Dissociating language and thought in large language models

Trends in Cognitive Sciences, Год журнала: 2024, Номер 28(6), С. 517 - 540

Опубликована: Март 19, 2024

Язык: Английский

---

Language writ large: LLMs, ChatGPT, meaning, and understanding

Frontiers in Artificial Intelligence, Год журнала: 2025, Номер 7

Опубликована: Фев. 12, 2025

Apart from what (little) OpenAI may be concealing us, we all know (roughly) how Large Language Models (LLMs) such as ChatGPT work (their vast text databases, statistics, vector representations, and huge number of parameters, next-word training, etc.). However, none us can say (hand on heart) that are not surprised by has proved to able do with these resources. This even driven some conclude actually understands. It is true it But also understand do. I will suggest hunches about benign “biases”—convergent constraints emerge at the LLM scale helping so much better than would have expected. These biases inherent in nature language itself, scale, they closely linked lacks , which direct sensorimotor grounding connect its words their referents propositions meanings. convergent related (1) parasitism indirect verbal grounding, (2) circularity definition, (3) “mirroring” production comprehension, (4) iconicity (5) computational counterparts human “categorical perception” category learning neural nets, perhaps (6) a conjecture Chomsky laws thought. The exposition form dialogue ChatGPT-4.

Язык: Английский

---

GPT and CLT: The impact of ChatGPT's level of abstraction on consumer recommendations

Journal of Retailing and Consumer Services, Год журнала: 2023, Номер 76, С. 103580 - 103580

Опубликована: Окт. 20, 2023

This study explores how ChatGPT interprets information through the lens of Construal Level Theory (CLT). The findings show that exhibits an abstraction bias, generating responses consistent with a high-level construal. bias results in prioritising construal features (e.g., desirability) over low-level feasibility) consumer evaluation scenarios. Thus, recommendations differ significantly from traditional based on human decision-making. Applying CLT concepts to large language models provides essential insights into behaviour may evolve increasing prevalence and capability AI offers many promising avenues for future research.

Язык: Английский

---

Still no lie detector for language models: probing empirical and conceptual roadblocks

Philosophical Studies, Год журнала: 2024, Номер unknown

Опубликована: Фев. 17, 2024

Язык: Английский

---

Do Language Models’ Words Refer?

Computational Linguistics, Год журнала: 2024, Номер unknown, С. 1 - 10

Опубликована: Июнь 10, 2024

Abstract What do language models (LMs) with language? They can produce sequences of (mostly) coherent strings closely resembling English. But those sentences mean something, or are LMs simply babbling in a convincing simulacrum use? We address one aspect this broad question: whether LMs’ words refer, that is, achieve “word-to-world” connections. There is prima facie reason to think they not, since not interact the world way ordinary users do. Drawing on externalist tradition philosophy language, we argue appearances misleading: Even if inputs text, text natural histories, and may suffice for refer.

Язык: Английский

---

Improving Generalization Beyond Training Data with Compositional Generalization in Large Language Models

Опубликована: Май 20, 2024

Enhancing compositional generalization in language models addresses a crucial challenge natural processing, significantly improving their ability to understand and generate novel combinations of known concepts. The investigation utilized the Mistral 7x8B model, employing advanced data augmentation refined training methodologies enhance performance. By incorporating diverse challenging compositions during training, model demonstrated substantial gains standard evaluation metrics, including accuracy, precision, recall, F1-score. Specialized metrics such as accuracy contextual coherence also showed marked improvement, reflecting model's enhanced capacity correct contextually relevant outputs when faced with compositions. study further highlighted significant reduction hallucination rates, underscoring increased logical consistency factual accuracy. This was statistically significant, indicating robust enhancement Qualitative analysis corroborated these findings, revealing more coherent narratives accurate information retrieval generated responses. These improvements are particularly important for real-world applications where reliability appropriateness essential. comprehensive effectiveness proposed techniques, providing valuable insights into underlying mechanisms that contribute improved findings underscore importance iterative experimentation validation refining architectures techniques. advancing capabilities models, this research contributes development robust, flexible, reliable AI systems capable handling broader range linguistic tasks greater understanding.

