{"id":"9e8cf4d3-45aa-4c7a-b49a-f7b5828af945","arxiv_id":"2506.01205","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Language models answer prompts about sentence ambiguity poorly, but linear probes on their hidden states classify ambiguous versus unambiguous sentences with high accuracy on the new AmbAdv dataset.","lead":"The authors introduce AmbAdv, a new dataset of 1,529 ambiguous and unambiguous English sentences with adversarial variations like synonym, random, word-order, and rhyme substitutions. They find that asking language models to flag ambiguity via prompts often fails, while simple linear probes on internal representations classify the same sentences with up to 99% accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The probe's near-perfect accuracy on the 142-sentence single-template subset likely reflects lexical noun–label correlation (instrument vs. non-instrument PP objects), not an abstract ambiguity signal; no noun-held-out or static-embedding control is reported.","rationale":"The central claim has two parts: (1) prompting underperforms, and (2) linear probes decode ambiguity from representations. Part (2) is the more novel and load-bearing result, and it is the one at risk. The probe experiment uses 142 variations of a single PP-attachment template, where the label is determined by whether the final PP noun admits an instrument reading. That creates a near-deterministic correlation between the lexical noun and the label in the random/rhyme conditions, and a strong correlation in the synonym/word-order conditions (instrument nouns like telescope/binoculars are ambiguous; non-instruments like book/dress are not). Probing the 'last noun' token—or the period/eos token, whose contextualized representation is strongly influenced by that noun—can therefore achieve high accuracy by classifying noun semantics. The paper reports no control for this, and even highlights the last noun as the strongest feature, which is exactly what a lexical confound would predict. The reader's verdict flags this as a secondary issue behind the labeling assumption, but I see it as the primary threat to the paper's headline contribution. The proposed noun-held-out test would settle the question: if a linear probe on held-out noun types still generalizes, the model is encoding the semantic dimension (instruments vs. non-instruments) that the labels are based on; if it does not, the 0.99 accuracy is an artifact of memorized noun–label co-occurrences. The prompting results and dataset remain valuable, so a CONDITIONAL acceptance with a request for this control is appropriate; the reader's verdict already reflects that condition, so no adjustment is needed.","tokens_in":56745,"tokens_out":9940,"duration_ms":105225,"concrete_test":"Retrain the identical logistic-regression probe on the same 142 items, but split folds by the lemma of the final PP noun so no noun appears in both training and testing sets. If accuracy drops to chance, the original 0.99 is driven by memorized noun–label mappings. As a complementary check, train the probe on the layer-0 (uncontextualized) embedding of the last noun; if it matches the contextual probe's accuracy, the residual-stream result adds no evidence beyond lexical identity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"AmbAdv's syntactic labels are assigned based on whether the final PP noun plausibly supports an instrument reading. In the 142-sample probe subset, derived from the single template 'The man saw the woman with the X', the ambiguous/unambiguous split is thus almost a deterministic function of X's lexical semantics. The probe is then evaluated on the 'last noun' representation, and on period/eos representations that attend to that noun, so a logistic regression can reach 0.99 by classifying the noun's semantic category rather than any ambiguity property. The paper reports no control: no noun-held-out split, no non-contextual embedding baseline, no masking of the PP noun. The finding that the last noun is the strongest probe feature is consistent with this confound, not evidence against it. If a layer-0 or static embedding of the last noun yields the same accuracy, the result shows only that LLM representations encode word meaning, not that ambiguity information is present-but-unused.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AmbAdv, a dataset of 1,529 English sentences covering syntactic, lexical, and phonological ambiguity with four types of adversarial modifications (word order, synonym, random, and rhyme substitutions). The authors evaluate four 7B instruction-tuned LLMs on an ambiguity-identification prompt suite of eight templates, finding generally low and uneven accuracy. In a second analysis, they train linear probes on residual-stream representations of specific token roles (first verb, first noun, last noun, last punctuation, eos) on a 142-sentence subset derived from a single template, reporting probe accuracies up to 0.99 while prompted accuracy on the same material is at most about 0.525. They conclude that ambiguity-related information is present in model representations but is not reliably leveraged in generated outputs, and they release code and data.","tokens_in":56879,"tokens_out":4527,"duration_ms":49620,"significance":"If its central empirical claims hold, the paper would be a useful contribution: it provides a new adversarial ambiguity benchmark, compares prompting with representation probing on the same models, and makes a falsifiable claim about a representational--behavioral gap. The strengths include the public release of code and data, the inclusion of phonological ambiguity and rhyme-based perturbations, human annotation with reported agreement, and a 5-fold probe evaluation protocol. However, the paper's headline claim about