--- name: design-baselines-and-ablations description: >- Design baseline and ablation plans that test a method's central claims and isolate which components matter. Use when a paper has a method sketch, claims, and research area but needs comparison structure, component ablations, and expected findings. Output is an experiment design plan, not a list of random benchmark names. metadata: owner_model: ParadoxGPT-Designer-4B skill_family: designer version: "0.1" output_type: generation default_language: zh trigger_keywords: [baseline, ablation, component, experiment design, 对比实验, 消融, 基线] required_inputs: [method_summary, core_components, claims] optional_inputs: [research_area, related_work, constraints, current_experiments] handoff_to: [map-claims-to-experiments, design-diagnostic-analysis, identify-fatal-concerns] --- # Design Baselines And Ablations ## Purpose 为方法型 claim 设计 baseline 和 ablation,使实验能回答“比什么强”“为什么强”“哪个组件必要”。这个 skill 不负责随机挑 dataset 名,而是确定对照结构和消融逻辑。 ## When to Use - 方法已有核心组件,但实验不知道该怎么对比。 - reviewer 可能质疑增益来自 trivial baseline 或某个组件无效。 - `map-claims-to-experiments` 发现 baseline / ablation 缺口。 ## Do Not Use - 还没有明确 claims → 先用 `map-claims-to-experiments`。 - 想解释机制为什么有效 → 用 `design-diagnostic-analysis`。 - 只想判断是否 fatal → 用 `identify-fatal-concerns`。 ## Required Inputs - `method_summary`: 方法概要。 - `core_components`: 方法核心组件清单。 - `claims`: 方法主张。 - `research_area`: optional,领域或任务类型。 - `related_work`: optional,相关方法类别。 - `constraints`: optional,算力/时间/数据限制。 ## Output Contract 输出必须至少包含: 1. **baseline_plan** - baseline 类型、比较目的、必须控制的变量。 2. **ablation_plan** - 每个核心组件如何消融、要检验什么。 3. **expected_findings** - 若 claim 成立应观察到什么。 4. **risk_if_missing** - 缺少某项实验会被 reviewer 如何攻击。 5. **minimal_viable_experiment_set** - 资源有限时的最小实验集。 ## Procedure 1. 先把 claims 按 effectiveness、component necessity、efficiency、robustness 分类。 2. 为 effectiveness claim 选择 baseline 类型:strong prior method、simple heuristic、same-budget variant、oracle/upper-bound 等。 3. 为每个 core_component 设计去除、替换或冻结式 ablation。 4. 判断需要控制的变量:compute、data、model size、training budget、prompt length。 5. 给出最小实验集和缺失风险。 ## Quality Bar 一个好的输出必须: - baseline 与 claim 对齐,不是越多越好。 - ablation 能隔离组件贡献,避免同时改变多个变量。 - expected_findings 能区分 claim 成立/不成立。 - minimal set 能在 constraints 下执行。 一个差的输出: - 只列“加 SOTA baseline、更多 ablation”。 - 没有说明每个 baseline 检验什么。 - 消融设计改变太多因素,无法归因。 ## Failure Modes - **baseline 过弱**:赢了也不能证明贡献。 - **baseline 过强但不公平**:比较预算或信息不一致。 - **消融不可解释**:删除组件导致系统无法运行,却被当作证据。 - **忽略 compute/data 控制**:增益可能来自资源而非方法。 ## Handoff - 如果 claim 到实验映射还不清 → `map-claims-to-experiments` - 如果 ablation 后仍无法解释机制 → `design-diagnostic-analysis` - 如果缺失实验可能导致 reject → `identify-fatal-concerns` ## Example Input: ``` method_summary: interaction scaling agent with planner and verifier core_components: planner, verifier, multi-round interaction budget claims: higher long-horizon success; gains come from interaction rounds ``` Output: ``` baseline_plan: B1 same model, single-round generation: 检验多轮是否必要。 B2 same token budget, fewer rounds but longer responses: 控制总生成量。 B3 prior memory/tool agent: 检验相对已有 agent 的收益。 ablation_plan: A1 remove verifier: 检验纠错机制是否贡献。 A2 fixed vs scaled interaction rounds: 检验 scaling 曲线。 A3 planner-only / verifier-only: 隔离组件作用。 expected_findings: - 如果 interaction claim 成立,same token budget 下多轮应优于单轮长生成。 risk_if_missing: - 无 B2 时 reviewer 会认为收益来自更多 tokens。 minimal_viable_experiment_set: - B1 + B2 + A2。 ```