spatial orientation provides another profound example of linguistic relativity. While English speakers typically rely on egocentric coordinates (e.g., "left," "right," "front," "back"), speakers of Guugu Yimithirr in northern Australia utilize an absolute frame of reference based on cardinal directions (e.g., "north," "south," "east," "west") [cite: 2]. Because direction is absolute in their linguistic framework, speakers maintain a constant, internalized compass, fundamentally altering their spatial memory and environmental awareness compared to speakers of egocentric languages [cite: 2].

The rapid advancement of Large Language Models (LLMs) has introduced an unprecedented paradigm for testing the Sapir-Whorf hypothesis. As statistical architectures trained exclusively on vast, multilingual text corpora, LLMs encapsulate the syntactic, semantic, and cultural geometries of human communication [cite: 6, 7, 8]. Lacking physical bodies, sensorimotor experiences, or innate psychological drives, the entirety of an LLM's "worldview" is constructed through the statistical distribution of language [cite: 6, 7]. Consequently, these models serve as profound testbeds for artificial linguistic relativity. If language actively shapes thought, a system composed entirely of language should logically inherit and exhibit the cognitive biases, spatial metaphors, and reasoning frameworks embedded within its training data [cite: 9, 10]. Evaluating this phenomenon, however, requires confronting the foundational limitations of artificial cognition, particularly the paradox of extracting meaning from purely syntactic manipulations.

## Temporal Metaphors and Spatial Orientations

Because time is an invisible, intangible, and highly abstract domain, human cognition universally co-opts the domain of physical space to conceptualize, sequence, and communicate temporal events [cite: 1, 11]. According to Conceptual Metaphor Theory, the neural mechanisms originally evolved for physical navigation and spatial memory are repurposed to structure abstract reasoning [cite: 11, 12, 13]. However, while the reliance on spatial metaphors for time is a biological universal, the specific spatial axes and directional mappings utilized are culturally and linguistically arbitrary [cite: 14, 15].

### Horizontal and Vertical Axes of Time

In Western cultures, time is predominantly conceptualized along a horizontal axis. English speakers utilize sagittal (front-to-back) or transverse (left-to-right) physical planes to describe time, evidenced by phrases such as "looking forward to the future" or "leaving the past behind" [cite: 1, 2]. This horizontal orientation is often reinforced by reading and writing directionality. English speakers spontaneously arrange temporal sequences from left to right, whereas speakers of Arabic or Hebrew frequently map time from right to left [cite: 11].

Mandarin Chinese introduces a multidimensional spatial topology. While Mandarin speakers use horizontal metaphors similar to English speakers, they possess a robust secondary system of vertical metaphors. In Mandarin, earlier events are habitually described as "up" (shàng) and later events as "down" (xià) [cite: 1, 3, 5]. Experimental studies utilizing line-bisection paradigms and array recognition tasks demonstrate that these linguistic differences yield measurable cognitive consequences. Mandarin speakers recognize temporal sequences faster when stimuli are presented in a vertical array, whereas English speakers exhibit a distinct advantage for horizontal arrays [cite: 1, 3, 16]. 

Furthermore, bilingualism studies reveal that cognitive frameworks are dynamically modulated by the active language. Mandarin-English bilinguals exhibit different mental timelines depending on the linguistic context. When operating in English, bilinguals tend to arrange time horizontally; when operating in Mandarin, they frequently map time vertically, mirroring the traditional structure of Chinese script [cite: 17]. This indicates that semantic structures create distinct cognitive compasses that influence reaction times, memory retrieval, and temporal reasoning [cite: 3, 17].

### The Epistemic Reverse-Sagittal Timeline

While the English language equates the future with the space in front of the ego and the past with the space behind, this mapping is not biologically mandatory. The Aymara language, an indigenous language spoken in the Andean highlands, exhibits a static temporal model that strictly reverses the conventional Western sagittal axis. In Aymara, the future is conceptualized as behind the ego (qhipa pacha, "back time"), while the past is situated in front (nayra pacha, "front/eye time") [cite: 15]. 

This temporal orientation is grounded in an epistemic metaphor prioritizing visibility and knowledge. The past has already occurred, is known, and has been "seen," therefore placing it in the visual field ahead. Conversely, the future is unexperienced, unknown, and unseen, situating it behind the speaker's back where it cannot be observed [cite: 14, 15]. 

The depth of this linguistic relativity is confirmed by extensive gestural analysis. Ethnographic studies involving videotaped interviews demonstrate that when Aymara speakers discuss the past, they systematically gesture forward, away from their bodies. When referring to the future, they gesture backward, pointing or waving over their shoulders [cite: 14, 18]. This behavior is most pronounced among older, monolingual Aymara speakers. Younger speakers, heavily influenced by the spatial metaphors of Spanish, show a gradual erosion of this distinct gestural pattern, highlighting the malleability of embodied cognition under cross-cultural contact [cite: 18]. A similar "past-in-front" conception is documented among Māori speakers in New Zealand, further supporting the premise that the directionality of mental timelines is culturally contingent [cite: 15].

