A high evidence score can reflect text edits
ChainCheck: Near-Perfect Evidence-Sufficiency Scores Can Conceal Sensitivity to Construction Artifacts -- Paired Conditional Contrasts that Expose an Evaluation Confound · arXiv:2611.11965
In plain words
A passage set can contain all the information needed to answer a question, or have a broken link between passages. Tests of whether a scorer can tell the difference often create broken examples by editing the text, so a scorer might notice the edit instead of following the links.
ChainCheck is an audit that compares original passages, edited passages with the link intact, and edited passages with the link broken. In an exploratory HotpotQA test, every evaluated system responded to whether bridge names matched, but a simple one-string checker reached the same test’s maximum. That response alone does not show multi-step reasoning.
The research
A high score on an evidence-sufficiency test can look like proof that a system follows supporting passages, even when it may be reacting to how test examples were edited.
ChainCheck makes that distinction visible, changing what the ordinary score can support: by itself, it is not enough to conclude that a system follows an evidence chain.
ChainCheck audits evidence-sufficiency tests by separating scorers’ conditional responses to a broken passage chain and to the edits used to create it.
ChainCheck compares original passage sets with edited sets that either keep the evidence chain intact or break it, then reports separate conditional chain and edit contrasts alongside the ordinary score. A deterministic validator admits only records meeting nine string and graph checks.
In an exploratory HotpotQA control, all 11 systems responded to bridge-string consistency, but a one-string checker reached the statistic’s maximum. An artifact-only scorer also achieved a near-perfect ordinary score without reading the chain, showing why that score alone cannot establish reasoning.
The protocol has no broken-chain condition without an edit, so its chain and edit contrasts are conditional and cannot separate their interaction.
The independent HotpotQA set is the only set here with an informative clustered verdict; its findings describe that set, and some underlying records and checkpoints are not public.
What the evidence shows
- A score can hide its cause
ChainCheck is an evaluation-validity audit for tests that ask whether a scorer can distinguish a complete passage set from one with a broken evidence chain. The ordinary pair AUC compares scores for original and broken sets.
Since creating the broken set also edits its text, that score can reflect chain detection, edit detection, or both; a high score does not tell these responses apart.
- Three passage conditions
The audit adds an intact-but-edited condition to the original and broken passage sets. This allows it to compare the response to a broken chain when an edit is present with the response to an edit when the chain remains intact.
The comparisons are conditional, but they make the ambiguity in the ordinary score visible instead of treating it as a direct measure of reasoning.
- String consistency is not reasoning
In the crossed control, the systems’ responses show sensitivity to whether bridge strings agree, but the one-string reference shows that this operational response alone does not establish multi-hop reasoning.
The control therefore supports a narrower conclusion about string consistency, not semantic reasoning across passages.
- Edits remain detectable
The control also found that systems responded to edit traces: an artifact-only language-model scorer that never reads the chain reached a nominal pair AUC of 0.997 on those exploratory records. Separately, three model judges rated edited later passages as less natural.
This shows that edits were detectable and could drive the ordinary score, but does not prove that the evaluated systems relied on them.
- Independent results changed the verdict
Transfer was limited on the independent HotpotQA set. The pre-specified edit-dominated pattern held for five of six trained systems, while the reference judge’s selectivity interval included zero.
This set had an informative clustered verdict, unlike the earlier MuSiQue and 2Wiki results, so those earlier passes do not establish that the patterns generalize.
Deep
- Original paper authors
- Hyojung Han, Ryangkyung Kang, Youngeun Kim, Seung-Hun Jeon
- Submitted to arXiv
- License
- CC BY 4.0