{"id":"db074bc7-e4f2-4a3b-8374-13b26706aa26","arxiv_id":"2509.01093","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"QA accuracy of open-weight LLMs drops as Wikipedia passages semantically drift from training-time content, while human accuracy stays flat.","lead":"This paper shows that large language models answer questions worse when the supporting text has been naturally edited away from the version they saw during training. It offers a new way to test whether models truly understand text versus merely recognizing familiar wording.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The similarity proxy is unvalidated: sentence-embedding similarity to same-title Wikipedia may not track the LLM's actual pretraining exposure; binning by LLM perplexity would test this.","rationale":"The reader's conditional verdict rests on the similarity proxy. I agree this is the weakest load-bearing assumption. The proposed test is direct and feasible: it replaces the external embedding with the LLM's own probability, removing the unvalidated mapping between sentence-embedding cosine and pretraining distance. Since the base model's perplexity is computed from the same weights that define pretraining, it is a natural measure of distributional closeness. If the trend persists, the paper's conclusion is robust; if it disappears, the central claim would need to be reframed as an effect of paraphrase/edit magnitude rather than pretraining divergence. This check is more informative than the alternative embedding ablation because all four embedding models are trained with similar objectives and may share the same bias. The check also controls for a possible confound: if perplexity is held constant but embedding similarity varies, any residual accuracy difference would isolate the proxy's contribution. I therefore recommend retaining the conditional verdict, pending this validation; no change to the reader's verdict is needed.","tokens_in":17786,"tokens_out":6730,"duration_ms":80047,"concrete_test":"For each edited passage in a subset (e.g., BoolQ and SQuAD 2.0), compute token-level perplexity under the corresponding base (non-instruction-tuned) LLM, using the same pretrained model that defines the training corpus (e.g., OLMo-2-1124-7B base). Bin passages by perplexity quantiles (e.g., deciles) and compute the same accuracy metric as in the paper. If accuracy monotonically declines with increasing perplexity (i.e., lower model likelihood), then the effect is not an artifact of the external sentence-embedding proxy and the central claim is supported. If the trend disappears, the proxy is the main driver. Additionally, regress accuracy on both perplexity and embedding similarity to see which better predicts performance; this isolates whether the LLM's own view of 'divergence from pretraining' reproduces the trend.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the binning variable—maximum semantic similarity between the edited passage and any same-title Wikipedia content in the training corpus, computed with all-MiniLM-L6-v2 (§2)—accurately reflects how close the passage is to what the LLM saw in pretraining. This is load-bearing because the entire monotonic trend is defined over these bins. The proxy has three unaddressed assumptions: (1) the sentence-embedding model's cosine similarity correlates with the LLM's own sensitivity; (2) the maximum over same-title content captures the relevant training exposure, even though the model saw a mixture of article versions and other web text that could contain the same information in different wording; (3) paragraph-level embedding pooling does not wash out the precise lexical/syntactic cues that the LLM relies on. The paper's infini-gram check (§4) only rules out verbatim inclusion, not semantic near-duplicates from non-Wikipedia sources. Because low-similarity bins are also likely to contain larger, more complex edits, the observed accuracy drop could be driven by general text difficulty or edit magnitude, not by 'distance from pretraining' per se. The human annotation control shows the passages remain answerable, but it does not disambiguate the mechanism. The three alternative embedding models (Appendix E) share the same training objective and thus do not break this confound. Without a model-internal measure of familiarity, the independent variable is not validated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for studying how natural, human-driven evolution of Wikipedia passages affects question answering in LLMs. It extracts edited versions of reading passages from six QA benchmarks using Wikipedia revision histories, measures each edited passage's maximum semantic similarity (via sentence embeddings) to same-title Wikipedia content in the model's publicly available pretraining corpus, bins passages by this similarity score, and correlates bin membership with LLM accuracy. The authors evaluate eight open instruction-tuned LLMs (plus base variants) and report a general decline in accuracy as similarity decreases, with a mean slope of 65.78 and mean Pearson correlation of 0.684. They include controls: filtering out instances the model can answer without the passage, human-annotation experiments showing stable human performance across bins, an infini-gram check for verbatim leakage, and robustness experiments with alternative embedding models, base (non-instruction-tuned) models, and chain-of-thought prompting. The central conclusion is that performance degrades as context passages diverge from pretraining content even when the question and all necessary information are present.","tokens_in":18095,"tokens_out":5395,"duration_ms":64262,"significance":"If the central claim holds, this is a valuable empirical contribution: it moves