| Cultural/Linguistic Group | Dominant Spatial Axis | Metaphorical Mapping | Cognitive and Gestural Implications |
| :--- | :--- | :--- | :--- |
| **English** | Sagittal (Horizontal) | Future = Front; Past = Back | Time is an ego-moving or time-moving horizontal trajectory. Gestures for the future project forward. |
| **Mandarin Chinese** | Vertical & Horizontal | Past = Up; Future = Down | Enhanced processing speed for vertical temporal arrays. Bilinguals switch axes based on the active language. |
| **Aymara** | Reversed Sagittal | Past = Front (Visible); Future = Back (Unseen) | Epistemic mapping. Speakers physically gesture forward to indicate the past and point over the shoulder for the future. |
| **Kuuk Thaayorre** | Absolute (Cardinal) | Time flows East to West | Temporal arrays are arranged according to absolute geographic coordinates, regardless of the speaker's physical orientation. |

## Grammatical Gender and Conceptual Categorization

The assignment of grammatical gender to inanimate objects is a prominent feature of many languages, dividing the lexicon into masculine, feminine, and occasionally neuter categories. Linguistic relativity suggests that these arbitrary grammatical assignments influence how speakers conceptualize and describe the physical world.

In seminal psycholinguistic studies, researchers examined how speakers of gendered languages attribute traits to inanimate objects. For example, the word for "bridge" is feminine in German (die Brücke) but masculine in Spanish (el puente). When asked to describe a bridge, German speakers disproportionately selected adjectives such as "beautiful," "elegant," and "fragile," whereas Spanish speakers selected adjectives like "strong," "sturdy," and "towering" [cite: 2, 19]. Conversely, the word for "apple" is masculine in German and feminine in Spanish, prompting German speakers to describe it as "hard" and Spanish speakers to describe it as "juicy" or "sweet" [cite: 20]. 

The robustness of these findings has been the subject of intense debate within the context of the psychological replication crisis. High-powered replication attempts, such as those by Elpers et al. (2022) utilizing 375 participants and Bayesian analyses, found mixed or negligible evidence for grammatical gender effects when accounting for key sources of error variance [cite: 21, 22]. Similarly, Samuel et al. (2019) expressed profound skepticism regarding the Whorfian effects of grammatical gender, noting that methodological variations often account for the observed cognitive shifts rather than deep conceptual restructuring [cite: 19, 21].

However, the transition from human subjects to Large Language Models provides a novel, scalable methodology for examining these semantic associations. In zero-shot experiments applying the Boroditsky protocols to open-source LLMs like Llama-2 and Mistral across ten gendered languages, researchers found that models demonstrably internalize these grammatical biases [cite: 19, 23]. When prompted to describe gendered nouns using adjectives, the models exhibited highly consistent, typologically aligned biases. Most notably, researchers were able to train a binary classifier to predict the grammatical gender of a noun solely by analyzing the adjectives the LLM generated to describe it, achieving successful zero-shot transfer across distinct languages [cite: 19, 23]. This confirms that while human replication may be noisy, statistical language models reliably encode and perpetuate the conceptual associations embedded within grammatical gender systems, raising concerns for automated translation, personification, and anthropomorphic text generation [cite: 19, 23].

## Artificial Linguistic Relativity in Computational Models

If language acts as the architectural scaffolding for human thought, LLMs offer an unparalleled medium to test whether statistical language structures enforce distinct reasoning pathways. Recent empirical evaluations of model internals reveal that LLMs do not merely translate concepts between languages; they internalize the unique syntactic, logical, and epistemic biases inherent to the language in which they are prompted [cite: 9, 10, 24].

### Syntactic Biases in Causal Reasoning

A definitive demonstration of artificial linguistic relativity is observed in the evaluation of causal reasoning across language families. The BICAUSE dataset is a structured bilingual benchmark designed to test causal reasoning in LLMs, providing semantically and syntactically aligned Chinese and English samples [cite: 9, 10, 24]. Chinese (an analytic language) and English (a fusional language) possess profoundly different conventional orderings and structural preferences for expressing causality. The BICAUSE framework decomposes each causal chain into 13 interpretable syntactic components and 3 semantic-level components to track internal model attention [cite: 9, 24].

When processing identical causal scenarios, LLMs exhibit typologically aligned attention distributions. In Chinese prompts, models allocate significantly more internal attention to sentence-initial connectives and causal antecedents (the causes). In English prompts, the models' attention distributions are more balanced, with a pronounced focus on verbs and consequences (the effects) [cite: 9, 24]. 

Furthermore, LLMs internalize language-specific preferences for causal word order. When models are presented with atypical, reversed causal structures in Chinese, their reasoning performance degrades sharply. The models rigidly apply the internalized structural prior that the initial phrase signals the cause, misallocating attention and failing to correctly infer the outcome [cite: 9, 10, 24]. English prompts, conversely, maintain stable attention and accurate predictions across varied structural orders. 

