Unverified50% confidenceSolutionExact time
处理分子任务时应使用骨架切分（Scaffold Split）而非随机切分，否则模型会因数据泄露导致效果虚高
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverified当co-folding技术跨越准确性阈值后，扩散模型将从被动筛选工具转变为主动生成理想结合特性全新分子的设计引擎69% similarUnverified扩散模型可对整个序列并行去噪，而非像自回归模型那样逐字推进，理论上能提升生成吞吐量64% similarUnverifiedChain-of-Thought（思维链）和 Tree-of-Thoughts（思维树）框架通过引导模型将大目标拆解为原子化推理步骤来降低幻觉发生概率62% similarUnverifiedSMOTE算法通过在少数类样本之间插值生成合成样本来实现过采样61% similarUnverified在污染基准上模型可直接从参数记忆提取答案而非依赖上下文检索，使报告的无损压缩结论失去意义60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/914906API
curl https://kongchang.com/api/v1/knowledge/claims/914906MCP
get_claim(id=914906)