Analysing the potential of ChatGPT to support plant disease risk forecasting systems

Smart Agricultural Technology, Journal Year: 2025, Volume and Issue: unknown, P. 100824 - 100824

Published: Feb. 1, 2025

Language: Английский

---

Attention heads of large language models

Patterns, Journal Year: 2025, Volume and Issue: 6(2), P. 101176 - 101176

Published: Feb. 1, 2025

Large language models (LLMs) have demonstrated performance approaching human levels in tasks such as long-text comprehension and mathematical reasoning, but they remain black-box systems. Understanding the reasoning bottlenecks of LLMs remains a critical challenge, these limitations are deeply tied to their internal architecture. Attention heads play pivotal role thought share similarities with brain functions. In this review, we explore roles mechanisms attention help demystify processes LLMs. We first introduce four-stage framework inspired by process. Using framework, review existing research identify categorize functions specific heads. Additionally, analyze experimental methodologies used discover special further summarize relevant evaluation methods benchmarks. Finally, discuss current propose several potential future directions.

Language: Английский

---

Text understanding in GPT-4 versus humans

Royal Society Open Science, Journal Year: 2025, Volume and Issue: 12(2)

Published: Feb. 1, 2025

We examine whether a leading AI system, GPT-4, understands text as well humans do, first using well-established standardized test of discourse comprehension. On this test, GPT-4 performs slightly, but not statistically significantly, better than given the very high level human performance. Both and make correct inferences about information that is explicitly stated in text, critical understanding. Next, we use more difficult passages to determine could allow larger differences between humans. does considerably on do school university students for whom these are designed, admission tests student reading Deeper exploration GPT-4's performance material from one reveals generally accepted signatures genuine understanding, namely generalization inference.

Language: Английский

---

An AI-powered approach to the semiotic reconstruction of narratives

Entertainment Computing, Journal Year: 2024, Volume and Issue: 52, P. 100810 - 100810

Published: July 4, 2024

This article presents a novel and highly interactive process to generate natural language narratives based on our ongoing work semiotic relations, providing four criteria for composing new from existing stories. The wide applicability of this reconstruction is suggested by reputed literary scholar's deconstructive claim that can often be shown tissue previous narratives. Along, respectively, three axes – syntagmatic, paradigmatic, meronymic stories yield the combination, imitation, or expansion an iconic scene; lastly, story may emerge through reversal via antithetic consideration, i.e., adoption opposite values. Targeting casual users, we present fully operational prototype with simple user-friendly interface incorporates AI agent, namely ChatGPT. prototype, in coauthor capacity, generates context-compatible sequences events storyboard format using backward-chaining abductive reasoning (employing Stable Diffusion draw scene illustrations), conforming as much possible user's authorial instructions. extensive repertoire book movie summaries available agent obviates need manually supply laborious error-prone context specifications. A user study was conducted evaluate experience satisfaction generated preliminary findings suggest approach has potential enhance quality while offering positive experience.

Language: Английский

---

Fusing Domain-Specific Content from Large Language Models into Knowledge Graphs for Enhanced Zero Shot Object State Classification

Proceedings of the AAAI Symposium Series, Journal Year: 2024, Volume and Issue: 3(1), P. 115 - 124

Published: May 20, 2024

Domain-specific knowledge can significantly contribute to addressing a wide variety of vision tasks. However, the generation such entails considerable human labor and time costs. This study investigates potential Large Language Models (LLMs) in generating providing domain-specific information through semantic embeddings. To achieve this, an LLM is integrated into pipeline that utilizes Knowledge Graphs pre-trained vectors context Vision-based Zero-shot Object State Classification task. We thoroughly examine behavior extensive ablation study. Our findings reveal integration LLM-based embeddings, combination with general-purpose leads substantial performance improvements. Drawing insights from this study, we conduct comparative analysis against competing models, thereby highlighting state-of-the-art achieved by proposed approach.

Language: Английский

---

Taking It Easy: Off-the-Shelf Versus Fine-Tuned Supervised Modeling of Performance Appraisal Text

Organizational Research Methods, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 28, 2024

When assessing text, supervised natural language processing (NLP) models have traditionally been used to measure targeted constructs in the organizational sciences. However, these require significant resources develop. Emerging “off-the-shelf” large (LLM) offer a way evaluate without building customized models. it is unclear whether off-the-shelf LLMs accurately score and what evidence necessary infer validity. In this study, we compared validity of NLP LLM (ChatGPT-3.5 ChatGPT-4). Across six datasets thousands comments, found that produced scores were more reliable than human coders. even though not specifically developed for purpose, produce similar psychometric properties as models, with slightly less favorable properties. We connect findings broader validation considerations present decision chart guide researchers practitioners on how they can use constructs, including guidance be “transported” new contexts.