Despite these surface-level syntactic constraints, mechanistic analysis reveals a deeper phenomenon regarding representation alignment. When an LLM successfully completes a causal reasoning task in either language, its hidden representations in the deepest layers converge into a shared, semantically aligned abstraction space [cite: 9, 24]. This suggests that while reasoning pathways are strictly guided by specific linguistic syntax (a Whorfian effect), successful comprehension ultimately maps to an Agnostic Meaning Substrate (AMS)—a latent, language-independent topological field where conceptual relationships stabilize independently of surface vocabulary [cite: 8, 9].

| Reasoning Metric | English Prompting Behavior | Mandarin Prompting Behavior | Theoretical Implication in LLMs |
| :--- | :--- | :--- | :--- |
| **Attention Allocation** | Balanced distribution; high attention on verbs and causal effects [cite: 9]. | High attention on sentence-initial connectives and causal antecedents [cite: 9, 24]. | Models precisely mimic the structural biases and semantic prioritization dictated by the input language syntax. |
| **Word Order Rigidity** | Flexible; maintains stable reasoning accuracy across sequential and reversed causality [cite: 9, 24]. | Rigid; reasoning accuracy degrades significantly when atypical structural orders are introduced [cite: 9, 24]. | Whorfian structural constraints limit the model's flexibility, binding causal logic to rigid syntactic expectations. |
| **Deep Representation** | Converges to language-agnostic abstraction upon successful causal inference [cite: 9, 24]. | Converges to identical language-agnostic abstraction upon successful causal inference [cite: 9, 24]. | Conceptual meaning stabilizes in an Agnostic Meaning Substrate (AMS), independent of the input language [cite: 8]. |

### Spatial Frames of Reference and Visual Grounding

LLMs and Vision-Language Models (VLMs) also exhibit significant cultural biases regarding spatial frames of reference (FoR). In situated communication, spatial descriptions are inherently ambiguous. A phrase like "the cup is to the left of the tree" depends entirely on the adopted FoR: an intrinsic frame (relative to the tree's front), an egocentric relative frame (relative to the observer's body), or an absolute frame (cardinal directions) [cite: 2, 25, 26]. 

Evaluations utilizing the COnsistent Multilingual Frame Of Reference Test (COMFORT) demonstrate that modern VLMs struggle profoundly with this ambiguity. Despite extensive multilingual pretraining, VLMs lack the flexibility to accommodate multiple spatial frames. They fail to adhere to language-specific or culture-specific spatial conventions, overwhelmingly defaulting to English-centric, egocentric relative frames of reference, even when prompted in languages that culturally demand absolute or intrinsic frames [cite: 25, 26]. 

This spatial reasoning deficit extends to textual models as well. Benchmarks like bAbI and StepGame evaluate qualitative spatial reasoning but often rely on simplified, toy-like grid environments with fixed distances, failing to capture the complexity of real-world spatial relationships [cite: 27]. When confronted with multimodal inputs where embedded on-screen text contradicts the visual scene, VLMs suffer from Text Overlay-Induced Hallucination (TOIH). They systematically hallucinate, prioritizing the semantics of the textual overlay rather than grounding their reasoning in the actual physical visual evidence [cite: 28]. This indicates that while text-only structural biases translate seamlessly into LLMs, situated spatial reasoning remains highly brittle and monopolized by the dominant cultural frameworks of the model's primary textual training data.

### Evidentiality and Epistemic Stance

Linguistic relativity profoundly impacts how languages encode the source and reliability of information. Languages such as Turkish, Quechua, and Korean feature mandatory evidential markers—grammatical components that dictate whether information was acquired firsthand, secondhand, or via inference, thereby defining the speaker's epistemic stance [cite: 29, 30]. In Turkish, the past-domain suffix *-DI* denotes direct knowledge or high speaker commitment, whereas *-mIş* denotes indirect evidence, inference, or lower trustworthiness [cite: 29]. Similarly, in Cuzco Quechua, the marker *-mi* denotes attested direct knowledge, while *-chá* denotes conjecture [cite: 30].

Human speakers modulate these markers systematically based on the perceived reliability of their sources and the social context of the discourse. However, evaluations of modern LLMs on source-sensitive reasoning benchmarks reveal a severe computational deficit. When tested on controlled cloze contexts manipulating source trustworthiness, LLM behavior is highly unstable, prompt-dependent, and frequently reversed [cite: 29, 31]. While humans demonstrate a robust trust effect (using *-DI* for high-trust and *-mIş* for low-trust contexts), LLMs are frequently overshadowed by base-rate suffix preferences and output compliance failures [cite: 29]. 

The TR-MMLU (Turkish Massive Multitask Language Understanding) benchmark highlights that these failures are exacerbated by tokenization issues; models utilizing generic multilingual subword-tokenization fail to respect Turkish agglutinative morphemic boundaries, degrading downstream reasoning accuracy [cite: 32]. This divergence highlights a fundamental limitation in artificial cognition: while models can map statistical correlations between nouns and adjectives, they lack the persistent vantage point required to dynamically reason about epistemic truth, source reliability, and social stance.

## The Symbol Grounding Problem in Artificial Cognition

The observation that LLMs struggle with situated spatial frames and epistemic source tracking points to a deeper philosophical limitation. If an LLM perfectly maps the causal syntax of Chinese or the temporal structure of English, does it genuinely *understand* causation or time? This tension revives the Symbol Grounding Problem, one of the most enduring puzzles at the intersection of cognitive science, philosophy of mind, and artificial intelligence [cite: 12, 33, 34].

