Converging Paradigms: The Synergy of Symbolic and Connectionist AI in LLM-Empowered Autonomous Agents

Traditionally considered distinct paradigms, connectionist AI focuses on neural networks, while symbolic AI emphasizes symbolic representation and logic. Recent advancements in large language models (LLMs), exemplified by ChatGPT and GPT-4, highlight the potential of connectionist architectures in handling human language as a form of symbols. The study argues that LLM-empowered Autonomous Agents (LAAs) embody this paradigm convergence. By utilizing LLMs for text-based knowledge modeling and representation, LAAs integrate neuro-symbolic AI principles, showcasing enhanced reasoning and decision-making capabilities. Comparing LAAs with Knowledge Graphs within the neuro-symbolic AI theme highlights the unique strengths of LAAs in mimicking human-like reasoning processes, scaling effectively with large datasets, and leveraging in-context samples without explicit retraining. The research underscores promising avenues in neurovector-symbolic integration, instructional encoding, and implicit reasoning, aimed at further enhancing LAA capabilities.

• Historical Context of Technology: This article provides an in-depth examination of the historical debate between connectionism and symbolism, contextualizing modern AI developments and highlighting the strengths of each approach. We present recent advancements in LLMs with Knowledge Graphs (KGs) [11] as references, discussing these techniques from the perspectives of symbolic, connectionist, and neuro-symbolic AI. The article also showcases the transformative impact of these techniques on knowledge modeling, acquisition, representation, and reasoning.

• Convergence of Paradigms: This article highlights the convergence of symbolic and connectionist approaches in developing LAAs, emphasizing their enhanced reasoning, decision-making, and efficiency. By contrasting LAAs with Knowledge Graphs (KGs) within neuro-symbolic AI, we examine distinct patterns and functionalities. While both integrate symbolic and neural methodologies, LAAs demonstrate unique advantages over KGs: (1) analogizing human reasoning with agentic workflows and various prompting techniques [12], [13], (2) scaling effectively on large datasets, adapting to in-context samples, and leveraging the emergent abilities of LLMs. These strengths drive the surge of a new wave of neuro-symbolic AI [14].

• Future Directions: The article highlights the trend of converging paradigms and current limitations of LAAs, pointing to two promising future directions. First is the development of neuro-vector-symbolic architectures, which integrate vector manipulation to enhance the reasoning capabilities of agents. Second is the approach known as program-proof-of-thoughts (P2oT) prompting. This involves breaking down complex reasoning processes into verifiable propositions, utilizing program proof languages (such as Dafny) for structured verification. It aims to provide rigorous reasoning by modeling propositions, integrating with theorem provers, and focusing on applications in specific domains.

Each approach has its limitations: connectionist AI is criticized for its black-box nature and lack of interpretability [20], while symbolic AI faced challenges with the labor-intensive knowledge acquisition process [21] and its limited adaptability [22]. Historical debates between figures, such as Yann LeCun, Yoshua Bengio, and Gary Marcus, have underscored these limitations [23]. However, the integration of both paradigms has led to robust hybrid models, combining neural networks’ pattern recognition with symbolic systems’ interpretability and logical reasoning [24].

RDF standardized data interchange on the web using triples (subject, predicate, object) for seamless data integration and interoperability [27]. This movement established the Semantic Web, aiming for a more intelligent and interconnected web [28]. Early adopters used RDF to build schemas and taxonomies, forming the basics of modern knowledge graphs [29]. As the field matured, the focus shifted towards capturing complex relationships and domain-specific knowledge. Ontologies, formal specifications of concepts and relationships, provided a framework for annotating and interlinking data,

GNNs adeptly leverage the graph structure for advanced pattern recognition and complex predictions. They excel in tasks such as node classification, link prediction, and the extraction of hidden patterns from graph-structured data [33]. This paradigm shift towards neural networks marks a convergence with modern machine learning techniques, enabling more nuanced and scalable interpretations of often massive and intricate datasets. The ability of GNNs to embed nodes and entire graphs numerically has significantly enhanced the computational handling of knowledge graphs

