- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper studies corpus-side detection of benchmark contamination, where benchmark items or their answers appear in a training corpus. The authors argue that existing methods rely on long exact matches or fixed retrieval units, and so can miss rewritten content or passages that sit inside long documents. They propose SpanAudit, which uses each benchmark item as a query and searches corpus spans of any length using character 4-gram MinHash with banded LSH, while sharing computation across overlapping spans to avoid enumerating all of them. Retrieved candidates are merged and localized into evidence spans, which are then judged by LLMs. In controlled experiments with planted rewrites and increasingly long documents, SpanAudit reaches the highest or tied-highest recall in every setting, with competitive latency. Audits of three previously decontaminated instruction corpora also find remaining contamination, much of it including answers or reasoning.
Critical strengths
1. A solution to a known problem. It is well known that exact n-gram matching breaks under small rewrites, and that fixed units such as paragraphs or windows can miss passages that are shorter, longer, or split across boundaries. The paper addresses both issues at once by searching spans of any length, and does so without the cost of enumerating all possible spans.
2. Good results on long documents. SpanAudit keeps recall near 99.8% as the surrounding context grows from 512 to 4,096 tokens per side, while the embedding-based baseline drops from 97.89% to 72.53% (Tab. 2). It also runs several times faster than both Infini-gram and LLM-Decontaminator in this setting, which matters for corpus-scale audits.
3. Transparency and inspectable outputs. The authors release code, data, prompts, and the retrieved evidence, so the results can be checked and reproduced.
Critical weaknesses
1. The paper measures retrieval, not contamination, and does not use ground truth where it matters. This is my strongest concern. The paper itself states that SpanAudit "returns localized evidence rather than final contamination labels"; however, its results are framed as contamination detection. In the controlled experiments, the planted positions are known, but recall is measured only at document level: "the planted pair is recovered if any retained evidence belongs to its carrier document." Localization is a central claim, yet it is never evaluated against the planted spans, and no precision is reported. In the real-corpus audits, "confirmed" means two LLM judges agreed, with no human validation, and the second judge "does not recover first-stage false negatives." Exclusive evidence is defined relative to the union of the five methods, so contamination missed by all of them is invisible. Without ground truth, it remains unclear whether the retrieved evidence reflects actual contamination or merely lexical overlap.
2. The controlled benchmark is built around the proposed method. The rewrites are constructed to hit Jaccard targets "under token-set Jaccard computed after the same normalization and character 4-gram tokenization used by SPANAUDIT," with the hardest level at [0.4, 0.6). Since SpanAudit uses the same measure with a threshold of τ = 0.4, every planted rewrite is guaranteed to lie at or above its threshold. This setup favors SpanAudit and may explain why it achieves the highest recall in every setting.
3. The rewrites may not reflect realistic paraphrasing. Sec. 5.1 states that hard examples "extensively rephrase and restructure the item." Yet the only example shown (Tab. 8) builds the moderate and hard versions by inserting noise tokens, such as "Write a(ignore) very angry(ignore) letter," and "Remove every parenthesized boundary marker before reading." This is closer to an obfuscation attack than to a paraphrase. It would help to know how many rewrites follow this pattern, and how the results change on natural paraphrases.
4. Some claims are not supported by their citations. The introduction states that corpus-side decontamination "removes, downweights, or flags suspected documents before training or evaluation," citing [1] and [2]. However, [1] attempts to remove overlapping training text (only partly, due to a bug) and then flags benchmark examples as clean or contaminated, while [2] only analyzes the contamination level of evaluation samples after training. Neither paper downweights or flags training documents. The later claim of policies "ranging from recording or downweighting suspected exposure to excluding it" is given without any citation.
References
[1] Language Models are Few-Shot Learners (Brown et al., NeurIPS 2020)
[2] Llama 2: Open Foundation and Fine-Tuned Chat Models (Touvron et al., arXiv 2023)
Questions for authors
-
How does benchmark contamination appear in post-training corpora? The introduction defines it as a case "in which benchmark-specific information appears in a pretraining or post-training corpus before evaluation," but post-training data is usually curated from selected tasks rather than scraped from the web.
-
All planted rewrites have a Jaccard similarity of at least 0.4, which matches SpanAudit's threshold of τ = 0.4. How does SpanAudit perform on rewrites with Jaccard below 0.4, for example in a [0.2, 0.4) range, compared with the other methods?