Multi-LLM-Agent Systems: Techniques and Business Perspectives

In the era of (multi-modal) large language models, most operational processes can be reformulated and reproduced using LLM agents. The LLM agents can perceive, control, and get feedback from the environment so as to accomplish the given tasks in an autonomous manner. Besides the environment-interaction property, the LLM agents can call various external tools to ease the task completion process. The tools can be regarded as a predefined operational process with private or real-time knowledge that does not exist in the parameters of LLMs. As a natural trend of development, the tools for calling are becoming autonomous agents, thus the full intelligent system turns out to be a multi-LLM-agent system (MLAS). This paper discusses the technical and business landscapes of MLAS. Compared to the previous single-LLM-agent system, a MLAS has the advantages of i) higher potential of task-solving performance, ii) higher flexibility for system changing, iii) proprietary data preserving for each participating entity, and iv) feasibility of monetization for each entity. To support the ecosystem of MLAS, we provide a preliminary version of such MLAS protocol considering technical requirements, data privacy, and business incentives. As such, MLAS would be a practical solution to achieve artificial collective intelligence in the near future.

P3: A Policy-Driven, Pace-Adaptive, and Diversity-Promoted Framework for Optimizing LLM Training

In the rapidly evolving field of Large Language Models (LLMs), selecting high-quality data for fine-tuning is essential. This paper focuses on task-specific data pruning and selection to enhance fine-tuning. We introduce an innovative framework, termed P3, which improves LLM performance through a dynamic, adaptive training strategy. Specifically, P3 comprises the following components: (1) Policy-driven Difficulty Measurement: we begin by measuring the difficulty of data based on the model's real-time performance, transitioning from static, predefined metrics to more dynamic and adaptable ones. (2) Pace-adaptive Selection: we employ self-paced learning (SPL) to gradually select increasingly challenging data, thereby progressively enhancing the model's performance. (3) Diversity Promotion: we integrate Determinantal Point Process (DPP) into the selection process to promote the diversity within and between samples, enriching the learning process. We have validated our method on two well-known LLM datasets, APPS and MATH, designed for logical reasoning scenarios. The results show that our P3 framework significantly improves training outcomes compared to traditional methods. By fundamentally refining data selection and utilization strategies, P3 not only advances theoretical understanding of dynamic training approaches but also provides a versatile framework that can revolutionize model training in natural language processing.

TRAD: Enhancing LLM Agents with Step-Wise Thought Retrieval and Aligned Decision

Numerous large language model (LLM) agents have been built for different tasks like web navigation and online shopping due to LLM's wide knowledge and text-understanding ability. Among these works, many of them utilize in-context examples to achieve generalization without the need for fine-tuning, while few of them have considered the problem of how to select and effectively utilize these examples. Recently, methods based on trajectory-level retrieval with task meta-data and using trajectories as in-context examples have been proposed to improve the agent's overall performance in some sequential decision making tasks. However, these methods can be problematic due to plausible examples retrieved without task-specific state transition dynamics and long input with plenty of irrelevant context. In this paper, we propose a novel framework (TRAD) to address these issues. TRAD first conducts Thought Retrieval, achieving step-level demonstration selection via thought matching, leading to more helpful demonstrations and less irrelevant input noise. Then, TRAD introduces Aligned Decision, complementing retrieved demonstration steps with their previous or subsequent steps, which enables tolerance for imperfect thought and provides a choice for balance between more context and less noise. Extensive experiments on ALFWorld and Mind2Web benchmarks show that TRAD not only outperforms state-of-the-art models but also effectively helps in reducing noise and promoting generalization. Furthermore, TRAD has been deployed in real-world scenarios of a global business insurance company and improves the success rate of robotic process automation.

Alioth: A Machine Learning Based Interference-Aware Performance Monitor for Multi-Tenancy Applications in Public Cloud

Multi-tenancy in public clouds may lead to co-location interference on shared resources, which possibly results in performance degradation of cloud applications. Cloud providers want to know when such events happen and how serious the degradation is, to perform interference-aware migrations and alleviate the problem. However, virtual machines (VM) in Infrastructure-as-a-Service public clouds are black-boxes to providers, where application-level performance information cannot be acquired. This makes performance monitoring intensely challenging as cloud providers can only rely on low-level metrics such as CPU usage and hardware counters. We propose a novel machine learning framework, Alioth, to monitor the performance degradation of cloud applications. To feed the data-hungry models, we first elaborate interference generators and conduct comprehensive co-location experiments on a testbed to build Alioth-dataset which reflects the complexity and dynamicity in real-world scenarios. Then we construct Alioth by (1) augmenting features via recovering low-level metrics under no interference using denoising auto-encoders, (2) devising a transfer learning model based on domain adaptation neural network to make models generalize on test cases unseen in offline training, and (3) developing a SHAP explainer to automate feature selection and enhance model interpretability. Experiments show that Alioth achieves an average mean absolute error of 5.29% offline and 10.8% when testing on applications unseen in the training stage, outperforming the baseline methods. Alioth is also robust in signaling quality-of-service violation under dynamicity. Finally, we demonstrate a possible application of Alioth's interpretability, providing insights to benefit the decision-making of cloud operators. The dataset and code of Alioth have been released on GitHub.