In general, every LLM undergoes a two-stage training process: pre-training and fine-tuning. Pre-training involves adjusting model parameters based on the statistical properties of a large text corpus, enabling an understanding of syntax, semantics, and linguistic nuances [25]. Fine-tuning then adapts the pre-trained model to specific tasks or domains using a smaller, task-specific dataset, optimizing performance for particular applications [44]. To ensure LLMs follow human’s instructions, align with human values and exhibit desired behaviors, instruction tuning and reinforcement learning from human feedback (RLHF) have been proposed on top of finetuning [45].

Central to the design of an agent is its neural sub-system– an LLM, which functions as the core controller or coordinator. The LLM orchestrates with the agent’s symbolic subsystem and external tools, including a planning and reasoning component for task decomposition and self-reflection, memory (both short-term and long-term), and a tool-use component that allows access to external information and functionalities.

• Agentic Workflow: An agentic workflow combines planning, reasoning, memory management, tool integration, and user interfaces with LLMs. Frameworks, such as LangChain [58] and LlamaIndex [59], help design these workflows.

• Planner and Reasoner: Advanced techniques such as chain-of-thought and tree-of-thought prompting [60] break down tasks into sub-tasks, with self-reflection allowing agents to critique and refine outputs [61].

• Memory Management: Incorporates short-term memory for context and long-term memory using external storage, such as vector databases, enabling efficient information retrieval and enhanced reasoning [62], [63].

• Tool-Use & Natural Language Interface (NLI) Integration: Agents can access external tools, APIs, and models, deciding when and how to utilize them based on task goals [64], [65]. In addition, An effective NLI interprets user requests and communicates actions [66]. Techniques, such as ReAct and MRKL, provide structured interaction steps (thought, action, action input, observation) [67], [68].

Symbolic Modeling and Neural Representation:

Classic symbolic AI represents knowledge using abstractions and symbols, utilizing explicit symbolic modeling such as rules and relationships to perform reasoning [70]. This approach typically involves well-defined logic and structured knowledge bases, enabling systems to behave based on pre-defined rules. In contrast, LAAs, driven by language models, represent knowledge in a more distributed and implicit manner. Instead of relying on explicit symbols and rules, these agents leverage vast amounts of corpus and self-supervised pre-training on language models to infer patterns and relationships from raw text [25]. The knowledge is embedded within the weights of LLMs,

Given a complex goal requiring multiple steps to achieve, existing agent technologies either harness symbolic AI to systematically explore the space of potential actions or employ reinforcement learning to optimize the trajectory of these actions, efficiently partitioning complex tasks into manageable subtasks [54]. Within a LLM-empowered agent, the Chain-of- Thought (CoT) method

Within a LLM-empowered agent, few-shot in-context learning (ICL) has been proposed to utilize given examples into a prompt to generate appropriate responses that solve problems without explicit re-training the LLM [73]. This approach mimics the case-based reasoning, a fundamental concept in symbolic AI, by leveraging explicit knowledge and experiences to tackle new problems.

d) Neuro-symbolic Integration Driven by Emergent Abilities: The emergent abilities of LLMs, such as contextual understanding, sequential reasoning, goal reformulation, and task decomposition, are surged by over-parameterized architectures and extensive pre-training corpora [46]. Combining well-designed rules with the emergent abilities of LLMs enables agents to create and follow complex workflows, known as agentic workflows. By prompting large language models with instructions like “let’s think step by step”, these models analogise human’s reasoning processes and can exhibit logical and mathematical reasoning, thereby enhancing their structured reasoning skills [12], [13]. This agentic approach allows LLMs to not only process but also proactively generate structured, logical, and adaptive reasoning pathways [56], significantly improving their problem-solving and decision-making