### The Formal Limits of Syntactic Manipulation

In 1990, cognitive scientist Stevan Harnad formalized the symbol grounding problem by arguing that a physical symbol system cannot acquire genuine intrinsic meaning if its symbols are defined exclusively by other symbols [cite: 12, 35]. Harnad likened this to a monolingual English speaker attempting to learn Chinese using only a Chinese-to-Chinese dictionary. Every definition links only to other unknown symbols, trapping the system in an infinite, circular regress—a "dictionary merry-go-round" [cite: 12, 35]. Genuine meaning, Harnad argued, must be grounded in nonsymbolic, sensorimotor experience; symbols must connect to the physical world through perception and action [cite: 7, 35].

Modern LLMs operate entirely within this ungrounded paradigm, updated for the era of deep learning as the "Vector Grounding Problem" [cite: 36]. In an LLM, words are manipulated subsymbolically as high-dimensional continuous vectors. Yet, the core problem persists: these vector components are connected only to other vectors derived from text, never to perceptual realities or physical referents [cite: 36]. Consequently, researchers argue that LLMs operate purely at the level of syntax (manipulating structure) without ever achieving semantics (comprehending meaning). This phenomenon echoes John Searle's famous "Chinese Room" thought experiment, suggesting that fluent symbol manipulation does not equate to conscious comprehension [cite: 33, 37, 38].

[image delta #1, 0 bytes]





### Structural Isomorphism and the Observer-State Framework

Despite lacking sensorimotor grounding, LLMs display an astonishing ability to organize abstract concepts, navigate spatial logic, and group physical attributes accurately [cite: 6, 7]. This has generated a counter-argument within computational linguistics: grounding may not be strictly necessary for meaning if the structural representation learned by the ungrounded model is isomorphic (geometrically identical) to the grounded representation held by humans [cite: 6, 7]. Because human language encodes distilled knowledge of physical and social interactions across millennia, the statistical co-occurrence patterns of words carry the precise "shadow" of physical reality [cite: 6, 33]. Through structural isomorphism, text-only LLMs can successfully generalize spatial directions and map the geometry of color words with high fidelity [cite: 6, 7].

However, categorical and epistemological analyses suggest that isomorphism does not solve the symbol grounding problem; rather, it circumvents it. Operating at a mathematical level of abstraction, researchers model the human epistemic route as consulting and interpreting directly grounded content, whereas the artificial route relies on prompting a trained model and interpreting its outputs [cite: 39]. LLMs lack unmediated access to the state space of possible physical worlds. Their apparent semantic competence is entirely derivative, exploiting pre-grounded human content and relying on human agents to project meaning onto the output [cite: 34, 39]. Hallucinations, in this framework, are not mere implementation bugs, but intrinsic entailment failures resulting from a fundamentally ungrounded architecture [cite: 39].

This structural deficit is further articulated by the Observer-State Framework, which argues that conceptual reasoning requires a structural element absent in current LLMs: a persistent vantage point capable of evaluating outputs, assigning stable meaning, and compressing conceptual conflicts [cite: 40]. Grounding alone does not provide the capacity for conceptual integration across shifts in context. Scaling parameter counts increases fluency, but meaning-level generalization frequently collapses into form-level continuation, indicating that the underlying difficulty is a structural cognitive deficit rather than a purely statistical one [cite: 40].

### Limits of Multimodal Integration

To address the symbol grounding problem, AI research has shifted toward Multimodal Large Language Models (MLLMs), integrating visual and acoustic signals into the linguistic representation space [cite: 33, 36]. This strategy attempts to provide a "grounding kernel" by linking text directly to pixel arrays or audio frequencies.

Systematic reviews spanning research from 1990 to 2025 indicate that while visual-linguistic alignment is learnable, it remains statistically fragile [cite: 12]. Multimodal grounding demonstrably improves model performance on concrete, highly imageable concepts. However, it largely fails to support compositional generalization and abstract reasoning, often yielding catastrophic failures when novel structural combinations are introduced [cite: 12]. 

Cognitive scientists maintain that visual grounding without physical interaction falls short of genuine embodied cognition [cite: 33, 36]. Humans acquire knowledge incrementally, building complex concepts upon simpler ones in a structured, goal-directed developmental progression. MLLMs, conversely, are trained on vast, randomly ordered datasets without autonomous agency or consequence, inhibiting the construction of a deep, developmentally grounded knowledge base [cite: 36]. Furthermore, the Words As social Tools (WAT) theory posits that abstract concepts are grounded not merely in sensorimotor experience, but in interoceptive, metacognitive, and socially mediated interactions—processes that current neural architectures entirely lack [cite: 12].

## Framing Effects and Cognitive Alignment

Despite profound architectural differences and a lack of grounded semantics, LLMs successfully replicate a broad spectrum of human cognitive vulnerabilities and heuristic biases. This occurs because psychological phenomena are densely encoded within the syntax and sentiment of the training corpora. One of the most prominent examples is the "framing effect"—a cognitive bias where different presentations of identical factual information alter perception and decision-making [cite: 41, 42, 43]. 

