{"id":"9e49adff-0e89-488b-9d12-07fd7dfb4dc3","arxiv_id":"2502.05793","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"RefNLI exposes a reference determinacy bias in NLI models that causes high false contradiction and entailment rates when verifying claims against retrieved evidence.","lead":"This paper shows that natural language inference models, trained on data collected under the assumption that premise and hypothesis refer to the same context, often fail to detect when a claim and a retrieved evidence sentence refer to different contexts, producing many false entailment and contradiction predictions in fact verification. The authors release RefNLI, a 1,143-example diagnostic benchmark with human labels for reference ambiguity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal attribution to the RD assumption is not established: it rests on a confounded cross-dataset comparison and a Jaccard-overlap filter that is never validated against human RD judgments.","rationale":"I read the paper as making three contributions: introducing RefNLI, showing that NLI models and LLMs make many false predictions on evidence that does not share the hypothesis' context, and attributing this behavior to the reference determinacy assumption in training data plus providing a mitigation. The RefNLI benchmark is a useful diagnostic, and the false-prediction result across models is reasonably supported by Table 3, though the sampling of examples from model-positive predictions means the reported error rates are conditional on such predictions. The load-bearing weakness is the causal step: the paper claims that training with the RD assumption has a 'trickle-down effect,' but the only direct evidence is the Jaccard-based filtering experiment in Section 4.4, which does not isolate RD. Low lexical overlap is not the same as RD-dependence, and the rebalancing step further confounds the comparison. The cross-dataset evidence in Section 4.2 is explicitly acknowledged by the authors as confounded. I therefore agree with the reader's identified weakest assumption. I also note an internal inconsistency in Section 5: the text says contradiction-neutral agreement improves from ambiguous to unambiguous cases, but Table 6 shows the opposite for both SNLI (-0.61 to -0.23) and MNLI (-0.25 to -0.03); this weakens the human-disagreement contribution but is secondary to the causal concern. These issues do not invalidate the benchmark, so a conditional acceptance remains appropriate, but the paper's interpretive claims about causation need to be revised or supported by direct evidence.","tokens_in":14700,"tokens_out":6286,"duration_ms":65956,"concrete_test":"Re-run the Section 4.4 experiment with a matched-size control and direct RD labels: annotate the Jaccard-filtered training examples with the RefNLI protocol, then train four T5-large models on (1) the full data, (2) the Jaccard-filtered data as in the paper, (3) data filtered by human RD labels, and (4) an equal number of randomly removed examples. If the Jaccard filter does not outperform random deletion on RefNLI precision/AUROC, or if human-RD filtering does not outperform the Jaccard filter, the causal attribution to the RD assumption is not supported by the data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that reference determinacy in training data causes the observed false entailment/contradiction behavior on RefNLI requires a causal link, and the paper's evidence for that link is weak. Section 4.2 compares models trained on different dataset mixtures and finds ANLI, where RD is least enforced, to be most useful for contradiction detection, but the authors explicitly concede that no definitive conclusion can be drawn due to confounders. The only direct manipulation, Section 4.4, filters training examples by a token-level Jaccard similarity of 0.15. This threshold is a proxy for lexical overlap, not for reference determinacy: low-overlap entailment/contradiction pairs can be fully deterministic, while some RD-dependent pairs may share entities and have high overlap. The filtering experiment also rebalances label proportions, so the precision improvements in Table 5 could arise from removing hard examples, changing class priors, or reducing noise rather than from removing RD-dependent labels. Without validating the Jaccard heuristic against human RD judgments or comparing it with a matched control, the causal attribution is unsupported. The benchmark itself shows that models make many false predictions on irrelevant or ambiguous evidence, but that pattern is compatible with more general explanations such as domain shift, low lexical overlap, or miscalibration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper revisits the reference determinacy (RD) assumption in NLI dataset creation and argues that models trained under this assumption fail when the premise and hypothesis refer to different contexts. The authors introduce RefNLI, a benchmark of 1,143 claim-evidence pairs sampled from FEVER and VitaminC retrieval settings, labeled by expert annotators as entailment, contradiction, neutral, or ambiguous. They evaluate finetuned T5-large models trained on different NLI dataset mixtures, plus Gemini 1.0 Ultra with few-shot prompting, and report many false contradiction and entailment predictions. They also propose a Jaccard-overlap filtering heuristic to mitigate the RD bias and present a ChaosNLI-based analysis linking reference ambiguity to human disagreement patterns.","tokens_in":14872,"tokens_out":6998,"duration_ms":71844,"significance":"The central observation that sentence-pair NLI models produce confident entailment and contradiction predictions on irrelevant retrieved evidence is practically important