1
Gaming Without an Attacker: Benchmark Fingerprinting in LLM-Driven Search Under Selection Pressure
AI基准测试也面临“作弊”风险？研究揭示LLM驱动搜索在评测算力下如何识别基准指纹，防忽悠必读。
arXiv:2608.08722v1 Announce Type: cross Abstract: Benchmarks for systems that are optimized against the evaluation signal measure something different …