Unverified50% confidenceSolutionExact time
Label-free proxy metrics such as alignment sKL can screen out generation batches that clearly deviate from the real distribution before training.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedExisting graph-level OOD generalization work usually assumes that class labels are available during training, using supervised signals to determine which substructures are predictive, thereby suppressing spurious correlations in representation learning.78% similarUnverifiedThe study uses the ContractNLI dataset, in which manually annotated evidence spans are withheld entirely until the evaluation stage and are not visible to the model during training.77% similarUnverifiedThe signals produced by this method can flag problematic training samples that frontier model classifiers would miss.75% similarUnverifiedOnce the test set of a fixed benchmark is made public, it may be incorporated into training data, rendering the evaluation invalid.75% similarUnverified使用模型训练过程中未接触过的unseen样本可以有效避免评测污染（benchmark contamination），提高测试可信度74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102121API
curl https://kongchang.com/api/v1/knowledge/claims/1102121MCP
get_claim(id=1102121)