TRINITY: An Evolved LLM Coordinator

2025-12-04

AI Summary

To address the limitations of static weight merging and inflexible task allocation in multi-LLM collaboration, this paper proposes a lightweight evolutionary coordination framework. It introduces a Thinker-Worker-Verifier tripartite architecture and employs Separable Covariance Matrix Adaptation Evolution Strategy (Sep-CMA-ES) to dynamically optimize cross-model role scheduling and context-enriched representation learning. Crucially, the framework avoids full-parameter fine-tuning by leveraging only a 0.6B-parameter backbone model augmented with a 10K-parameter lightweight head. Extensive evaluation demonstrates substantial improvements over both single-LLM baselines and mainstream collaborative approaches across programming, mathematical reasoning, logical inference, and knowledge-intensive tasks—achieving 86.2% on LiveCodeBench. Moreover, it exhibits strong out-of-distribution generalization, validating the efficacy and efficiency of evolutionary strategies for high-dimensional, parameter-constrained LLM coordination optimization.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Multiagent Systems: Coordination and Collaboration
• Game Theory and Economic Paradigms: Coordination and Collaboration

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Combining diverse foundation models is promising, but weight-merging is limited by mismatched architectures and closed APIs. Trinity addresses this with a lightweight coordinator that orchestrates collaboration among large language models (LLMs). The coordinator, comprising a compact language model (approximately $0.6$B parameters) and a lightweight head (approximately $10$K parameters), is optimized with an evolutionary strategy for efficient and adaptive delegation. Trinity processes queries over multiple turns, where at each turn the coordinator assigns one of three roles (Thinker, Worker, or Verifier) to a selected LLM, effectively offloading complex skill acquisition from the coordinator itself. Experiments show that Trinity consistently outperforms individual models and existing methods across coding, math, reasoning, and domain knowledge tasks, and generalizes robustly to out-of-distribution tasks. On standard benchmarks, Trinity achieves state-of-the-art results, including a score of 86.2% on LiveCodeBench. Theoretical and empirical analyses identify two main factors behind this performance: (1) the coordinator's hidden-state representations provide rich contextualization of inputs, and (2) under high dimensionality and strict budget constraints, the separable Covariance Matrix Adaptation Evolution Strategy offers advantages over reinforcement learning, imitation learning, and random search by exploiting potential block-epsilon-separability.

Problem

• Combines diverse LLMs despite mismatched architectures and closed APIs.
• Orchestrates LLM collaboration via a lightweight coordinator with evolutionary optimization.
• Assigns dynamic roles to LLMs for complex task handling across domains.

Innovation

• Lightweight coordinator orchestrates multiple LLM collaboration
• Evolutionary strategy optimizes compact model for adaptive delegation
• Role assignment mechanism offloads complex skill acquisition