Efficient Data Selection for Multimodal Models via Incremental Optimization Utility

Researchers introduce One-Step-Train (OST), a new data selection framework for Large Multimodal Models that uses incremental optimization to identify high-quality training samples. The method reduces computational costs by 43% while outperforming existing approaches like LLM-as-a-Judge, demonstrating significant efficiency gains in multimodal model training.

The development of OST addresses a critical bottleneck in scaling Large Multimodal Models: the quality-quantity trade-off in synthetic data. As LMMs become increasingly resource-intensive, the ability to train effectively on smaller, curated datasets directly impacts their commercial viability and accessibility. OST's core innovation lies in reformulating data selection as an optimization utility ranking problem rather than relying on semantic heuristics, which typically require expensive LLM inference passes. This computational efficiency breakthrough matters because it lowers barriers to entry for organizations developing multimodal AI systems.

The research context reflects broader industry trends where data efficiency has become as important as raw model scale. Previous methods like LLM-as-a-Judge provided effective filtering but at prohibitive cost. OST's use of lightweight proxy models for marginal utility estimation represents an elegant architectural solution that maintains performance while reducing overhead. The experimental validation across Qwen series models on mathematical reasoning tasks provides credible benchmarking evidence.

For the AI development community, these results have immediate practical implications. The ability to achieve 5.6-point performance gains with 20% of data while reducing total training time by 17% creates tangible economic incentives to adopt optimization-based selection methods. Additionally, OST's demonstrated capability to identify and filter toxic samples addresses a persistent challenge in complex reasoning tasks where noise causes performance degradation. This directly benefits developers building commercial multimodal systems seeking cost-efficient scaling strategies without sacrificing output quality.

• →OST reduces training costs by 43% while outperforming LLM-as-a-Judge baseline by 1.8 points on multimodal reasoning tasks
• →Using only the top-20% data subset achieves 5.6-point gains over existing filtering methods under fixed compute budgets
• →The framework uses lightweight proxy models to estimate marginal utility rather than expensive semantic heuristics
• →OST effectively identifies and filters toxic samples, reversing negative transfer in complex reasoning tasks
• →Pareto-optimal efficiency gains make the method commercially viable for scaling multimodal model development