Language: Английский

---

CERN for AI: a theoretical framework for autonomous simulation-based artificial intelligence testing and alignment

European Journal of Futures Research, Journal Year: 2024, Volume and Issue: 12(1)

Published: Aug. 24, 2024

Abstract This paper explores the potential of a multidisciplinary approach to testing and aligning artificial intelligence (AI), specifically focusing on large language models (LLMs). Due rapid development wide application LLMs, challenges such as ethical alignment, controllability, predictability these emerged global risks. study investigates an innovative simulation-based multi-agent system within virtual reality framework that replicates real-world environment. The is populated by automated 'digital citizens,' simulating complex social structures interactions examine optimize AI. Application various theories from fields sociology, psychology, computer science, physics, biology, economics demonstrates possibility more human-aligned socially responsible purpose digital environment provide dynamic platform where advanced AI agents can interact make independent decisions, thereby mimicking realistic scenarios. actors in this city, operated serve primary agents, exhibiting high degrees autonomy. While shows immense potential, there are notable limitations, most significantly unpredictable nature dynamics. research endeavors contribute refinement AI, emphasizing integration social, ethical, theoretical dimensions for future research.

Language: Английский

---

Can LLMs Mimic Human-Like Mental Accounting and Behavioral Biases?

SSRN Electronic Journal, Journal Year: 2024, Volume and Issue: unknown

Published: Jan. 1, 2024

This paper delves into behavioral biases in the economic decision-making of LLMs across English, Chinese, Spanish, and French, focusing specifically on mental accounting. Our investigation comprises three distinct components. First, we analyze LLMs' arithmetic using prospect theory, revealing that Spanish French exhibit human-like value functions, while English Chinese lack loss aversion. Additionally, explore application hedonic framing, noting their alignment with human single-type outcomes but divergence mixed loss-gain scenarios. The second component evaluates influence accounting financial decision-making, particularly contexts where consumption is either concurrent or separate from transactions. findings reveal only derive utility consistently demonstrate a ability to distinguish between items intended for immediate use those future consumption. In final component, examine dynamic tend temporally segregate losses, prefer separating gains, deviating humans segregating both. indicate mimic certain aspects, significant differences persist. These insights underscore need caution when employing understand consumer preferences simulate decision-making.

Language: Английский

---

An Ai-Powered Approach to the Semiotic Reconstruction of Narratives

Published: Jan. 1, 2024

Download This Paper Open PDF in Browser Add to My Library Share: Permalink Using these links will ensure access this page indefinitely Copy URL DOI

Language: Английский

---

Exploring the Synergy of Grammar-Aware Prompt Engineering and Formal Methods for Mitigating Hallucinations in LLMs

East African Journal of Information Technology, Journal Year: 2024, Volume and Issue: 7(1), P. 188 - 201

Published: Aug. 15, 2024

Recent advancements in Artificial Intelligence (AI), particularly the advanced machine learning for Natural Language Processing (NLP) paradigm, have led to development of powerful Large Models (LLMs) capable impressive feats tasks like translation, text summarisation, generation and code generation. However, a critical challenge hindering their real-world deployment is susceptibility hallucinations, where they generate plausible looking but factually incorrect outputs. These limitations come with adverse effects, such as propagation misinformation reducing user trustworthiness related technologies, even when possess transformative potential various sectors. This study aims enhance performance LLMs by presenting new strategy that combines grammar-aware prompt engineering (GAPE) formal methods (FMs) leverage synergy LLM process logic. We argue combining linguistic principles using GAPE constructing basis structures FMs, we could improve LLM's ability analyse language, decrease ambiguity prompts, consistency output, eventually, greatly diminish hallucinations. To do this, propose collaboration between linguists AI experts while also providing specialised training emphasises precision. Additionally, suggest implementing iterative design procedures use FM continuously LLMs. By following these techniques, may create future which are more trustworthy wide range users cases reliable technologies efficient practical situations

Language: Английский

---