模型融合已成为一种实用方法,用于在不额外训练的情况下结合专用大型语言模型(LLM)的能力。在长到短(L2S)场景中,将基础模型与长链推理模型融合旨在保留推理准确性的同时减少输出长度。现有方法依赖于任务算术及其变种,隐含假设模型输出与融合系数线性变化——我们证明在L2S设置中这一假设系统性地被违反。我们提供了首个理论依据:证明融合误差被每层Hessian范数成比例的项所限制(命题1),并通过局部极值点处的Fisher-Hessian等价性,建立Fisher信息矩阵(FIM)作为此界限的可行代理。基于此理论,我们提出FIM-Merging,仅使用随机标记输入计算对角FIM(无需领域特定校准数据),并利用其分配每层融合系数。在7B L2S基准上,FIM-TIES在六个评估基准中的五个中取得最佳性能,包括在MATH500上比ACM-TIES(90.2 vs. 84.0)多6.2个点,同时无需校准数据。在1.5B基准上,FIM-TIES的平均准确率为47.3,超过之前的最佳ACM-TIES(43.3)3.9个点,同时将响应长度减少91.9%相对长CoT模型。我们的框架还提供了现有层自适应方法如ACM为何在经验上优于均匀融合的统一理论解释。
英文摘要:
Model merging has emerged as a practical approach to combine capabilities of specialized large language models (LLMs) without additional training. In the Long-to-Short (L2S) scenario, merging a base model with a long-chain-of-thought reasoning model aims to preserve reasoning accuracy while reducing output length. Existing methods rely on Task Arithmetic and its variants, which implicitly assume that model outputs vary linearly with the merging coefficient -- an assumption we show is systematically violated in L2S settings. We provide the first theoretical justification for layer-adaptive merging: we prove that merging error is bounded by a term proportional to the per-layer Hessian norm (Proposition~1), and establish that the Fisher Information Matrix (FIM) is a principled, computable proxy for this bound via the Fisher-Hessian equivalence at local optima. Building on this theory, we propose \textbf{FIM-Merging}, which computes diagonal FIM using only random token inputs (no domain-specific calibration data required) and uses it to assign per-layer merging coefficients. On the 7B L2S benchmark, FIM-TIES achieves state-of-the-art performance on five out of six evaluation benchmarks, including a \textbf{+6.2} point gain on MATH500 over ACM-TIES (90.2 vs.\ 84.0), while requiring no calibration data. On the 1.5B benchmark, FIM-TIES achieves an average accuracy of \textbf{47.3}, surpassing the previous best ACM-TIES (43.3) by \textbf{+3.9} points, while reducing average response length by \textbf{91.9\%} relative to the long-CoT model. Our framework also provides a unified theoretical explanation for why existing layer-adaptive methods such as ACM empirically outperform uniform merging.
FRISM: Fine-Grained Reasoning Injection via Subspace-Level Model Merging for Vision-Language Models
College of Future Information Technology, Fudan University, Shanghai, China(复旦大学未来信息科技学院,中国);Shanghai Innovation Institute, China(上海创新研究院,中国);The Chinese University of Hong Kong, China(香港中文大学,中国);Harbin Institute of Technology, China(哈尔滨工业大学,中国);Shanghai Artificial Intelligence Laboratory, China(上海人工智能实验室,中国);Sun Yat-Sen University, Shenzhen, China(暨南大学深圳校区,中国)
RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking Format
Institute of Image Processing and Pattern Recoginition, School of Automation and Intelligent Sensing, Shanghai Jiao Tong University(图像处理与模式识别研究所,自动化与智能感知学院,上海交通大学);The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));MoE Key Laboratory of System Control and Information Processing (Shanghai)(系统控制与信息处理国家重点实验室(上海))
Sparsity Curse: Understanding RLVR Model Parameter Space from Model Merging
Zhejiang University(浙江大学);Simon Fraser University(西蒙菲莎大学);The Chinese University of Hong Kong(香港中文大学);Zhejiang Key Lab of Accessible Perception and Intelligent Systems(浙江省可感知智能系统重点实验室)
Bridging Training and Merging Through Momentum-Aware Optimization
Swinburne Business School, Swinburne University of Technology, Australia(斯威本科技大学斯威本商学院,澳大利亚);Data Science & Big Data Lab, Universidad Pablo de Olavide, Seville, Spain(巴勃罗·德奥拉维德大学数据科学与大数据实验室,西班牙塞维利亚)
BlendRL: A Framework for Merging Symbolic and Neural Policy Learning
Technical University of Darmstadt(达姆施塔特工业大学);Eindhoven University of Technology(埃因霍温理工大学);Hessian Center for Artificial Intelligence (hessian.AI)(黑森人工智能中心);German Research Center for Artificial Intelligence (DFKI)(德国人工智能研究中心);Centre for Cognitive Science, Technical University of Darmstadt(达姆施塔特工业大学认知科学中心)
Conflict-Free Replicated Data Types for Neural Network Model Merging: A Two-Layer Architecture Enabling CRDT-Compliant Model Merging Across 26 Strategies
DeepMerge II: Building Robust Deep Learning Algorithms for Merging Galaxy Identification Across Domains
Fermi National Accelerator Laboratory(费米国家加速器实验室);University of Chicago(芝加哥大学);Argonne National Laboratory(阿贡国家实验室);Oak Ridge National Laboratory(橡树岭国家实验室);Space Telescope Science Institute(太空望远镜科学研究所);Kavli Institute for Cosmological Physics(卡夫利宇宙物理研究所)
Learning to Think: Information-Theoretic Reinforcement Fine-Tuning for LLMs
Institute of Software, Chinese Academy of Sciences(中国科学院软件研究所);University of Chinese Academy of Sciences(中国科学院大学);The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州))
MAny: Merge Anything for Multimodal Continual Instruction Tuning
College of Computer Science and Technology, National University of Defense Technology(国防科技大学计算机科学与技术学院);National Key Laboratory of Parallel and Distributed Computing, National University of Defense Technology(国防科技大学并行与分布式计算国家重点实验室);State Key Laboratory of Complex & Critical Software Environment(复杂与关键软件环境国家重点实验室);School of Computer Science, Tsinghua University(清华大学计算机科学与技术系)
AP-BMM: Approximating Capability-Cost Pareto Sets of LLMs via Asynchronous Prior-Guided Bayesian Model Merging
Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies(广东新型安全智能技术重点实验室);Institute of Cyberspace Security(网络空间安全研究院);School of Computer Science and Technology(计算机科学与技术学院)
Mediator: Memory-efficient LLM Merging with Less Parameter Conflicts and Uncertainty Based Routing
The Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));The Hong Kong University of Science and Technology(香港科技大学);Platform and Content Group, Tencent(腾讯平台与内容事业群);UCAS(中国科学院大学);Sun Yat-sen University(中山大学)