{"id":"2c464a53-afc5-468c-937c-175410cedad3","arxiv_id":"2411.16433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Language models exhibit human-like structural priming and can generalize some linguistic rules from indirect evidence, but their adjective-order preferences largely track training-data bigram frequencies.","lead":"This PhD thesis asks whether large language models learn grammatical structure like humans do, using behavioral tests, training-data analysis, and synthetic languages. It finds that models show structure-sensitive priming and can generalize from indirect evidence, but much of their behavior tracks simple corpus statistics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FiCT's indirect-evidence conclusion depends on filters removing all direct evidence and leaving the distribution intact; without a leakage audit or matched-corpus control, contamination or distribution shift remain live alternatives.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: FiCT's validity requires filters to remove all direct evidence while leaving the rest of the distribution intact. My stress-test converges on the same concern, sharpened into two testable failure modes: leakage and distribution shift. The rest of the thesis is comparatively well supported: the structural priming chapters use carefully controlled corpora and report both positive and negative results, and the adjective-order chapter explicitly checks n-gram correlations and unseen-pair generalization, so it already acknowledges the memorization/generalization trade-off. The FiCT claim is the one place where the central conclusion about indirect evidence would collapse under either failure mode. A concrete leakage audit and a matched-deletion control would settle whether the filters are clean; without those checks, the conditional verdict is appropriate, which is why I recommend no change to the reader's verdict.","tokens_in":49541,"tokens_out":3357,"duration_ms":37777,"concrete_test":"Perform a leakage audit on each FiCT filtered corpus using a second, independent detector, such as a different parser or a manually validated grammar, to estimate the fraction of retained sentences that contain the target construction. If the fraction is above zero, retrain on a stricter filter and recompute accuracy delta. Additionally, train control models on corpora where the same number of length- and frequency-matched sentences are removed by an unrelated filter; if the control accuracy delta is comparable to the FiCT delta, the indirect-evidence interpretation is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Chapter 5 (§5.5, §5.7) is that models trained with FiCT generalize from indirect evidence, because they judge correctly on constructions whose direct instances were removed. This conclusion is valid only if the filters in §5.3 and Figure 5.1 actually remove all direct evidence and do not systematically distort the training distribution. Two unaddressed threats remain. First, leakage: the filters are implemented with parsing and surface heuristics, so parse errors, coordination, or alternative syntactic realizations can leave target constructions in the 'filtered' corpus, making the evaluation items not truly unseen. The visible text reports no independent audit of the filtered corpora. Second, distribution shift: removing all sentences containing a construction also removes correlated material; for example, filtering out PP-modified subjects changes the frequency of specific prepositions, noun-preposition bigrams, and subject length. If a control corpus with the same amount of unrelated, length- and frequency-matched deletion is not used, the measured 'indirect evidence' improvement could simply reflect a shift in baseline statistics. The accuracy-delta metric in §5.5 compares filtered against unfiltered models, but without such a matched deletion control it cannot separate genuine generalization from these confounds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The thesis investigates whether Transformer language models encode abstract grammatical structure, using three complementary strands: structural priming as a psycholinguistic paradigm (Chapters 2–3), analyses connecting training-data statistics to model behaviour including the Filtered Corpus Training (FiCT) intervention (Chapters 4–6), and controlled synthetic-language experiments for evaluating interpretability methods (Chapters 7–8). The central positive claims are that language models exhibit structural priming modulated by lexical, semantic, and frequency factors in ways partly analogous to humans, and that FiCT-trained models can judge constructions that were filtered out of their training data, indicating generalization from indirect evidence. The adjective-order chapter is more cautious, showing that simple bigram frequencies from The Pile predict most adjective-order preferences, with limited generalization to unseen pairs and with context effects that go beyond n-gram counts.","tokens_in":49739,"tokens_out":6274,"duration_ms":59327,"significance":"If the results hold, the thesis makes a substantial contribution to the debate on memorisation versus generalisation in language models and demonstrates the value of repurposing psycholinguistic paradigms in NLP. Concrete strengths include: the carefully constructed and publicly released Prime-LM and CAP corpora; the Priming Effect metric of Chapter 2, which compares two