所属模块：
M5 · NLP 与大语言模型 (NLP & Large Language Models)| 专题分类：模型合并与蒸馏 (Model Merging & Distillation)| 难度等级：Easy
一、核心一句话结论 (One-Sentence Summary)
微调得到的’任务向量’（微调权重 − 基座权重）可加减组合，实现多任务能力的合并或去除。
Task arithmetic represents domain specializations as linear vectors in parameter space, enabling zero-training model composition, multi-task merging, and capability unlearning via basic vector addition and subtraction.
二、核心考点要义 (Key Insights)
- 📌 任务向量 τ = 微调权重 − 基座权重（表示’这个任务学了什么’）
- 📌 合并：基座 + Σ λ·τ（多个任务向量相加）
- 📌 负任务向量可’遗忘’某能力（θ − λτ）
English Insights:
– Task vector definition: the parameter difference $,tau_i = theta_{text{ft}, i} – theta_{text{base}},$, isolating the specialized knowledge acquired during fine-tuning from a shared base initialization
– Vector algebraic operations: multi-task merging via linear combination $,theta_{text{merged}} = theta_{text{base}} + sum_i lambda_i tau_i,$, and targeted unlearning via negation $,-lambda tau_j,$
– Theoretical premise: operates on the empirical linearity of fine-tuning trajectories and shared loss basins across models derived from the same base checkpoint
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$tau=theta_{text{ft}}-theta_{text{base}};qquad theta_{text{merged}}=theta_{text{base}}+lambdasum_itau_i$$
数学机理：任务算术（Ilharco 等 2022）——(1) 任务向量（task vector） 定义为微调权重与基座权重之差：τ_i = θ_ft,i − θ_base；它表示’为任务 i 学到的权重变化’（可视为参数空间中的一个’方向’）。(2) 多任务合并——把多个任务向量相加（可加权）：θ_merged = θ_base + Σ_i λ_i·τ_i；这样合并后的模型同时具备多个任务的能力（无需联合训练）。(3) 能力去除（forgetting）——用负的任务向量：θ’ = θ_base − λ·τ_i 可’遗忘’任务 i 的能力（如去掉有害行为、去掉某风格）。(4) 类比（analogy）——τ_A + τ_B − τ_C 可做’任务类比’（如 ‘A 的能力 + B 的风格 − C 的风格’）。为什么可以相加——(a) 线性化假设——微调在参数空间中移动的’方向’近似正交（不同任务学到的变化方向不太冲突），故相加不太互相干扰；(b) 损失面的平坦性——微调的损失面较平坦，故’中间的插值点’仍性能良好（类似’模式连通性’现象）；(c) LoRA 的天然契合——LoRA 的增量本身是低秩矩阵，多个 LoRA 可直接相加（W + Σ B_iA_i），这是’任务算术’在 PEFT 上的自然实现。局限——(a) 干扰（interference）——若任务向量冲突（方向相反或高度相关），相加会损害性能（这就是 TIES/DARE 等方法要解决的）；(b) 缩放敏感——λ 需调（λ 太大则过拟合某任务、太小则学不到）；(c) 不是所有能力都能合并（容量有限、任务冲突）；(d) 不增加容量（合并的模型参数与基座相同，故容量受限于基座）。与’集成’的区别——合并是’单模型’（参数层面），集成是’多模型’（预测层面）；前者推理成本不变，后者 ∝ 模型数。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical Mechanism: 1. Task Vector Definition (Ilharco et al., 2022): Given a pre-trained base model $theta_{text{base}} in mathbb{R}^d$ and a model fine-tuned on task $i$, $theta_{text{ft}, i} in mathbb{R}^d$, the task vector is defined as: $$tau_i = theta_{text{ft}, i} – theta_{text{base}}$$ 2. Multi-Task Composition: Combining $M$ distinct fine-tuned capabilities into a unified model without retraining: $$theta_{text{merged}} = theta_{text{base}} + sum_{i=1}^M lambda_i tau_i$$ where $lambda_i in (0, 1]$ is a task-specific scaling hyperparameter. 3. Capability Unlearning / Forgetting: Removing an undesirable behavior (e.g., toxic generation, proprietary memorization, biased style) via negative task arithmetic: $$theta_{text{unlearned}} = theta_{text{base}} – alpha tau_{text{toxic}}$$ 4. Analogical Reasoning in Weight Space: Performing conceptual arithmetic across model weights: $$tau_{text{target}} = tau_A + tau_B – tau_C$$ (e.g., French coding assistant = Base + French Task Vector + Python Task Vector – English Conversational Bias). 5. Theoretical Foundation: Fine-tuning trajectories from a common pre-trained initialization remain within the same low-loss basin and explore nearly orthogonal parameter subspaces: $langle tau_i, tau_j rangle approx 0$ for $i neq j$, preserving individual task functionalities when superimposed.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
