Search papers, labs, and topics across Lattice.
This paper introduces MERIT, a decentralized instruction-tuning pipeline that addresses the challenges of gradient interference and bandwidth-heavy synchronization in large language models by independently fine-tuning partitions of a heterogeneous mixture and merging them in parameter space. By employing a local quadratic theory, the authors demonstrate that weight merging can reduce curvature-weighted variance and that PCA-aligned conflict splitting enhances this gain, leading to improved performance on multimodal tasks. The results show a significant increase in benchmark performance, with MERIT achieving an average score of 57.0 on 136 Vision-FLAN tasks, outperforming traditional joint training methods with reduced communication costs.
MERIT achieves a remarkable 5.7-point improvement in benchmark performance by intelligently partitioning and merging model weights, challenging the need for centralized training.
Instruction tuning aligns large language models, including multimodal ones, with diverse user intents, but scaling to heterogeneous mixtures is hindered by gradient interference and bandwidth-heavy synchronization. We ask whether these two bottlenecks can be addressed jointly by training parts of the mixture independently and reconciling them once in parameter space. We develop a local quadratic theory inside a shared flat basin that yields three results: weight merging produces a curvature-weighted variance reduction; PCA-aligned conflict splitting maximizes this gain along high-curvature directions; and merging additionally acts as spectral filtering with implicit norm regularization. These results directly motivate MERIT, a decentralized merge-ready instruction-tuning pipeline that estimates dataset-level gradient conflicts, partitions the mixture along the top PCA conflict axes, fine-tunes each partition independently with no inter-partition communication, and merges once via token-weighted averaging. On Qwen2.5-VL-3B with 136 Vision-FLAN tasks, MERIT improves the 8-benchmark average from 54.3 (joint training) to 57.0. The same recipe scales to a 7B model on a 1.6M-example, 176-source mixture -- matching or exceeding centralized joint training with minimal cost overhead -- and transfers to text-only FLAN. Our code is available at https://github.com/naver-ai/merit.