Язык: Английский

---

Synergistic Integration of Large Language Models and Cognitive Architectures for Robust AI: An Exploratory Analysis

Proceedings of the AAAI Symposium Series, Год журнала: 2024, Номер 2(1), С. 396 - 405

Опубликована: Янв. 22, 2024

This paper explores the integration of two AI subdisciplines employed in development artificial agents that exhibit intelligent behavior: Large Language Models (LLMs) and Cognitive Architectures (CAs). We present three approaches, each grounded theoretical models supported by preliminary empirical evidence. The modular approach, which introduces four with varying degrees integration, makes use chain-of-thought prompting, draws inspiration from augmented LLMs, Common Model Cognition, simulation theory cognition. agency motivated Society Mind LIDA cognitive architecture, proposes formation agent collections interact at micro macro levels, driven either LLMs or symbolic components. neuro-symbolic takes CLARION a model where bottom-up learning extracts representations an LLM layer top-down guidance utilizes to direct prompt engineering layer. These approaches aim harness strengths both CAs, while mitigating their weaknesses, thereby advancing more robust systems. discuss tradeoffs challenges associated approach.

Язык: Английский

---

Can large language models help augment English psycholinguistic datasets?

Behavior Research Methods, Год журнала: 2024, Номер 56(6), С. 6082 - 6100

Опубликована: Янв. 23, 2024

Research on language and cognition relies extensively psycholinguistic datasets or "norms". These contain judgments of lexical properties like concreteness age acquisition, can be used to norm experimental stimuli, discover empirical relationships in the lexicon, stress-test computational models. However, collecting human at scale is both time-consuming expensive. This issue compounded for multi-dimensional norms those incorporating context. The current work asks whether large models (LLMs) leveraged augment creation large, English. I use GPT-4 collect multiple kinds semantic (e.g., word similarity, contextualized sensorimotor associations, iconicity) English words compare these against "gold standard". For each dataset, find that GPT-4's are positively correlated with judgments, some cases rivaling even exceeding average inter-annotator agreement displayed by humans. then identify several ways which LLM-generated differ from human-generated systematically. also perform "substitution analyses", demonstrate replacing a statistical model does not change sign parameter estimates (though select cases, there significant changes their magnitude). conclude discussing considerations limitations associated general, including concerns data contamination, choice LLM, external validity, construct quality. Additionally, all (over 30,000 total) made available online further analysis.

Язык: Английский

---

The new paradigm in machine learning – foundation models, large language models and beyond: a primer for physicians

Internal Medicine Journal, Год журнала: 2024, Номер 54(5), С. 705 - 715

Опубликована: Май 1, 2024

Abstract Foundation machine learning models are deep capable of performing many different tasks using data modalities such as text, audio, images and video. They represent a major shift from traditional task‐specific prediction models. Large language (LLM), brought to wide public prominence in the form ChatGPT, text‐based foundational that have potential transform medicine by enabling automation range tasks, including writing discharge summaries, answering patients questions assisting clinical decision‐making. However, not without risk can potentially cause harm if their development, evaluation use devoid proper scrutiny. This narrative review describes types LLM, emerging applications limitations bias likely future translation into practice.

Язык: Английский

---

Nanjing Yunjin intelligent question-answering system based on knowledge graphs and retrieval augmented generation technology

Heritage Science, Год журнала: 2024, Номер 12(1)

Опубликована: Апрель 9, 2024

Abstract Nanjing Yunjin, a traditional Chinese silk weaving craft, is celebrated globally for its unique local characteristics and exquisite workmanship, forming an integral part of the world's intangible cultural heritage. However, with advancement information technology, experiential knowledge Yunjin production process predominantly stored in text format. As highly specialized vertical domain, this not readily convert into usable data. Previous studies on graph-based Question-Answering System have partially addressed issue. graphs need to be constantly updated rely predefined entities relationship types. Faced ambiguous or complex natural language problems, graph retrieval faces some challenges. Therefore, study proposes that integrates Knowledge Graphs Retrieval Augmented Generation techniques. In system, ROBERTA model first utilized vectorize textual information, delving deep semantics unveil profound connotations. Additionally, FAISS vector database employed efficient storage achieving semantic match between questions answers. Ultimately, related results are fed Large Language Model enhanced generation, aiming more accurate generation outcomes improving interpretability logic System. This research merges technologies like embedding, vectorized retrieval, overcome limitations graphs-based terms updating, dependency types, understanding. implementation testing shown Intelligent System, constructed basis Generation, possesses broader base considers context, resolving issues polysemy, vague language, sentence ambiguity, efficiently accurately generates answers queries. significantly facilitates utilization knowledge, providing paradigm constructing other heritages, holds substantial theoretical practical significance exploration discovery structure human heritage, promoting inheritance protection.

Язык: Английский

---