Assignment

1. Research & Literature Review:

a. Fundamentals of Prompt Engineering and the ReAct Technique

Prompt Engineering:

Prompt engineering focuses on designing structured and clear inputs for

large language models (LLMs) to elicit desired responses. Effective
prompts include explicit instructions, context, examples, and structured
reasoning.

Types of Prompting:

1. Few-shot Prompting: Provide examples to guide the LLM's output.

2. Chain-of-Thought (CoT): Use step-by-step reasoning to improve LLM

accuracy on complex tasks.

3. Tree of thoughts (ToT): Tree of Thought builds on Chain-of-Thought

prompting but introduces a structured exploration of multiple
reasoning paths.

4. Prompt Function: Prompt Functions refer to modular, reusable

predefined prompts that automate prompting for specific tasks

5. Retrieval Augmented Generation (RAG): RAG improves LLM

performance by integrating retrieval of relevant external data into the
generation process

6. ReAct (Reasoning + Acting): Combines intermediate reasoning steps

(thoughts) and actions (API calls, tool usage) to achieve logical
decision-making.

Key Concept of ReAct:

ReAct allows LLMs to:

1. Reason about an input problem iteratively.

2. Act based on its reasoning (e.g., calling APIs, searching tools).

Assignment 1
 a. Respond with informed conclusions.

Example: If an LLM agent needs weather data:

Thought: "I need to check today’s weather."

Action: Query a weather API.

Result: "It is 25°C and sunny."

b. Review existing literature

Large Language Model based Multi-Agents: A Survey of Progress and

Challenges:

Evolution from Single to Multi-Agent Systems: Initially, LLMs were

employed as individual agents for tasks like planning and decision-making.
The field has since progressed to multi-agent systems, enhancing
capabilities in complex problem-solving and world simulations.

Domains and Environments: LLM-based multi-agent systems are applied

across various domains, including social simulations, gaming, and
collaborative problem-solving, effectively modeling intricate human
interactions and behaviors.

Agent Profiling and Communication: These systems define agents with

specific roles and characteristics, facilitating natural language
communication among them. This setup enables sophisticated interactions
and collaborative strategies.

Enhancing Agent Capabilities: Mechanisms such as reinforcement

learning, memory integration, and tool use are employed to improve agent
performance, adaptability, and decision-making processes.

Challenges: The paper discusses several challenges, including ensuring

coherent communication among agents, maintaining consistent behavior,
and addressing ethical considerations in autonomous decision-making.

Multi-Agent Consensus Seeking via Large Language Models:

Consensus-Seeking Task: The authors examine a scenario where each

agent possesses an initial numerical state and engages in negotiations to

Assignment 2
 reach a common consensus value. This setup serves as a simplified model
for understanding more intricate collaborative tasks.

Emergent Strategies: Without explicit instructions, LLM-driven agents

predominantly adopt an averaging strategy, adjusting their states to the
mean of all agents' current states. This behavior indicates a natural
inclination towards equitable collaboration.

Influencing Factors: The study analyzes how variables such as the

number of agents, individual agent personalities, and the network topology
impact the negotiation process and the efficiency of reaching consensus.

Application to Multi-Robot Systems: The consensus-seeking framework

is applied to a multi-robot aggregation task, demonstrating the potential of
LLM-driven agents to achieve autonomous planning and coordination
without prior task-specific training.

c. Irrelevant input impact system performance:

Case Study: Impact of Irrelevant Inputs in a Weather Forecasting System

Scenario: We submit to the system the reports of weather to the system.

Task: Filter and summarize weather-related inputs.

Challenge: The system receives a mix of relevant inputs (weather-related),

irrelevant inputs (unrelated content), and noisy inputs (gibberish or
malformed data).

Goal: Ensure irrelevant inputs do not trigger incorrect reasoning or actions in a

ReAct-based system.

2. Understanding the React Technique

a. React

React combine Reasoning and Acting:

Reasoning:

The model generates step-by-step reasoning traces to logically

explain its thought process.

Assignment 3
 These traces resemble "chain-of-thought" reasoning, where each
step builds upon the previous one to move toward a solution.

Acting:

The model simultaneously produces actions that interact with

external sources, such as APIs, databases, or simulated
environments.

Actions allow the model to gather new information or take steps

toward achieving its goals.

b. Common Challenges in multi-agent system

Scalability:

As the number of agents increases, managing their interactions

becomes exponentially complex.

Ensuring the system remains efficient and coordinated, especially in

large-scale applications like smart cities or supply chain
management, is a significant challenge.

Traditional system architectures can become overwhelmed by the

sheer number of interacting agents.

Interoperability:

Agents may be developed on different platforms or by various teams,

leading to compatibility issues.

The lack of standardized protocols can prevent agents from

effectively communicating and collaborating.

Solving this requires ensuring agents “speak the same language” to

enable seamless integration and information sharing.

Managing Complex Interactions:

As systems grow more sophisticated, the relationships and

dependencies between agents become increasingly intricate.

Coordinating actions, resolving conflicts, and maintaining system-

wide coherence while balancing agents’ individual goals is a major
challenge.

Assignment 4
 This requires technical expertise and a deep understanding of
autonomous collaboration dynamics.

Personal Opinion: Based on what I test in my previous project, if we choose a

good base llm model for agent, we wont meet some problems such as irrelevant
or noisy input triggering actions)

3. Problem Analysis
Analysis of given multi-agent system:
Let's analyze the simple multi-agent system I designed earlier where the
agents using the ReAct framework. The system has three agents:

Orchestration Agent

DataProcessing Agent

Analysis Agent

ReportGeneration Agent

Orchestration Agent will take the responsability of deciding agent to run the
task. For some big task, we will split it in some small tasks and each agents
will take care each small task.

Scenarios where irrelevant or noisy inputs trigger unintended actions:

An agent framework where reasoning (thoughts) triggers actions like API

calls, database queries, or tool usage.

For some simple question of user but the system trigger the tool calling,
leading to latency and wasted resources.

Boundaries of “relevant” and “irrelevant” messages in the system:

Relevant Messages: User inputs that align with the system's purpose,
intent, or tasks, enabling meaningful reasoning and actions.

Irrelevant Messages: Inputs that are off-topic, ambiguous, contain noise,

or do not contribute to the system's defined goals or operations.

Assignment 5
 Boundary: Relevance is determined by:

Intent Matching: Input aligns with expected tasks or queries.

Context Awareness: Input relates to the current conversation flow or

system state.

Task Contribution: Input helps achieve the system's defined

objectives.

My thought: React agent boost the accuracy when solving task a little bit and it is
easy for us to keep track of what system thinking and acting.

Assignment 6