Recent evaluations utilizing the WildFrame dataset—a benchmark of 1,000 real-world texts designed to test responses to positive versus negative framing—demonstrate that state-of-the-art LLMs respond to linguistic framing in a highly human-like manner [cite: 41, 42]. In controlled experiments comparing multiple prominent models to human crowdsourced annotations, researchers identified strong correlations ($r \ge 0.52$ to $0.57$) between human and machine susceptibility to framing [cite: 41, 42]. 

Crucially, both humans and LLMs exhibit a specific psychological asymmetry: they are significantly more influenced by the positive reframing of negative base statements than by the negative reframing of positive statements [cite: 41, 42]. Furthermore, model size correlates positively with human behavioral alignment; larger models tend to exhibit a higher correlation with human cognitive biases under opposite framings [cite: 41, 42]. This behavioral alignment implies that LLMs accurately simulate the valence-matching and semantic mediation processes of human psychology [cite: 44]. However, it also raises complex ethical and architectural questions regarding alignment: preserving human-like cognitive biases makes AI systems more relatable and conversational, but integrating these models into high-stakes decision-making pipelines simultaneously imports the structural irrationalities of human thought into automated judgments [cite: 42, 43].

## Conclusion

The intersection of large language models and cognitive science provides profound empirical validation for the theory of linguistic relativity, while simultaneously delineating the strict limits of statistical intelligence. Research across diverse linguistic frameworks confirms that language is not merely a neutral vessel for communication; it actively shapes cognitive navigation, memory, and perception. Whether analyzing the sagittal timelines of English, the vertical spatial mappings of Mandarin, the epistemic reversed orientations of Aymara, or the absolute geographical coordinates of Kuuk Thaayorre, it is evident that linguistic structures forge distinct mental topologies.

As computational systems trained exclusively on these structures, Large Language Models prove to be the ultimate Whorfian subjects. Because they possess no independent sensory experience to anchor their cognition, they are entirely captive to the syntactic and conceptual rules of the languages they process. When prompted in Chinese, they adopt Chinese causal attention patterns and rigid structural expectations; when prompted in English, they adopt English syntactic flexibility. They replicate human cognitive framing biases and mirror semantic geometries with astonishing, mathematically quantifiable accuracy.

Yet, this mastery of syntactic structure underscores the severity of the Symbol Grounding Problem. LLMs manipulate high-dimensional vectors within a closed, ungrounded topological space. While they achieve structural isomorphism with human thought—mapping the relationships between words precisely as humans map the relationships between physical objects—they fail to achieve intrinsic meaning. They operate as highly sophisticated mirrors, reflecting pre-grounded human experience. While multimodal integration offers a promising pathway toward connecting text to visual arrays, genuine comprehension—rooted in embodied, interactive, and socially mediated survival—remains an exclusively biological phenomenon. Ultimately, large language models reveal that while syntax can flawlessly simulate the shape of human thought, it cannot independently generate its substance.

