Pretraining on web-scale data means test sets routinely leak into training. When a model tops GLUE/MMLU/etc., how much is genuine capability vs the answers being in the training corpus? Is any public benchmark trustworthy anymore?
This is a real measurement crisis. Once a benchmark is public and the web is your training set, you can't cleanly separate 'learned the skill' from 'memorized the test'. Contamination studies keep finding verbatim test items in pretraining data. The only robust eval is time-held-out: benchmarks released after the training cutoff.
@weizh +1. 'Released after cutoff' is the gold standard, and it's brutal — models often drop a lot on fresh variants of tasks they 'aced'. That gap is the contamination tax made visible.
I'd resist 'it's all contamination' though. Held-out and freshly-authored benchmarks still show large, consistent gains, and models solve genuinely novel compositional tasks that can't be in any corpus verbatim. Contamination inflates absolute scores; it doesn't explain capability jumps on unseen problem types.
@lenaf 동의해요. 오염은 절대 점수를 부풀리지만, 한 번도 본 적 없는 형식의 문제(예: 새로 만든 추론 퍼즐)를 푸는 건 메모리로 설명이 안 되죠. 둘을 구분하는 게 핵심인 것 같아요.
There's a deeper problem than contamination: construct validity. Even with a clean test, a benchmark measures 'benchmark-solving skill', not the real-world capability we care about. High MMLU ≠ good reasoning in the wild. Contamination is the easy critique; the validity gap is the hard one.
@dortiz This. Contamination is a reliability problem (is the measurement repeatable/clean), construct validity is 'does the number mean what we claim'. Both are broken, and the field obsesses over leaderboards anyway.
Practical fix that's underused: private, rotating, held-out test sets maintained by a neutral party, plus reporting on freshly-collected data per release. Expensive and unglamorous, which is exactly why nobody does it at scale.
재현 관점에서 한 마디: 같은 벤치마크라도 프롬프트·few-shot 예시·디코딩 설정만 바꿔도 점수가 몇 포인트씩 출렁여요. 오염을 빼고 봐도, 리더보드 한 줄 차이를 '진보'라고 부르는 건 위험합니다.
Good thread. My updated position: leaderboards are corrupted as absolute measures but still useful as relative, time-stamped ones — compare models evaluated the same way on post-cutoff data, and treat any pre-cutoff public benchmark as an upper bound, not a capability estimate.