CORAM introduces a new approach to merging fine‑tuned models by partitioning each target weight matrix into row slices and representing each slice with its singular value decomposition in the base‑model’s SVD frame. The method performs manifold averaging of task‑specific factors and applies an amplification coefficient to counteract contraction, with the coefficient’s scale estimated from update norms and its restoration strength chosen from expert update dispersion. Across multiple model families and scales, CORAM outperforms the prior OrthoMerge technique by up to 1.35 points and matches or exceeds the strongest weight‑space baselines.
ResMerge is a new framework for merging large language models trained via reinforcement learning. It separates each model’s task vector into a leading spectral head and a residual component, finding that both parts contain valuable behavior knowledge but behave differently during merging. The method builds a stable residual backbone using Spherical Residual Consensus Adaptation and then adds a lightweight head correction module that activates only when experts agree, leading to better preservation of expert capabilities compared to existing merging baselines.
By Yandu Sun, Zhiyan Hou, Hongyan An, Weizhen Wang, Haokai Ma, Yuheng Jia, Junfeng Fang, Haiyun Guo, Jinqiao Wang
The paper introduces DiGA, a Disentangled Geometry-Aware framework for merging pretrained models. DiGA orthogonally decomposes each task vector into components tied to distinct geometric attributes, aggregates these components independently, and then recombines them, thereby preserving each component’s geometric identity. Experiments across various models, tasks, and merging methods show that DiGA improves merged-model performance and reduces capability degradation.
By Zijing Wang, Yongkang Liu, Mingyang Wang, Ercong Nie, Mengjie Zhao, Yunpu Ma, Kang Liu, Zihan Wang, Shi Feng, Daling Wang, Hinrich Sch\"utze
arXiv:2608. 07814v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) language models deliver high capacity at low per-token compute, but deploying them cheaply requires compressing their many expert weight matrices.
By Inesh Chakrabarti, Sourjya Roy, Bowen Bao, Thiago Crepaldi, Spandan Tiwari, Ashish Sirasao
arXiv:2606. 01717v1 Announce Type: new Abstract: 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.
By Minsik Choi, Geewook Kim
arXiv:2606. 19549v1 Announce Type: new Abstract: Low-rank adaptation (LoRA) makes it cheap to train many domain- and task-specific language model adapters, but whether two adapters can be merged is usually discovered only after both have been fully trained and evaluated.
By Lin Tang, Wei Zhang, Jing Li, Hongyu Chen, Ming Zhao, Yuxuan Wang