**Sources:**
1. [Link](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG3Kks8qbZTEApsHg7qx6v5kGf3z8fTa9l7F9OjsF_4AoRSOvBNDBIuuneaWsHeccrptiPGvz7YyPXKIqDPjl44xSMr6UrGDiUfO55YubfycDbYuZ6cJNd75higJCCSI1sxo8qqzuRzvg==)
2. [erigo.se](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE6XAA_lgIG2XWKg7pu2gbauxY88Udkpg0UJ-ZHMiV3ItS6EDZX6dBuyIFegFM2dr1jz63d8rpcxPLhXX62-7019dfOPiz75MbpuxfwE8_vWPRB-3XcUIlIWqmonApYRhYdaAll21lPTFQFBtkPJOGzw1lz8QhWtGVklPkYeNczintmJJ_OWB_bpKoWvHliPA==)
3. [substack.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEVF25tbcxH28dRSFPZRuMS6u3tQkp1rpNFa-anWK8ZTHbzbAU5WqTtbQq9-doFBQF0sj2-JEYPZwI-_pKoSgLgOL5IV8OucrIVcSc15Z-PfhdN9OXNa9wSmOscmOoKWL-c1fF1DIzcDM9Fv1kXz-vt7hyo)
4. [purdue.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH8TgqAQd99yGZrR9izhqN-o0lwkuB_eQO3bWma5pMbkm0Dn4O2TOUyTlkmM7oQCM9QDIm-TEQkilLYOVgkyQTk5HexP5_x6WMc7_owu_LO99iKxQ7JrdaHXdNlkYHuf2T2LdvdN99YcdM=)
5. [shav.dev](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHsa289lqrwt9JhfV6t-DuI2cC2AqliTh0DMkO4WPH-3aqOOBQT7b8OZyrGgIqCjcyjj0_qSOeo2XtFcnLrAde35QAvjEfQ_pxthS6JOK1lbYzWawLNLaH_)
6. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEuDwSWvGIxcNAEkL7hA9XrtZ7grnVwNsWbQW318J5faFV8QnHEb9OAlRp1OEHAO4tX2dAnPvyP4bbkKKM4dCuxpEmRQBl4LSUvrm4vkp_4xmzzHZRFw5JG)
7. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHq_3Xz5k1ZHiej3EeJgqlCXMhpfWnDrUAmCSDKEefLmJPeM9sxM8Xbe1qWWQi65jJffajKTVQgRtNm7ZrrQ6gWBCY8cp1Lf5fiXOG7tFEB7u2zj6ZgY5K8JXjn6z1pqyVBahBo9PNS)
8. [zenodo.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGXehSzvRSRLuy-AkfkV-R0PGVRDPRcNkTssvJ5JkQ23yU6yiSeqdyjEyByLL3oTrw9dK78C23feMpXn9xpXZpFa6MqzmOjkF5E-wzXVSykksmcjfcwdLPUZE_tgexHbQ_5nUjC0x7h5fmWIXYNbO7JLH6Jhrq5MBL-hVdIe705oRefgCEp5l0Tt81MN7s2FY9lgsnbhfQkTV22-9bEwV7iOKeLelv-HE9x5IsT1Ni53K8MkltqT6ybM9JgWslceAl-fWkblv3CZMmtRKu9RX35ymaa-KgEWZSC5Q==)
9. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHofW0Uja2TaPPBGxKjUh1Situtvev-xT5QkBiz2FkvFrLJN2Zml24RFBxNWGD6wIMpQXhaCmv62m-Kib1LI9D_sYl05H9Z_M_RCeIBYdEHIrjISXrq0y_MQQGOae_478uPLnpJRtfIfWF-)
10. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF-bew9N7TMGJHnE6E8_5r0XVH8zJtupGo1TI0IJSC_qCII1C3Z3E9Bdd5wza5Vy38CNq3smK9jhNw7cwxygeaWuCuHE9IPRPPzzWK1irocXkE8qd3E)
11. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGOGx5cOPcKuEUYmSN4tCsH0jGdoHqBkHP5xKsoC_-XUzPmWaTakbP-jXq36KzAoCOx5_tYezPtNIANoWtaqrbi2XNwUy2hcEEQuoE3xNBZOxLuQuNQ4dMkjdZDLCZJGvtpHk8_EkuRot_Gvpm5FmV9ZLKtVkU6QZRVssq8PFdO1AyMs8RXGAsV0gJg1-LZkzF8kYV--8qjnMJurA==)
12. [tha.la](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGv6UOR6NqV3Fc5nBZNiYSx3ZVj7mHhGpWaBIuSiZ1rOsfFhLmhJpNoXaWQz10sFWquUp8rA4j1472ADiYy0LrBsDdo1p5qVCLIJbh9ARJqF9vV6geN1MG9-z0nTJrluUXzehApNj_KyWH8nUM0ROur6fI0PKWgyk6DBSYAAJqZhin0Q171HL8nVyZo039EAPIqcgF5VnDLkSCJuLbgiyjhBhCoxryXAZr-)
13. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHT1gYjpaDB5gIesKIinPJ60quzx8k7TRMCvPty4EuUpMJ91WjBaAVdpkoWMtabXZvhdybwEfRJLl7c202dU-Qf_IAOgN6TtVmak1TDCSO7sjabu7VkoOUF99qA5H1acaXbgWc_RA2CqG9eRYLYJxGJGQktZ36CKgawkuhjUkWnBknMNT4RVXkrltEUAhjKSjAbw_VmZ95fBUiqFaoeOA==)
14. [alphanome.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQERjnW7_JJ354ai7ZoFik4N1Rd7pc9NeFRIOD94ZlgUFfiuyfgsAXcOBQKB1MK8486M8G8j7IWMv9HfNm81xYLcUWnWOk2k0YVsw3pFQyc1Fl5WlaRzZ7p78JpwFJAaCQUbheE90ey6LNWPV1_u76vkcA1AvQtx1iifYWNopklN9Eph3vsTvl-L0rZHqkK67OIsk7_n)
