Akses gratis untuk semua konten hanya login saja.

Jadwal Sholat

Memuat jadwal sholat…

Memuat

← All journals Search index

Ilmu Komputer & AI editorial

Open AccessOA2025

Large Language Models as General Purpose Intelligence Systems for Reasoning, Planning and Decision Making

A comprehensive review of LLM capabilities across reasoning, planning, and decision-making domains, with emphasis on architectural foundations, real-world applications, and open challenges

Feng-Yuan Zhang; Biting Wu· American journal of artificial intelligence and neural networks· 2025· DOI 10.71465/ajainn473

The core problem

The emergence of large language models (LLMs) has fundamentally transformed artificial intelligence (AI) research and applications, positioning these systems as potential candidates for general purpose intelligence. LLMs are deep neural networks trained on massive text corpora that demonstrate remarkable capabilities across diverse cognitive tasks without task-specific fine-tuning. This review examines how LLMs function as general intelligence systems, with particular emphasis on three core cognitive domains: reasoning, planning, and decision making. The authors, Feng-Yuan Zhang and Biting Wu, synthesize recent advances to assess both the promise and the limitations of LLMs as general-purpose intelligence systems. The central question addressed is whether the remarkable task-agnostic performance of LLMs constitutes genuine progress toward artificial general intelligence (AGI) or represents a set of narrow, albeit broad, capabilities that lack fundamental cognitive grounding. The review is structured around the IMRAD framework, analyzing architectural foundations, methodological innovations, empirical results across domains, and the critical challenges that remain open research prob

Innovation

The review reports significant findings across the three core cognitive domains. In reasoning, LLMs demonstrate remarkable capabilities through chain-of-thought prompting, which enables step-by-step problem decomposition and improves performance on complex arithmetic, logical, and commonsense reasoning tasks. In-context learning allows LLMs to adapt to new tasks from just a few examples without parameter updates, a form of meta-learning that emerges at scale. Emergent abilities—capabilities that are absent in smaller models but appear abruptly in larger ones—have been documented across various benchmarks, suggesting qualitative shifts in model competence as parameter counts increase. In planning, LLMs show proficiency in multi-step problem decomposition, goal-oriented task completion, and strategic action sequencing. They can generate coherent plans for tasks such as travel itinerary creation, project management, and even scientific experiment design. However, the review notes that planning coherence degrades over long horizons, with models often losing track of intermediate goals or producing logically inconsistent steps. In decision making, LLMs function effectively as autonomo

The emergence of large language models (LLMs) has fundamentally transformed artificial intelligence (AI) research and applications, positioning these systems as potential candidates for general purpose intelligence. LLMs are deep neural networks trained on massive text corpora that demonstrate remarkable capabilities across diverse cognitive tasks without task-specific fine-tuning. This review examines how LLMs function as general intelligence systems, with particular emphasis on three core cognitive domains: reasoning, planning, and decision making. The authors, Feng-Yuan Zhang and Biting Wu, synthesize recent advances to assess both the promise and the limitations of LLMs as general-purpose intelligence systems. The central question addressed is whether the remarkable task-agnostic performance of LLMs constitutes genuine progress toward artificial general intelligence (AGI) or represents a set of narrow, albeit broad, capabilities that lack fundamental cognitive grounding. The review is structured around the IMRAD framework, analyzing architectural foundations, methodological innovations, empirical results across domains, and the critical challenges that remain open research problems.

This study employs a comprehensive literature review methodology, synthesizing recent advances in LLM research across multiple disciplines. The authors analyze architectural foundations that enable LLMs to perform complex reasoning tasks, including chain-of-thought prompting (CoT), in-context learning (ICL), and emergent abilities that arise from scale. The review evaluates planning capabilities through performance on multi-step problem decomposition, goal-oriented task completion, and strategic action sequencing. Decision-making frameworks are investigated where LLMs serve as autonomous agents, policy advisors, and collaborative systems integrating human expertise with machine intelligence. The methodology includes examination of prompt engineering techniques, retrieval-augmented generation (RAG), and multimodal integration that enhance LLM capabilities for general intelligence tasks. Real-world applications spanning healthcare diagnosis, financial analysis, scientific discovery, and autonomous systems management are examined to ground the theoretical analysis in practical outcomes. Critical challenges including hallucination, reasoning consistency, computational efficiency, and ethical considerations are thoroughly discussed as part of the evaluative framework. The review does not conduct primary experiments but rather synthesizes existing empirical findings to construct a holistic assessment of LLM general intelligence capabilities.

Why it matters

The review critically analyzes the implications of these findings for the pursuit of general purpose intelligence. While LLMs exhibit significant progress, fundamental limitations remain. Causal understanding is a key weakness: LLMs excel at correlational patterns but struggle with counterfactual reasoning and interventions. Long-term planning coherence is another challenge, as models tend to drift from original goals over extended sequences. Adaptive learning—the ability to continuously update knowledge and skills from new experiences—is largely absent in static LLMs, which require retraining or fine-tuning to incorporate new information. Hallucination remains a persistent problem, where models generate plausible but incorrect information, undermining trust in high-stakes applications. Reasoning consistency is also variable; models may produce different answers to logically equivalent questions. Computational efficiency is a concern, as inference costs for large models are substantial, limiting deployment in resource-constrained environments. Ethical considerations include bias, privacy, and the potential for misuse, necessitating robust governance frameworks. The authors argue that addressing these challenges requires continued innovation in architecture design, training methodologies, and evaluation frameworks. Hybrid systems that combine LLMs with symbolic reasoning, causal inference engines, and memory mechanisms may offer a path forward. The review concludes that while LLMs represent a significant step toward general purpose intelligence, they are not yet AGI, and fundamental research is needed to bridge the gap between narrow task performance and genuine general intelligence. A conceptual framework for integrating LLMs with complementary modules is depicted below: This architecture suggests that augmenting LLMs with specialized components can mitigate some limitations, though integration complexity and scalability remain open questions.

Who should read this

CS practitioners and researchers

Opening member content…