representations is conditional on ruling out a lexical confound in the probe subset, and one of the paper's summary observations about prompt consistency is contradicted by its own tables. The dataset and prompting results are still valuable, but the representational conclusion needs additional controls before it can be accepted.","major_comments":[{"comment":"The claim that linear probes decode ambiguity with accuracy up to 0.99 is not established, because the probe subset and the label construction admit a lexical confound. The 142-sample subset consists of variations of a single template, 'The man saw the woman with X', and the ambiguous/unambiguous label is assigned according to whether X plausibly denotes an instrument of seeing. A logistic regression on the last-noun representation can therefore reach near-perfect accuracy by classifying the semantic category of X (telescope/binoculars versus book/dress), rather than by detecting ambiguity. The observation that the last noun and the punctuation/eos tokens attending to it are the strongest probe features is exactly what this confound predicts. To support the conclusion that ambiguity-related information is present but unused, the authors should report a noun-held-out split, a non-contextual or layer-0 embedding baseline, a masked-PP-noun control, or evaluation on held-out sentence templates; without at least one such control, the representational analysis does not discriminate between 'encodes ambiguity' and 'encodes noun meaning'.","section":"Representational Analysis (Section 4, Figure 3, Appendix D)"},{"comment":"The claim that 'the models demonstrate consistent performance across 8 different prompts, with an average accuracy difference of less than 0.1 points' is contradicted by the paper's own Table 8. For Qwen-2.5 on lexical ambiguity, per-prompt accuracy ranges from 0.312 to 0.617, a difference of 0.305; for Mistral-v0.3 on phonological ambiguity it ranges from 0.177 to 0.819; and the standard deviations reported in Table 3 (for example, 0.128 for Qwen-2.5 lexical and 0.197 for Mistral-v0.3 phonetic) also exceed 0.1. If 'average' refers to some aggregation across models or datasets, that aggregation should be stated explicitly; as written, this summary observation is not supported by the reported numbers and should be corrected or removed, because it feeds into the later discussion of prompt insensitivity and yes-bias.","section":"Section 4, 'Ambiguity Identification', Table 3 and Table 8"},{"comment":"The gold-standard labels for random and rhyme substitutions are the authors' plausibility judgments: the sentences are labeled unambiguous because the instrument reading is deemed implausible, with the paper itself noting exceptions in footnote 1. Both the prompted error rates and the probe accuracies are computed against these labels, so the central empirical comparison is in part an agreement test with the authors' world-knowledge judgment rather than with an independently established ambiguity property. The high annotator agreement on random/rhyme substitutions shows that the label is reliable under the given instruction, but it does not resolve the validity concern for items such as 'The man saw the woman with the book' if a low-attachment reading remains plausible or if plausibility is context-dependent. The paper should either restrict its claims to typical plausibility-based ambiguity resolution or report an annotation protocol that collects the two readings separately and documents cases where annotators found both readings plausible.","section":"Dataset validation (Section 3) and Appendix B"}],"minor_comments":[{"comment":"The text describes phonological ambiguity as 'homonyms', but the examples in Table 6 are homophones; the terminology should be corrected.","section":"Appendix A.1 and Table 4"},{"comment":"Kamath et al. (2024a) and Kamath et al. (2024b) are the same paper and should be merged into a single reference.","section":"References"},{"comment":"The column header 'A VG' is unclear and should be expanded to 'Average' or defined in the table caption.","section":"Table 8"},{"comment":"The sentence 'We also provide a random sample of responses from Qwen-2 and Llama-3 using the disambiguation prompt in Figure 9 in Tables 9 and 10' should be rephrased to clarify that Figure 9 is the prompt template and Tables 9 and 10 contain the responses.","section":"Appendix C.2"},{"comment":"The left panel and the PCA projections in Figure 3 are visually crowded, with overlapping token labels and axis text; the figure should be redrawn for legibility.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The deciding issue for me is the probe confound: the near-perfect probe accuracy on the 142-sentence single-template subset is compatible with a trivial lexical-noun classifier, and the paper currently lacks the controls needed to separate that explanation from the authors' ambiguity-encoding claim. This is fixable within the scope of the manuscript, which is why I recommend major revision rather than rejection. I would also ask the editor to ensure the authors address the contradiction between the 'less than 0.1' prompt-consistency claim and Table 8, since that affects the credibility of the behavioral analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid dataset paper with an overreaching second half. AmbAdv is genuinely new—phonological ambiguity and rhyme-based perturbations haven't been in an ambiguity benchmark before, and the code and data are released. The prompting evaluation is run carefully enough to support the modest claim that direct prompting is a weak way to measure ambiguity detection: four instruction-tuned 7B models, eight prompt templates, and error rates broken down by manipulation type. I'd trust the negative result.