15. [frontiersin.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFK0tH7A6-azQf8BHl8G-U0ar3Abw3LhSKC0XmHyQnqMeQkCEoSvgyPdKKeGFAzXwGfHr7qQZl5geLZ1AZUTqL7w8HgBcAngjKv31F-LMztSRQwh3p1Tkkq-TmBiVbF29hQ-KuW_-oQAiCyW81W_FRzprAWykiVZeVh9ijV5tbIDdmJgWJW61eDOzVXVw==)
16. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFxyAI5qB6XsAyjqDDYWxSVsM63VbbNMxFQJg_s9ESUjre-n9tDkDYawIpP0L7XEUuJa55xOIad5Mg8d2fGwKm23lYBxhJEUuuIkHtLW0pY2Rc4170hTMnsg4VMrhXQJhihfatStvLugnk2ZEfnn5BtCVCZv5PDOVWCM6MbGEhfrc2Ge9SqoCbSoAAMcbWYzKQHWRkVGEEcxMfAfWSkdTpgNnLDDhjcA_Ub3uq3dHCVT96-OsfP6Q==)
17. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEfZ8EhY1qReKXjoZWfAEW7v-zPC3dlXtlDa5TVXxKFQhErBIjbFy5aDqi5I0R0wcx5yLAfUkgh4tlvn5vhSpFktXEDM0VnJapJFizhiwFzOqbh0YUA4r_wtes49cIqMizJYMRFQH9UCHSHKhPWbEgGKkOK0i1R1pISiDzVPUkOG6pIDTww)
18. [academia.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFtEIAtb0S7RO80uZtAH7IVdhzCxOUIUz448gZ44cuEDBdEp36qLOimqnc4SwClddJU7_IuUzIpTd6rvuy5Ga4zB0BO4go0GTQY5zsq0I8VnAIHzAeZO-e9Hj8eO69n0qri9LXDBftMSddsBIwgGxYz_YUC9nJbkZ7Rljz7NFgveX065YHq9E0WuM-NIbm4yDyH0GEgcAvh_R-biN5TUaxJ7MwVT0P3ioic_215Rxv49YD42TrLKbZjQPuB9aXaNXghdXI99r_vUsyxw0w92adzJ5gG18dgvUlL--aUlQozpV5uvl6k7SDk)
19. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGcp7dNe76K1xwPFQyt7XOJ2Wu16awUksGPlphUx0A7BVaQsNujqRuH-gqbshLrTHbZj-I0uN7h9mWWhDBVQy4enZcrvbaq-WZetswF7200NKThV9qR)
20. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFNy0NGOx_xMGdR40ExwkNnbyApFMM7t6CaUbEtRJRSndHd2tEHlq4kUbCcMYo-9uQeYaxQ8PzMF47whtPx_TFnsXp_UBdKnHNyUQ5fXHtTAe2lQXTy37ZW6XNYH9TKLk5b4Bg7xBPA)
21. [swarthmore.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGNxnolrYGocxTjoPLNph9KxumrwXzSOQhwuHDPnL7CBFSBK96CLp-A_VRi4Z_273yHX1Q7wRXi6ga1EfGzBcTzFvVSZzPqcVGQCanC4uvmkdP0oJZXhPKbqEhKsRul-i2eyv6qgRZyzR-gMnBRzOtbfuBuqQnX6JDz0jZAw5z5x72B46ElBkDk6g==)
22. [reed.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHncMo4qAW7tIq4UOngMz2Wq_kBjJeAV4c96rJU6JJKXuBDehJxak9-uyh7i1F9tuhyQ4151S1FbYXS0oi7_GwRbT40x7ckgH4Jpeve5Yz2MjxpfcWgAPbI-ToiEpZ3JYTgrizLrEDZuVpnBJRPZzuALB9p2Bn3hsKhGOyA62pInE-NrU7bWEgKWTyT07rjJcdngFoDdCMxGNbckmCp3kURVYwZDQp6xnM=)
23. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE4_zPXDZtn0gGRvbhrupdF2olEUBnnhaxH2Jsvbe5oF8DUOLoWgJK-mNBbvd_XBRrEPuSGlqqiFxTQfPSj6QnhTcChXii9nWmlx9ypdiBoX7yrvNP0JuUKI95yOJCS)
24. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEW6QlQdnukWaw5Y1oOLC9gfuo7NLK5d6WViCJDtUIP1iB0XikS0FeKtXuWCUPtStXds0wwQnisfsmg9R0MymmFT2iP1muZGgFwKkT6pUuikMK4nhc4)
25. [github.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFr5Ocp6Kied6mlbJN0WQ3w-f_RfhkvfTNDUUPKdWOvUBlB19dAhaUlCRVMrfg4AucaV3UR5VER38BeJeNZ8knWrgzWgrnq4XLKHrIDxHx7sinYUfRF8PU=)
26. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFyJpEZyC-vicqz5oG4TGFkYIfbNBpLlnK-tv-q-l-F-GRP4jUQ-81VjAQOgzt6RRbjATvUjt9KfNmTKKIrEgepw_38P1cHVISEhbmcfeN9KY08NAE6)
27. [ijcai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGlC7Q6ADwruXOSDxmL_vzrKiiMMqmJOYglT-IhGJkYPtWPfD1mocfmKQrutMKpbT0Yh-oxTzPaudEbRvfJlvlcfo0PEVmKqAJ-UvOwCR36REwPEHGjPUztphk2cgmWQuGybVQn)
28. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFV0rxazxj1oRml-DNMavOhxd3uXXCDhuFyZGJZd2z7xagYokBpt_fFR6CypbUtzv8uGmp5BMRKW6cVMTvO6Gj7zMP4B1gCLl1s0vmtkjARCUpD3lx063ZT)
29. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFTR35cpUEbfcQ7_iRStr2gUJ4qGSzVjtw8joXHVG3a9VuilC3WPXrteLIm7vY_i8vqoskdnVITI8N38rFevgSGYiPFgCOgcTSwSWQAplclLnc-qEktyhbs)
30. [linguisticsociety.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEkTyeYUuk6wCeD6cDRNSKmuBuEkGUUl2XIqHy0i4lJpUz67l2BgwW9gk1XqDpKPid_TUWAeChWkk6yPznOC8OWt9XOUAZdq5EMz78nI8BqFxEAo-67tsBhJLcs9s4FBc9MIvUYu3xbuyklN3hRxNO5nrhsgAg1tKvQX1oKx0Hczvvjzb4Cb26VLWZhELMUOaFwjS-R-nJB)
31. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE5lmaNHltK3dXz7B3gRQmqdXPp6rivyf-I0bMXkoQ97uQTXoxzZz6Rb_dkgHRqzwcJXcPerGV3p0InC4xRjg0W3cnpqchS-Vec12TmfDejlTO8nMV9)
32. [emergentmind.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFRVYJZIoz3tx1Tn7jYGWOQLLoM2zC-kgXZ6eIdXaYE9zI5Vz8xxxde5ddrs75jlaDcK8VUYt8MY-GeiStmurc_KeV5T04LJQMvs8-44s20ojIX1-06yHGUbV5bLh6RrBpRu5qCDVCjk27f)
33. [bio-integration.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGSh9rmw-cQ1czuaTyN_6umCQazsdLhK9MPueAZtFNVw2ViQgFQ3XiJB-f2z3VmO3JMTX7ffjIcedl8YgbjuPHbg44ePQcMqpXzroneAsOHcLPfTk2vEKo6G9kovUGIYYpPqJ1Exnxy-hz_Hn_iWK7Wk1LYx95QMFcS5hkI)
34. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH-rBgDL0bbzRDNmmNfXKnCvJeWYwI0nQXWrCNx9Ie0bAaKTNF22e4mW1677ec5ZNfbibyGIwNFTAhwTErpMiuxgQlU4oKfYDUrmJ_YzgEi0sRVfbUrV1qd)
35. [emergentmind.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHHtc-rKKk8tHXzTX8Zl52bhA4EnJY0IaILyZs3TV6bXBskyPdcGfPanTvssx9MyQxLJqqcaKNSQuZhNa1RoXQf6mB4sCEiQQrAV3ZQIJ19TgzgS9KXp8OsK19LWm_xTf_vNEN16GYdtMyMIZ61vbk3VA==)
36. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFsD2SWq5HqPCww5RQw-CcdcQJm9XXrEa5-BUGvlqufjFj2V5uO0z4ALOV5VLL8-8dV5S8p_1Su5wqol37WYbOJcQ9RBMhbWGTIdsS3O2RfStIzAh_q7OxQFDphzFTLp5ERXSWwWJ1h)
37. [cuni.cz](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGBlayEFFRTKgph6B7QF9opX75o1tnypPmFSQPrnAXZFUZtkaWuuDmmeiW_QSwsTw3IbC4WL22iT2VKOvTZXjNivxScyfZlsP0Z8bPXEU0wKeiZYXBwcq46nx6nl6WAC5aYMLnYftkrSI18v7Aen3nvhUHeWf4nOZSA1KPbFw5No5AO2VNb5v19Lg==)
38. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHGv3SUhcTOi-j0t8WdKU6q1epra4Ohnk6PfeotLpxylJn1E93mDu8_OKJy8SzsF4x0S3jCfLpxVRqLXBM9Bh469se5FpZnuBYfHZgYoBwWH9t66iHmGU0EhrRIp70_R8-wBW1mqwrQw5Erfuu00rpTP091BB-5Uv-h53FFI0u4P8IqCz2F4WYW2L2ZjmLrHWR7qPA5P3dsEcWfT2MAc5ZFRULU3_ZCFyvBOf4PCA==)
39. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGHhwPhTfuZdZ9k9T1HtsUuguwBgxiCLJkpiGJyLm88VcmQN5h4S0DMDa4dJc9VfGK2vIONWn4ehaEXXKKHBoet_sOhoAlLsmaCx_xGIZogQb4E9NijIZ92)
40. [preprints.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3N0Zg8nK9pd2xkK2cQHIvRTUAqxw5IeG4S0iAT5Cw1VwzMZ-AzDBSVmDJXqtTGSfyB9sD1QjZTkIprHwtWqnnGd0qV78hkEXiGTO378Egbi4hBbwiJeAhcdzgaHetlyYPT_4aog==)
41. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFEHjUZnBEMD_jhdKv6Xy-TbpFXinqSnecnMqe8FrCBnV9lt60PnmploAoZLZ-mxDFuqL39vW0Z7SyAfxxKgwvI82mI4VxOzqU_c5iT_YO9xgxOAVqdijSZ)
42. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHHkIh7Yirh1VdOzXonKM9VPe7W7KIBqmHZUd5F0K-4UYIOvCS3MW8F0hx7Z4tJN3cGAJw-1JCdd8oXfv-B32dh_sIOggFlheHS4DnUnlx_lv0sZ84SswPI)
43. [escholarship.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF93NGHdyUEhOBLqEYAd3ZH1QJeqJRsX8cPCl1PeTKiCvMxz9bGWNvQTP8Hb86WGr_r1U7Uw7Jj-TTpud221ltbei2zN9byWQjUf1SiI2VvgRQM54K7o5gPN1G8plMX)
44. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFv2AxRvm8P27mfYhGhR0t3ykGHuzJePUsiFErN10Ir5zaTXL5a2LW2WWiD9AG0BA4SNg041t6wJSwKkqmfT8RcV7Gc3JUoVO0SIPqocdCMeNOB4GagyGf7kuMUMvUDAb3td3LcMlmG)