for fact verification and other downstream applications. The paper contributes a new diagnostic benchmark, reports a reasonable inter-annotator agreement (0.83 Fleiss kappa) on a subsample, evaluates multiple training-dataset mixtures, and includes candid limitation statements. The benchmark itself is a useful resource. However, the significance of the paper depends on whether the evaluation protocol treats the Ambiguous class correctly, whether the sampling procedure supports the headline false-prediction rates, and whether the causal attribution to the RD assumption is supported by the experiments; in the current version those load-bearing points need further work.","major_comments":[{"comment":"The benchmark construction introduces a selection bias that is not accounted for in the reported statistics. The annotation sample was drawn only from pairs where the RoBERTa NLI model predicted entailment or contradiction and where the evidence did not come from the correct Wikipedia page, so the benchmark is conditional on model-positive predictions on irrelevant pages. The abstract's statement that models lead to '>80% false contradiction and >50% entailment predictions' should be framed as behavior on this filtered set, not on randomly retrieved evidence. Please report the size of the sampling pool, the distribution of model predictions on the full retrieved set, and present the RefNLI results with this conditioning made explicit.","section":"Section 3.2 / Table 3"},{"comment":"The evaluation protocol does not state how the 135 Ambiguous examples in RefNLI are treated. Table 3 reports metrics only for Contradiction, Neutral, and Entailment, and the label counts in the table sum to 1,008, not 1,143. If ambiguous examples are excluded from all metrics, the reported precision and recall numbers are computed on a subset of the benchmark and the headline false-prediction rates are not directly comparable to a 3-way NLI evaluation; if they are mapped to another label, that mapping should be justified. The paper should report a full confusion matrix over all four labels and clarify the exact evaluation denominator.","section":"Section 4.1 / Table 3"},{"comment":"The causal attribution from the RD assumption in training data to the observed model behavior is not established. Section 4.2 compares models trained on different dataset mixtures and notes that ANLI, where RD is least enforced, is most useful for contradiction detection, but the paper concedes that 'no definitive conclusion can ever be drawn here due to the existence of many other confounders.' The only direct manipulation, the Jaccard filter in Section 4.4, uses a threshold of 0.15 that is never validated against human RD judgments, and the experimental protocol also rebalances label proportions, so the precision improvements in Table 5 could result from removing hard examples, changing class priors, or reducing label noise rather than from removing RD-dependent examples. A matched control condition (e.g., random filtering of the same size, or a threshold sweep calibrated on human-annotated ambiguity) is needed before the abstract and conclusion can claim a causal role for the RD assumption.","section":"Section 4.2 / Section 4.4"},{"comment":"The interpretation of the correlation analysis appears inconsistent with the reported numbers. The paper states that a correlation of -1 indicates perfect agreement among raters on the distinction between two labels, so a more negative correlation means less confounding between labels. With that definition, the SNLI contradiction-neutral correlation moves from -0.61 for ambiguous examples to -0.23 for unambiguous examples, meaning raters actually confound contradiction and neutral more, not less, when the reference is unambiguous. The text claims the opposite ('the rater agreement between contradiction and neutral improves when we go from ambiguous to unambiguous cases'). This inversion undermines the human-disagreement contribution as currently written and requires re-analysis and correction.","section":"Section 5 / Table 6"}],"minor_comments":[{"comment":"There are several typos: 'parathenses' in the Table 3 caption, 'discrepency' in Section 4.2, 'applciations' in Section 6, and 'the the effect' in the contributions list in Section 1.","section":"Throughout"},{"comment":"The sentence '˜ 47%in SNLI, ˜ 46%in SNLI' appears to contain a typo and should read '˜47% in SNLI, ˜46% in MNLI'.","section":"Section 5"},{"comment":"The figure would be more informative if it reported the number of predictions in each bar and the number of retrieved sentences per claim; currently the reader cannot assess whether the contradiction-heavy distribution is driven by a small number of claims.","section":"Section 3.1 / Figure 1"},{"comment":"The AUROC computation is described only as 'one-label-vs-rest'; please specify the exact probability aggregation over the T5 output space and how ties or low-confidence predictions are handled.","section":"Section 4.1"},{"comment":"The annotation protocol says one author produced the initial label and another verified it, with a separate 102-example subsample used for the kappa computation; please clarify whether the kappa reflects agreement among three authors on the same 102 examples, which is how it appears in the text.