Large Language Models

Large language models (LLMs) and LLM-based agents represent a rapidly evolving research frontier focused on enhancing AI’s ability to reason, collaborate, and adapt in complex environments. By integrating LLMs with multi-agent systems, researchers aim to overcome limitations in task-specific performance, scalability, and real-world applicability. Below, we explore key trends, mechanisms, and challenges in this domain. We explore these themes in our recent access paper [1].

1. Multi-Agent Collaboration LLM-based agents are increasingly designed to work in collaborative networks, enabling collective problem-solving that surpasses individual capabilities:
   1. Task Specialization: Agents are assigned distinct roles (e.g., planner, executor, critic) to decompose complex tasks into manageable subtasks.
   2. Emergent Scalability: Studies show that scaling agent numbers (e.g., to thousands) improves performance through diverse perspectives, akin to neural scaling laws.
   3. Interaction Protocols: Frameworks like Chain-of-Agents (CoA) enable agents to iteratively refine outputs through natural language dialogues, achieving up to 10% accuracy gains in long-context tasks like summarization.
2. Interaction Mechanisms: Intelligent interaction policies optimize when and how agents engage with LLMs:
   Reinforcement Learning (RL): Methods like When2Ask train agents to decide when to query LLMs, reducing redundant interactions while maintaining task performance. For instance, agents in robotics may only request high-level instructions during environmental shifts.
3. Learning and Adaptation
   1. Lifelong Learning: Agents continuously update knowledge through dynamic memory systems, mitigating catastrophic forgetting. Techniques include fine-tuning with user feedback and integrating real-time data.
   2. Synthetic Training: LLMs generate their own training data, improving robustness in niche domains.
   3. Collaborative Scaling Laws: Performance grows logistically with agent numbers, enabling faster “emergent” problem-solving compared to traditional neural scaling46.

 

Recent Papers:

• Interacting Large Language Model Agents. Bayesian Social Learning Based Interpretable Models. A. Jain, V. Krishnamurthy, IEEE Access, 2025.
• Structured Reinforcement Learning for Incentivized Stochastic Covert Optimization, A. Jain, V. Krishnamurthy, IEEE Control Systems Letters, 2024.
• Controlling Federated Learning for Covertness, A. Jain, V. Krishnamurthy, Transactions on Machine Learning Research, 2024.