beyond synthetic perturbations to real human edits and connects a concretely observable phenomenon—Wikipedia text evolution—to LLM robustness and training-data proximity. The study is unusually careful in several respects: it uses transparent models with known pretraining corpora, evaluates multiple datasets and model families, includes human annotation to rule out unanswerability, and reports leakage checks. The accuracy trend is measured rather than derived from the similarity scores, so the finding is not forced by construction. However, the independent variable—the maximum semantic similarity to same-title Wikipedia content—is not validated as a model-internal measure of pretraining exposure. Since the entire monotonic trend is defined over this proxy, this is the main threat to the paper's interpretation. The framework is reusable and the paper is worth publishing after the proxy concern is addressed with additional evidence or explicit reframing of the claims.","major_comments":[{"comment":"The independent variable is the maximum all-MiniLM-L6-v2 similarity between an edited passage and any same-title Wikipedia content in the training corpus, and the abstract interprets this as 'how closely each variant aligns with content seen during pretraining.' This is load-bearing, yet the proxy is never validated against a model-internal signal of familiarity. The infini-gram check in §4/Table 1 rules out only verbatim inclusion; it does not address semantically near-duplicate content in non-Wikipedia web text or later Wikipedia snapshots. Low-similarity bins are also likely to contain larger or more complex edits, so the observed accuracy drop could reflect edit magnitude/text difficulty rather than distance from pretraining per se. The three alternative embedding models in Appendix E share the same sentence-embedding training objective and therefore do not break this confound. Pleas","section":"§2 (Semantic similarity-LLM average accuracy correlation analysis)"},{"comment":"The aggregate claim that the relationship is 'consistent and statistically grounded' is not supported by the reported statistics. The mean slope is 65.78 ± 41.55 and the mean Pearson correlation is 0.684 ± 0.291, but the per-model/per-dataset values vary widely and include negative or near-zero slopes (e.g., OLMoE-1B-7B-0125-Instruct on D(RoBERTa): slope = -18.90, Figure 5(c); OLMoE on WikiWhy: slope = 4.71, Figure 5(e)). The paper does not report confidence intervals, p-values, or per-bin sample sizes for these slopes, and the regression appears to be fit on only ten bin means, some of which may have very few instances after filtering. Please provide per-dataset/per-model significance tests (e.g., permutation tests) and a clear statement of how many of the fitted slopes are significantly positive.","section":"§4 (slope analysis and Figure 5)"},{"comment":"Two analysis choices are not justified with robustness evidence. First, the WikiWhy correctness threshold of 0.6 on all-MiniLM-L6-v2 similarity between the model response and the ground-truth explanation is hand-chosen; no sensitivity analysis is given. Second, the context-free-answerability filter removes up to 61.40% of BoolQ instances for OLMo-2-1124-13B-Instruct (Table 3). Filtering is a sensible control, but the paper does not report how the per-bin sample sizes or the monotonic trend change as a result of this filtering. These choices could affect the magnitude and even the direction of the reported trends, so they should be tested explicitly.","section":"§3 (WikiWhy correctness; Appendix B/Table 3)"}],"minor_comments":[{"comment":"Typo: 'BOOL Q' should be 'BoolQ'.","section":"Abstract"},{"comment":"Human annotation results are shown without error bars, per-bin sample sizes, or inter-annotator agreement. Since the human control is used to argue that edited passages remain answerable, these details should be reported.","section":"§4 / Figure 4 / Appendix H"},{"comment":"Table 2's column header 'A vg. per Passages' appears to be a formatting error; also the table layout makes it hard to distinguish original passage counts from edited passage counts. Please clarify the column definitions.","section":"Appendix A / Table 2"},{"comment":"For the non-OLMo models, only three datasets (SQuAD 2.0, WikiWhy, HotpotQA) are reported, while the abstract and §4 imply coverage of all six datasets across all models. Please state explicitly which models were evaluated on which datasets and why.","section":"§3 / Figure 3"},{"comment":"The infini-gram percentages in Table 1 are small but not uniformly negligible (up to 8.69% for SQuAD 2.0 on OLMo-2/OLMoE), and several cells are missing (e.g., D(RoBERTa) for dolly, HotpotQA for all models). The statement that the impact 'may not be significant' should be substantiated with a clearer threshold or additional checks.","section":"§4 (leakage discussion)"},{"comment":"The slope analysis plots linear regressions over bin midpoints but does not show the underlying data points or per-bin confidence intervals. Adding these would help assess the reliability of the slopes, especially for bins with sparse data.","section":"Appendix D / Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the unvalidated similarity proxy; this is fixable with additional analyses (e.g., perplexity-based binning, edit-size controls, or a regression with covariates) or by softening the causal wording to a claim about sentence-embedding similarity. I would not recommend rejection, because the paper's empirical design is otherwise strong and the finding is plausible and interesting. Please also ensure the authors provide enough detail on the extraction pipeline from Wu et al. (2025) so the study is reproducible without consulting that paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper establishes a new robustness failure mode—QA accuracy degrades as passages naturally drift away from the version seen in pretraining, even when the passage still contains the answer. The effect shows up across six datasets, eight transparent LLMs, base and instruct variants, and multiple sentence embedders. It's a genuine contribution to the evaluation literature, and likely to inspire replication.