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The benchmark is potentially a useful diagnostic resource, and the behavioral observation about false entailment/contradiction predictions on irrelevant evidence is interesting. The main risks are the unstated treatment of the Ambiguous class, the selective sampling procedure, the lack of a controlled test for the causal RD claim, and the apparent inversion in Table 6. If the authors fix the evaluation protocol and recalibrate the causal claims, this could become a solid contribution; in its current form, the evidence is too conditional to support the abstract's conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper builds a small, carefully annotated benchmark (RefNLI) that shows NLI models, both fine-tuned and few-shot LLMs, frequently call 'contradiction' on evidence that is simply about a different context. That result is real and worth having. The 4-way label set (entailment/contradiction/neutral/ambiguous) is a good idea, and the annotation protocol is serious: 1,143 pairs, two-author adjudication, and 0.83 Fleiss kappa on a subsample.\n\nWhat it does well: it makes a concrete, reproducible artifact. The GitHub link and the detailed sampling procedure mean others can build on this. The observation that models produce many false contradictions on irrelevant Wikipedia pages (Figure 1) is a clean demonstration of a known but underspecified problem.\n\nNow the soft spots. The causal narrative — that the reference determinacy (RD) assumption in training data causes the bias — is not established. Section 4.2 compares models trained on different dataset mixtures and finds ANLI best, but the authors themselves concede confounders. The only direct manipulation is the Jaccard filter in Section 4.4, which is a proxy for lexical overlap, not reference determinacy. Low-overlap pairs can be perfectly deterministic; high-overlap pairs can be ambiguous. The filter also rebalances labels, so the precision gains could come from removing hard examples or changing class priors. Without validating the heuristic against human RD judgments, it doesn't carry the causal weight the paper puts on it.\n\nThere is also an internal inconsistency in Section 5. Table 6 shows that on both SNLI and MNLI, the contradiction-vs-neutral correlation is more negative (i.e., raters distinguish them better) for ambiguous examples than for unambiguous ones (-0.61 vs -0.23 on SNLI; -0.25 vs -0.03 on MNLI). That is the opposite of what the text claims. The paper says agreement improves when going from ambiguous to unambiguous, but the numbers say the reverse. Either the interpretation of the correlation sign is wrong or the rows are mislabeled. This needs fixing.\n\nMinor issues: the benchmark is sampled from model-positive predictions on 'irrelevant' pages, so it is a diagnostic, not an estimate of error rates in the wild. The contradiction and entailment cells are small (66 and 37), so percentages come with wide intervals. The paper also has a couple of typos, including a duplicated 'SNLI' where MNLI is meant.\n\nWho is this for: researchers working on NLI, fact verification, and annotation artifacts. The benchmark itself is a solid contribution, and the mitigation experiment, even with its flaws, is a reasonable starting point. It deserves a serious referee, but the authors should be pushed to either soften the causal claims or validate the filtering heuristic.","headline":"Useful diagnostic benchmark, but the causal story about reference determinacy is overclaimed and the human-disagreement analysis has a table/text mismatch.","tokens_in":15453,"tokens_out":2646,"would_cite":true,"duration_ms":24722,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that NLI models trained under the reference determinacy assumption — that premise and hypothesis always describe the same context — systematically fail when applied to retrieved evidence that may concern a different…","keywords":["natural language inference","reference determinacy","fact verification","benchmark","human disagreement","entailment","contradiction","retrieval-augmented verification"],"falsifier":"Train a model on an NLI dataset re-annotated without the RD assumption (or on ANLI, where RD is least enforced) and evaluate on RefNLI; if false contradiction and entailment rates remain above 80% and 50%, the RD-assumption causal story is not supported. A more targeted test would vary the Jaccard filtering threshold in the mitigation experiment and show that the precision gain disappears at thresholds not chosen by the authors.","tokens_in":14474,"feed_emoji":"🧩","tokens_out":3887,"duration_ms":36111,"temperature":0.7,"pith_summary":"The paper argues that the reference determinacy (RD) assumption baked into most NLI dataset creation — that premise and hypothesis always describe the same context — leaves trained models unable to recognize when a retrieved evidence sentence is irrelevant to a claim. To show this, the authors build RefNLI, a benchmark of 1,143 claim-evidence pairs drawn from fact-verification data, labeled with a four-way scheme that separates true entailment/contradiction from neutral and reference-ambiguous cases. Finetuned NLI models and few-shot LLMs both misclassify these pairs, producing over 80% false contradictions and over 50% false entailments. The paper also finds that reference ambiguity explains part of the human disagreement documented in NLI annotation, and that filtering training examples that rely on the RD assumption improves downstream precision.","feed_headline":"NLI models fail when premise and claim diverge","feed_subtitle":"Trained under a same-context assumption, they label unrelated evidence as contradiction or entailment in RefNLI tests.","key_machinery":"The load-bearing object is the reference determinacy (RD) assumption — the convention that premise and hypothesis are interpreted as referring to the same context when annotators assign an NLI label. The paper's diagnostic instrument is RefNLI, a four-way labeled benchmark (entailment, contradiction, neutral, ambiguous) built by sampling claims from FEVER and VitaminC and retrieving