深度剖析与工程权衡：① ‘参数空间的方向可以加减’是核心洞察——它把’能力’视为参数空间中的’向量’，从而支持’合并’与’去除’；这是’模型可编辑性’的重要进展。② ‘LoRA + 任务算术’是天然组合——多个 LoRA 可 (a) 直接相加（合并多任务）、(b) 相减（去除某能力）、(c) 用不同的 λ 调权重；这使’适配器管理’非常灵活（无需重训）。③ ‘干扰’是主要障碍——任务向量冲突时合并会掉点；TIES（修剪小值 + 符号一致时合并）与 DARE（随机丢弃 + 重缩放）正是为解决干扰而设计。④ ‘能力去除’的安全价值——可用于’去毒’（去掉有害输出）、’去偏见’；但需注意（a）可能损害相关能力、(b) 效果有限（能力分布式存储）。⑤ ‘与联邦学习/多任务的关系’——任务算术提供了’先分任务训、再合并’的流程（避免多任务联合训练的复杂度）；这对’多团队协作’有工程价值。⑥ 面试要点——被问’模型合并是什么’，应给出’任务向量（τ = θ_ft − θ_base）+ 相加（多任务）/ 相减（去除）+ 与 LoRA 的天然契合‘与’干扰是主要障碍（TIES/DARE 解决）‘；能指出’合并 vs 集成的成本差异’是深度理解的标志。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Deep Dive & Engineering Trade-offs: ① Parameter Interference and Cross-Task Destruction: When tasks require conflicting parameter adaptations (e.g., task $A$ increases weight $W_{jk}$ while task $B$ decreases it), simple linear addition results in destructive interference and catastrophic performance degradation on both tasks. This fundamental limitation motivated advanced merging algorithms like TIES and DARE. ② Hyperparameter Sensitivity: Merging performance is highly sensitive to the scaling coefficients $lambda_i$. Over-scaling induces activation blow-ups and model incoherence, while under-scaling fails to transfer the desired capability; coefficients must be calibrated via grid search or Bayesian optimization on held-out validation splits. ③ Model Merging vs Ensemble Inference: Merged models preserve $1times$ parameter size, latency, and memory footprint during serving, whereas ensemble inference multiplies serving compute and memory linearly by model count $M$. ④ Strict Pre-training Lineage Requirement: Task arithmetic is mathematically invalid between models trained from different base weights or random initializations (e.g., Llama-3 and Mistral cannot be merged via task arithmetic because their parameter coordinates reside in disjoint, non-aligned loss basins). ⑤ Interview Strategy: Define the task vector $tau = theta_{text{ft}} – theta_{text{base}}$, present equations for multi-task addition and capability unlearning, explain the orthogonal subspace assumption, and contrast single-model serving efficiency against ensemble inference.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 认为合并总是不掉点（任务冲突时会掉）
- ⚠️ 把模型合并与集成混为一谈
English Pitfalls:
– Attempting to merge models fine-tuned from entirely distinct pre-trained base checkpoints or different architectures
– Applying uniform scaling coefficients $lambda_i = 1.0$ across all task vectors without tuning for parameter interference
– Evaluating merged models only on average aggregate scores, masking severe capability collapse on individual constituent tasks
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么任务向量可以相加？
- Why is the shared pre-trained base initialization an absolute mathematical prerequisite for task vector addition?
- 合并时为什么需要缩放 λ？
- How does catastrophic parameter interference manifest in weight space, and how do sign conflicts cause capability collapse?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
模型权重合并技术：SLERP 球面插值、Ties-Merging 与 Task Arithmetic(Model Merging: SLERP, Ties-Merging & Task Vectors) - 🗺️ 知识图谱模块：
大语言模型全景图谱
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。