\n\nWhat's new: the correlation of accuracy with semantic similarity to the training Wikipedia subset, computed per model from its actual pretraining data. That's a step beyond the usual synthetic paraphrasing. The human annotation in Figure 4 is a strong control: humans don't degrade, so the drop is not just \"passages became unanswerable.\" The infini-gram verbatim leakage check is also a reasonable first pass.\n\nWhere it's soft:\n\nThe independent variable is the max semantic similarity between the edited passage and same-title Wikipedia content in the training corpus, computed with all-MiniLM-L6-v2. That score is used to bin passages and define the entire trend. It's never validated against a model-internal notion of familiarity—no perplexity, no token-level probing. The alternative embeddings in Appendix E help rule out a specific embedding artifact, but they are all sentence-transformers trained on similar objectives, so they don't break the core confound. This is the load-bearing assumption, and the paper doesn't test it. I'd want to see, at minimum, that the similarity bins track the perplexity of the target LLM on those passages.\n\nThe context-free filtering is reasonable but heavy: on BoolQ it drops up to 61% of instances for some models (Table 3). That changes the composition of the remaining set, and you could worry it interacts with the similarity bins. I don't see it killing the result, but it deserves a robustness check where those instances are kept and analyzed separately.\n\nMinor: WikiWhy correctness threshold (0.6) is hand-set, and no inter-annotator agreement is reported for the human evaluation. Also, the mean slope is 65.78 ± 41.55 with at least one negative slope (OLMoE on D(Roberta), Figure 5), so \"consistent\" is doing some work. A mixed-effects model with per-bin confidence intervals would tighten the claim.\n\nBottom line: the finding is real, the methodology is mostly transparent, and the limitations are stated honestly. The proxy validation is the one serious gap. I would send this out; the authors should be asked to add a perplexity-based validation and address the BoolQ filtering interaction.\n\nCheers.","headline":"Real robustness finding with an unvalidated independent variable; worth serious refereeing, but the similarity proxy needs a model-internal check.","tokens_in":18576,"tokens_out":2558,"would_cite":true,"duration_ms":28816,"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":"LLM accuracy falls as Wikipedia passages drift from the text seen in training, even when questions stay answerable.","keywords":["natural language understanding","question answering","context drift","semantic similarity","reading comprehension","Wikipedia revision history","LLM robustness","benchmark evaluation"],"falsifier":"Take the edited BoolQ passages in the lowest similarity bin that models answer incorrectly, have human editors rewrite them so they become semantically similar to the pretraining Wikipedia version without changing facts or answerability, and check whether model accuracy recovers to the high-bin level. If accuracy stays low, the decline is not caused by semantic distance to pretraining text; if it recovers, the similarity proxy is what carries the effect.","tokens_in":17673,"feed_emoji":"📉","tokens_out":8569,"duration_ms":92109,"temperature":0.7,"pith_summary":"This paper asks what happens when the reading passages given to a question-answering LLM are not the passages it saw during pretraining but human-edited later versions of the same Wikipedia articles. The authors build a framework that digs each passage out of Wikipedia's revision history, measures its maximum semantic similarity to same-title Wikipedia text in the model's own training corpus, and tracks accuracy across similarity bins. Across six QA datasets and eight LLMs with publicly available training data, accuracy declines as passages drift: for example, BoolQ accuracy drops from about 71% to 12.5% between the most and least similar bins, and the aggregate regression slope is 65.78 ± 41.55. Human readers do not show the same decline, so the edited passages remain answerable. The paper's conclusion is that natural context evolution is a genuine and measurable challenge to LLM reading comprehension, not a benchmark artifact.","feed_headline":"LLM accuracy falls as Wikipedia passages drift from pretraining text","feed_subtitle":"Real human edits leave questions answerable, yet model performance drops as similarity to training text declines.","key_machinery":"The load-bearing object is a semantic similarity score: for each edited passage, the maximum similarity between that passage and any same-title Wikipedia content in the LLM's training corpus, computed with a sentence-embedding model (all-MiniLM-L6-v2, with three alternative embeddings tested as ablations). The authors group these scores into ten bins, plot average accuracy per bin, and fit a linear slope. Two filters make the comparison clean: answer-preservation checking ensures edited passages still contain the answer, and a context-free question pass removes instances the model can already answer from parametric memory. The natural-perturbation pipeline producing the edited variants is ad","core_discovery":"The central claim is that a generative LLM's ability to answer questions