Wikipedia evidence with BM25. The ambiguous label is the key addition: it marks cases where the premise could support or contradict the claim under some resolution of its references, so the correct three-way label cannot be determined without the RD assumption. The machinery works by measuring how often models trained under RD assign entailment or contradiction to such pairs.","core_discovery":"The central discovery is that current NLI systems carry a reference determinacy bias: because they are finetuned exclusively on premise-hypothesis pairs annotated under the assumption that both texts refer to the same context, they systematically over-predict entailment and contradiction when applied to inputs where the premise is retrieved evidence that may concern a different entity or event. On the RefNLI benchmark, every training-mixture tested shows low precision on contradiction and entailment predictions, with the majority of true cases being neutral. The bias appears across finetuned T5 models and in few-shot prompted Gemini, and it transfers to fact verification: models predict contradictions against irrelevant Wikipedia pages far more often than entailments. A simple lexical-overlap filter on training data reduces the bias, and the paper shows that human raters also disagree more on examples with reference ambiguity, linking the RD assumption to known annotation disagreements.","pith_inferences":["The paper leaves implicit that the bias likely extends beyond fact verification to any retrieval-augmented pipeline that uses NLI-style judgments, such as hallucination detection and text attribution, since those also pass retrieved snippets as premises.","A design fix the paper does not explore is adding an explicit relevance-checking head or an 'ambiguous/neutral' output category to NLI models trained for downstream use, rather than relying on lexical filtering.","The observed correlation between ambiguity and human disagreement suggests that disagreement-aware training, which models label distributions rather than single majority labels, may partly absorb the RD effect.","RefNLI could be extended to other languages and to multi-sentence evidence to test whether the bias scales with context length."],"forward_implications":["NLI models used in retrieval-based fact verification will keep producing false contradictions against irrelevant evidence unless trained or prompted to check reference determinacy.","Filtering low-lexical-overlap contradiction and entailment examples from training data improves precision on RefNLI, so dataset curation can partially mitigate the bias.","The four-way label set including 'ambiguous' gives a practical protocol for flagging claims that should not be verified from retrieved evidence.","Human disagreement on NLI labels, especially contradiction versus neutral, is partly explained by reference ambiguity rather than purely subjective judgment.","Any downstream task that feeds retrieved text into an NLI model inherits the reference determinacy bias."],"supporting_citations":[{"why":"Introduced SNLI and the reference determinacy assumption in NLI annotation, which the paper identifies as the origin of the bias.","marker":"Bowman et al., 2015"},{"why":"Provided MNLI, another RD-based dataset used in the training mixtures and disagreement analysis.","marker":"Williams et al., 2018"},{"why":"Supplied the ANLI dataset and the RoBERTa NLI mixture model used for the initial false-contradiction analysis.","marker":"Nie et al., 2020a"},{"why":"Provided FEVER claims, a source for RefNLI's claim-evidence pairs and the fact-verification setting.","marker":"Thorne et al., 2018"},{"why":"Provided VitaminC claims and a training dataset, linking NLI to fact verification with contrastive evidence.","marker":"Schuster et al., 2021"},{"why":"Supplied ChaosNLI, whose 100 human votes per example ground the paper's analysis of reference ambiguity and human disagreement.","marker":"Nie et al., 2020b"},{"why":"Documented inherent human disagreements in NLI labels, which the paper connects to reference indeterminacy.","marker":"Pavlick and Kwiatkowski, 2019"}],"fun_headline_variants":["RefNLI shows NLI models miss context shifts","NLI systems overfit to reference determinacy","Reference ambiguity trips NLI and fact-checkers","Same-context assumption breaks NLI in the wild"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the reference determinacy assumption in training data, rather than other differences between datasets such as annotation style, domain, or label balance, is what causes the observed false contradiction and entailment behavior.","fun_headline_variants_meta":{"raw":{"variants":["RefNLI shows NLI models miss context shifts","NLI systems overfit to reference determinacy","Reference ambiguity trips NLI and fact-checkers","Same-context assumption breaks NLI in the wild"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00043,"raw_usage":{"total_tokens":2194,"prompt_tokens":938,"completion_tokens":1256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1195}},"tokens_in":554,"tokens_out":1256,"duration_ms":11039,"temperature":1.0,"reasoning_tokens":1195,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:57:03.439857+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model on an NLI dataset re-annotated without the RD assumption (or on ANLI, where RD is least enforced) and evaluate on RefNLI; if false contradiction and entailment rates remain above 80% and 50%, the RD-assumption causal story is not supported. A more targeted test would vary the Jaccard filtering threshold in the mitigation experiment and show that the precision gain disappears at thresholds not chosen by the authors.","supporting_citations":[],"review_version":1}