Unverified50% confidenceFactExact time
Benchmark Contamination refers to a model intentionally or unintentionally being exposed to evaluation set data during training, thereby achieving inflated scores on tests.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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77% similarUnverifiedBenchmark optimization suffers from training data contamination, with many models boosting their scores by having 'seen' the test questions before.73% similarUnverified评测游戏（Benchmark Gaming）指AI公司通过选择性使用特定基准测试来最大化模型表现分数的行为，包括数据污染和针对特定评测集过度优化73% similarUnverifiedBenchmark hacking can artificially inflate benchmark scores through methods such as data contamination (test set leakage into training data), targeted training on the specific question formats of particular benchmarks, and selective reporting of the best benchmark results.71% similarUnverifiedOnce the test set of a fixed benchmark is made public, it may be incorporated into training data, rendering the evaluation invalid.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/894802API
curl https://kongchang.com/api/v1/knowledge/claims/894802MCP
get_claim(id=894802)