MERGE$^3$: Efficient Evolutionary Merging on Consumer-grade GPUs

2025-02-09

AI Summary

This work addresses the prohibitively high computational cost of evolutionary model merging on consumer-grade GPUs. We propose the first efficient multi-task/multilingual model merging framework grounded in Item Response Theory (IRT). Methodologically, we integrate IRT-based capability modeling, dataset compression, and subset distillation to construct a lightweight fitness estimator—reducing fitness evaluation cost by 50×—and design an IRT-driven evolutionary search strategy enabling cross-lingual knowledge transfer. Our key contribution is the novel application of IRT to model merging evaluation, providing theoretical interpretability and convergence guarantees. Experiments demonstrate state-of-the-art performance on multilingual and cross-lingual model merging tasks, achieving competitive results using only a single consumer GPU. The open-source framework ensures full reproducibility, substantially lowering the hardware barrier for high-quality model merging.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Evolutionary model merging enables the creation of high-performing multi-task models but remains computationally prohibitive for consumer hardware. We introduce MERGE$^3$, an efficient framework that makes evolutionary merging feasible on a single GPU by reducing fitness computation costs 50$ imes$ while preserving performance. MERGE$^3$ achieves this by Extracting a reduced dataset for evaluation, Estimating model abilities using Item Response Theory (IRT), and Evolving optimal merges via IRT-based performance estimators. Our method enables state-of-the-art multilingual and cross-lingual merging, transferring knowledge across languages with significantly lower computational overhead. We provide theoretical guarantees and an open-source library, democratizing high-quality model merging.

Problem

• Efficient evolutionary model merging
• Reduced fitness computation costs
• State-of-the-art multilingual merging

Innovation

• Efficient evolutionary merging framework
• Reduced dataset for evaluation
• IRT-based performance estimators