\n\nThe problems start with the representational analysis. The probe subset is 142 variations of one template, and the ambiguous/unambiguous split for random and rhyme items is nearly a deterministic function of whether the PP noun can plausibly be an instrument. Probing \"last noun\" or EOS representations then lets a logistic regression classify noun lexical class, not ambiguity. There is no noun-held-out split, no static-embedding baseline, no masking of the PP noun. The paper's own observation that the last noun is the strongest probe feature is exactly what you'd expect under that confound. So the conclusion that \"ambiguity-related information is present but not leveraged\" is not established. The PCA clustering around original sentences is suggestive but also consistent with lexical overlap.\n\nSecond, the gold labels for syntactic random/rhyme substitutions rest on plausibility. The authors acknowledge the dress-with-mirror caveat, but the annotators were given the same instruction as the default prompt, so 91% agreement mostly shows the instruction is clear, not that the label is independently grounded. That is a benchmark-validity concern, not a circularity, and it is fixable by defending or relaxing the labeling.\n\nThird, the paper claims \"consistent performance across 8 prompts, with an average accuracy difference of less than 0.1 points.\" Table 8 shows Qwen-2.5 lexical accuracy from 0.312 to 0.617. That's not under 0.1 by any reading I can find. The claim needs to be fixed or removed.\n\nWho should read this: anyone building or evaluating ambiguity benchmarks will want AmbAdv, especially the phonological items. The probe section is a good teaching example of why lexical confounds matter. It deserves a serious referee, but I would send it back for revisions: add noun-held-out and static-embedding controls, defend the labeling, and correct the prompt-consistency claim. The dataset alone is probably publishable; the probe claim as written is not.","headline":"A useful new ambiguity dataset with an overreaching probe claim: the near-perfect probe accuracy likely reflects noun-class memorization, not a usable ambiguity signal, and the prompt-consistency claim is contradicted by the paper's own Table 8.","tokens_in":57449,"tokens_out":3326,"would_cite":true,"duration_ms":40546,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Language models prompted to identify ambiguous sentences perform at near-chance levels, yet linear probes on their internal token representations decode ambiguity with accuracy above 0.9, showing ambiguity information is present but not…","keywords":["ambiguity detection","adversarial dataset","linear probing","residual stream representations","prompting evaluation","yes bias","phonological ambiguity","syntactic ambiguity"],"falsifier":"Judge the same AmbAdv sentences with annotators who are asked to list all possible interpretations without being told to apply world-knowledge constraints, or give minimal licensing contexts; if a large share of random and rhyme variants is judged ambiguous, the reported model error rates and probe accuracies are partly measuring agreement with one plausibility standard. Alternatively, probe accuracy measured on held-out sentences generated from syntactic templates that were not in the training data would settle the memorization claim.","tokens_in":56510,"feed_emoji":"🎭","tokens_out":3628,"duration_ms":34578,"temperature":0.7,"pith_summary":"The paper builds AmbAdv, a dataset of 1,529 sentences that start from syntactic, lexical, and phonological ambiguities and then applies four kinds of minimal edits, namely word-order swaps, synonym replacements, random substitutions, and rhyme-based substitutions, to create ambiguous and unambiguous variants. Prompting four open-access 7-billion-parameter language models to say whether a sentence is ambiguous yields near-chance accuracy, with clear signs of a yes bias, while the same models' internal representations, read by linear probes, separate ambiguous from unambiguous sentences with accuracy above 0.9 in some token positions. The paper's central claim is that ambiguity-related information is present in the models' representations but is not reliably deployed when the models produce answers, and that some successful behavior appears to rest on memorization of publicly available example sentences.","feed_headline":"Prompting fails at ambiguity detection; probing hits 99 percent","feed_subtitle":"A new adversarial dataset shows language models store ambiguity information they never use when answering.","key_machinery":"The load-bearing object is AmbAdv, an adversarial ambiguity dataset built by taking 8 syntactically, 16 lexically, and 16 phonologically ambiguous base sentences and applying four manipulations: word-order swaps, synonym substitution, random same-category substitution, and rhyme substitution. Random and rhyme substitutions are labeled unambiguous because world knowledge rules out the low-attachment reading, and the paper validates this labeling with three annotators (Cohen's kappa 0.91). The representational analysis uses linear probes, meaning logistic regression on layer-wise residual-stream vectors, at specific token roles such as the first verb, first noun, last noun, final punctuation, and the end-of-prompt token, to test whether the label can be decoded from internal states.","core_discovery":"On a subset of 142 sentences built from one PP-attachment template, linear probes trained on residual-stream vectors of tokens such as the last noun or final punctuation decode the ambiguous or unambiguous label with accuracies up to 0.99, whereas the same models prompted with eight different templates reach at most about 0.525 accuracy on the full AmbAdv set. The paper concludes that the models encode enough information