研究想法孵化
这里存放从论文阅读和 Survey 分析中孵化出的研究想法。每个想法文件包含：
- Hypothesis: 可证伪的核心假设
- Motivation: 知识空白 + 现有方法局限
- Proposed Method: 核心方法设计
- Experiments: 验证假设的实验设计
- Risk Analysis: 潜在失败点和应对策略
想法状态
raw: 初始想法，需要进一步调研验证validated: 已通过文献调研验证，有差异化空间prototyping: 正在原型验证archived: 放弃或合并到其他方向
当前活跃想法
| 想法 | 状态 | 来源 | 核心假设 |
|---|---|---|---|
| ForkPoint-CreditAssignment-GUI | raw | CUA-Survey | Fork-point detection for long-horizon credit assignment |
| ScaleInvariant-Grounding-GUI | raw | CUA-Survey | FPN-like scale handling for GUI grounding |
| AdversarialVerification-SelfImproving-GUI | raw | CUA-Survey | Adversarial verifier for self-improving bias correction |
| AgentFacing-WebRuntime | validated | CUA-Survey | Non-oracle runtime affordances for zero-training web agents |
| HybridVerifier-GUIRuntime | raw | GUI Environment / WeaveBench | Agent-facing cross-channel verifier to reduce reward hacking |
| EvidenceDependence-GUIGrounding | raw | VisualFLIP / GUI grounding | Action Collapse Rate for counterfactual GUI evidence dependence |
| PersonalizedSafety-CUA | raw | MyPCBench / BraveGuard | Task-scoped permission + trajectory privacy guard for personal CUA |
| CounterfactualProbe-EvolutionGate | archived (13/25) | SelfEvolvingAgents-Survey / ABot-AgentOS | |
| RetrievalMediated-MemoryMisevolution | validated (19/25) | Misevolution / Memoir | Memory reward hacking 由检索正反馈介导，改检索不洗内容即可防御（Self-Improving literature-only 方向 lead idea；机制获 2606.23075 独立命名佐证） |
| MismatchTriage-LongHorizonRecovery-GUI | developing → MismatchTriage | Reliability-Survey / TRIAGE | 最优恢复动作随 mismatch 点而变，固定策略留下可测量的 recovery-selection gap（2026-07-20 算法强化：分叉标签 = 免疫 PRM-hacking 的 counterfactual 监督；train-with-fork deploy-fork-free） |
| StateSufficiency-AmnesiaProbe-GUI | raw | EvoCUA / context management | 用遗忘探针测 GUI 任务的历史信息充分性边界 |
| SelfInitiatedFork-GUI | veto | AgentRuntimePrimitives-Survey | |
| ExecutionPriorDrift-BenignGate | raw | ExperienceSafetyRisks / GRASP | 良性经验的安全退化由执行倾向漂移中介，可用全良性探针在准入时预测并拦截（不接触 harmful probe；prior-art PARTIALLY_ADDRESSED，最近邻 FATE 2605.11882） |
| RestorationFidelity-BranchGains | veto | AgentRuntimePrimitives-Survey OP4 |