from a passage degrades as the passage moves, through ordinary human editing rather than adversarial perturbation, away from the version of that article the model saw in pretraining. This holds even though the question and all necessary information remain present in the edited passage, and despite differences in model family, size, training corpus, and instruction tuning. The effect is specific to LLMs: human annotators' accuracy stays roughly flat across the same similarity bins. Aggregated across tasks and models, the mean slope is 65.78 ± 41.55 and the mean Pearson correlation is 0.684 ± 0.291; BoolQ and SQuAD 2.0 sh","pith_inferences":["The similarity proxy only compares edited passages with the Wikipedia subset of training data; a natural extension is to search the full training corpus for near-duplicates, which would tell whether low similarity really means far from all pretraining exposure.","Because the effect is weakest on multi-hop reasoning, the phenomenon may reflect shallow lexical or surface matching rather than a general loss of comprehension; a controlled paraphrase test that varies wording while holding content fixed could isolate that mechanism.","This suggests a deployable diagnostic: measure the similarity of incoming documents to a model's known training distribution and flag low-similarity contexts as high-risk for accuracy degradation before answering.","The framework could transfer to other evolving corpora with revision logs, such as news article archives, code documentation, or product manuals, where similar drift curves might appear."],"forward_implications":["Static QA benchmarks drawn from a single Wikipedia snapshot will overstate how well LLMs read in the wild, where articles are continually edited.","Robustness evaluations should include drifted or temporally shifted splits, ordering passages by their distance to training text rather than only by random noise or paraphrasing.","Surface-form-dependent tasks such as extractive QA and yes/no decisions are the most vulnerable, while reasoning-heavy tasks are less susceptible, suggesting different failure modes.","Instruction tuning and chain-of-thought prompting do not remove the degradation, so fixes will need training-time exposure to evolving text or explicit invariance to harmless rewording."],"supporting_citations":[{"why":"Contributes the natural perturbation pipeline adapted to curate human-edited passage variants from Wikipedia revision histories.","marker":"Wu et al. (2025)"},{"why":"Supplies the sentence-embedding models used to compute the semantic similarity scores that order the passages.","marker":"(Reimers and Gurevych, 2019)"},{"why":"Defines Dolma, the corpus used to locate same-title Wikipedia content for the OLMo models' pretraining reference.","marker":"(Soldaini et al., 2024)"},{"why":"Provides SQuAD 1.1, one of the six QA benchmarks whose passages are traced through edits.","marker":"(Rajpurkar et al., 2016)"},{"why":"Provides SQuAD 2.0 with its unanswerable-question setup, a dataset showing one of the steepest accuracy declines.","marker":"(Rajpurkar et al., 2018)"},{"why":"Provides BoolQ, the yes/no benchmark showing the largest drift-related accuracy drop.","marker":"(Clark et al., 2019)"},{"why":"Adds the adversarial human-annotation reading comprehension benchmark D(RoBERTa) used in the evaluation.","marker":"(Bartolo et al., 2020)"},{"why":"Provides WikiWhy, a cause-effect reasoning benchmark whose smaller slope supports the claim that reasoning tasks are more resistant.","marker":"(Ho et al., 2023)"},{"why":"Provides HotpotQA, a multi-hop reasoning benchmark used to show drift effects are shallower for reasoning-heavy tasks.","marker":"(Yang et al., 2018)"},{"why":"Documents the OLMo models with fully open training data, which makes the pretraining reference passages identifiable.","marker":"(Groeneveld et al., 2024)"}],"fun_headline_variants":["Natural human edits slash LLM QA accuracy","LLMs stumble as passages drift from pretraining","Passage drift, not missing info, trips LLMs","Even when answers stay, LLM accuracy drops","Real-world text edits expose LLM fragility"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The result depends on the idea that the highest semantic similarity between an edited passage and same-title Wikipedia content in the model's training corpus really measures how close the passage is to what the model saw during pretraining; if edited passages resemble other, non-Wikipedia training text, the similarity bins misclassify passages and the apparent drift effect could be an artifact of the embedding model.","fun_headline_variants_meta":{"raw":{"variants":["Natural human edits slash LLM QA accuracy","LLMs stumble as passages drift from pretraining","Passage drift, not missing info, trips LLMs","Even when answers stay, LLM accuracy drops","Real-world text edits expose LLM fragility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1065,"prompt_tokens":687,"completion_tokens":378,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":307}},"tokens_in":431,"tokens_out":378,"duration_ms":5138,"temperature":1.0,"reasoning_tokens":307,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:53:08.728305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the edited BoolQ passages in the lowest similarity bin that models answer incorrectly, have human editors rewrite them so they become semantically similar to the pretraining Wikipedia version without changing facts or answerability, and check whether model accuracy recovers to the high-bin level. If accuracy stays low, the decline is not caused by semantic distance to pretraining text; if it recovers, the similarity proxy is what carries the effect.","supporting_citations":[],"review_version":1}