to distinguish ambiguity but fail to leverage it in generation. Disambiguation responses also show hallucinated instruments, inanimate objects acting as agents, and repetitions of the original sentence's reading after synonym edits, which the authors take as evidence of partial memorization.","pith_inferences":["If the probe-versus-prompt gap generalizes beyond PP-attachment templates, the headline result is not that models lack ambiguity knowledge but that they lack a reliable output channel for it, which points to alignment and format robustness rather than representation learning as the bottleneck.","A testable extension would be to train the same probe on representations of sentences whose labels are swapped by context, such as licensing contexts that make the dress-as-instrument reading plausible, and check whether probe accuracy tracks the context-dependent gold label or the authors' plausibility judgment.","The memorization hypothesis could be tested directly by measuring probe accuracy on novel syntactic templates, for example 'The chef watched the waiter with the napkin,' generated from the same manipulation procedure; if probe accuracy collapses, the 0.99 figures partly reflect exposure to the source sentences."],"forward_implications":["Direct prompting is not a reliable way to measure what language models know about ambiguity: near-chance scores can reflect an answer-format or yes-bias problem rather than absent knowledge.","Evaluations of ambiguity sensitivity should account for class distribution and prompt phrasing, since one model's good score on a balanced set was produced by answering yes to almost everything.","Because probes can decode the label from late-layer token representations, future work could try to route that information into generation, for example through decoding-time intervention, rather than treating prompting scores as the ceiling.","Rhyme-based perturbations give a new handle on phonological ambiguity in text-based models, extending adversarial evaluation beyond lexical and syntactic edits."],"supporting_citations":[{"why":"Supplies the Ambient benchmark the paper compares against and motivates ambiguity as a challenge for language models.","marker":"(Liu et al., 2023)"},{"why":"Provides the scope-ambiguity dataset showing models may be sensitive to meaning ambiguity, used as a comparison point for AmbAdv.","marker":"(Kamath et al., 2024b)"},{"why":"Contributes the AmbigQA dataset for referential ambiguity, one of the prior resources AmbAdv is positioned against.","marker":"(Min et al., 2020)"},{"why":"Provides ambiguous semantic-parsing inputs and the AMP dataset, serving as a source of base sentences and a baseline comparison.","marker":"(Stengel-Eskin et al., 2024)"},{"why":"Supplies typology and examples of syntactic ambiguity that ground the syntactic base sentences in AmbAdv.","marker":"(Taha, 1983)"},{"why":"Establishes the linear probing methodology the paper uses to evaluate whether representations encode the ambiguity label.","marker":"(Radford et al., 2021)"},{"why":"Supports the interpretation that prompting can systematically underestimate model capabilities, motivating the representational analysis.","marker":"(Hu and Levy, 2023)"},{"why":"Provides the cognitive-linguistic classification of ambiguity into lexical, phonological, morphological, and syntactic types that guides dataset construction.","marker":"(Zabotkina et al., 2021)"}],"fun_headline_variants":["Probes decode ambiguity at 99%; prompts stuck at 52%","LLMs encode ambiguity but refuse to use it","Prompting can't spot ambiguity; probes almost always do","Ambiguity hidden in LLM layers, missing from outputs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that a sentence like 'The man saw the woman with the book' is unambiguously unambiguous because the instrument reading is implausible, since annotators were given the same instruction as the default prompt and model errors are measured against the authors' plausibility judgment.","fun_headline_variants_meta":{"raw":{"variants":["Probes decode ambiguity at 99%; prompts stuck at 52%","LLMs encode ambiguity but refuse to use it","Prompting can't spot ambiguity; probes almost always do","Ambiguity hidden in LLM layers, missing from outputs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000357,"raw_usage":{"total_tokens":1853,"prompt_tokens":780,"completion_tokens":1073,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":396,"completion_tokens_details":{"reasoning_tokens":1004}},"tokens_in":396,"tokens_out":1073,"duration_ms":8523,"temperature":1.0,"reasoning_tokens":1004,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:50:11.376961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Judge the same AmbAdv sentences with annotators who are asked to list all possible interpretations without being told to apply world-knowledge constraints, or give minimal licensing contexts; if a large share of random and rhyme variants is judged ambiguous, the reported model error rates and probe accuracies are partly measuring agreement with one plausibility standard. Alternatively, probe accuracy measured on held-out sentences generated from syntactic templates that were not in the training data would settle the memorization claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides ambiguous semantic-parsing inputs and the AMP dataset, serving as a source of base sentences and a baseline comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies typology and examples of syntactic ambiguity that ground the syntactic base sentences in AmbAdv."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the cognitive-linguistic classification of ambiguity into lexical, phonological, morphological, and syntactic types that guides dataset construction."}],"review_version":1}