primes against the same target and thereby avoids the prior-perplexity confound of earlier adaptation-based measures; the use of intermediate checkpoints and batch-level n-gram counts in Chapter 4 to relate model behaviour directly to its training signal; the genuinely causal FiCT training-intervention design; and the FIDAM framework in Chapter 7, which explicitly addresses faithfulness of feature-interaction attributions. The thesis is also appropriately self-critical, for example in reporting that bigram statistics explain 90.3% of adjective-order accuracy and in showing that priming asymmetries are opposite in direction to typical human asymmetries.","major_comments":[{"comment":"The central claim of Chapter 5—that FiCT-trained models 'have never seen instances of their evaluation targets' and therefore generalize from indirect evidence—requires that the filters actually remove all direct evidence of each target construction. The visible text describes filters based on parsing and surface heuristics (Figure 5.1; §5.3) but reports no independent audit of the filtered corpora: parse errors, coordination, or alternative syntactic realizations could leave fragments or full instances of the target construction in the 'filtered' training data, making the evaluation items partially seen. The authors should quantify filter precision and recall on a held-out sample and, ideally, verify with an n-gram or dependency search that the TSE evaluation sentences or their structural instantiations do not appear in the filtered corpus. Without such an audit, the strong 'never seen' formulation is unsupported.","section":"§5.3, Fig. 5.1, §5.5"},{"comment":"Even if the filters are perfectly precise, the accuracy-delta comparison between filtered and unfiltered models cannot separate indirect-evidence generalization from distribution shift induced by filtering. Removing all sentences containing PP-modified subjects, for example, also changes the frequencies of specific prepositions, noun-preposition bigrams, and subject lengths, which can shift baseline statistics for the TSE items. The chapter does not include a matched-deletion control in which an equal amount of unrelated, length- and frequency-matched material is removed. Adding such a control—or showing via re-weighting that the filtered corpus preserves the relevant marginal distributions—is necessary to support the conclusion that the measured accuracy deltas reflect genuine generalization to unseen constructions rather than a shift in the training distribution.","section":"§5.5, §5.7"}],"minor_comments":[{"comment":"Equation (2.2) writes the conditioning context as 'tx_i' in the final term; it should be the previous target tokens tx_{<i}, as correctly written in Eq. (3.2).","section":"§2.3, Eq. (2.2)"},{"comment":"Several cross-references in Chapter 2 are stale: the captions of Figures 2.3–2.5 refer to '§7.1.1', '§7.2', '§7.3.1', '§7.3.2' and '§7.4', which should be Sections 2.6.1–2.6.5, and Table 2.1 is referenced as 'T able2.1'.","section":"Chapter 2 captions"},{"comment":"There is a duplicated word in the sentence 'unless unless explicitly mentioned otherwise'.","section":"§3.2.1"},{"comment":"Figure 5.1 labels the method as 'FICT' in the overview box while the text consistently uses 'FiCT'; the capitalization should be standardised.","section":"Fig. 5.1"},{"comment":"The token labels in the legend of Figure 3.2 (e.g., 'NPO2 × NDO2') are difficult to parse; the caption would benefit from listing the corresponding word positions in the example sentence.","section":"Fig. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The thesis is a collection of previously published papers, and the chapters are largely reprints of the author's published work. The most novel and defensible contribution is Chapter 5's FiCT design, but it is also the claim most likely to be scrutinized: the 'indirect evidence' conclusion depends on filter completeness and on separating distribution shift from generalization. I recommend requesting the filter audit and matched-deletion control described in the major comments. The remaining chapters appear sound and, if Chapter 5 is strengthened, the thesis would be a solid contribution to the computational linguistics / interpretability literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-executed thesis compilation of seven published papers. The two chapters I could read in full (structural priming, adjective order) are careful and honest; the FiCT chapter's central claim, however, is not yet supported by the evidence shown in the text.\n\nWhat's new: the thesis itself is a compilation, so the novelty sits in the individual contributions. Chapter 2's Priming Effect metric and the Prime-LM corpus are reusable, controlled tools, and Chapter 3's token-level decomposition is a genuine conceptual addition. Chapter 4's CAP corpus and the finding that simple bigram frequencies predict 90.3% of human-preferred adjective orders while a 12B model sits at 94.1% is a clean, honest result — the authors explicitly report that n-gram statistics carry most of the signal and that generalization to unseen pairs is limited. The three-stage learning dynamic is a nice observation.\n\nThe soft spot is Chapter 5. The FiCT conclusion that models 'generalize from indirect evidence' depends on filters actually removing all direct instances of the target construction. The visible text does not report a leakage audit (e.g., how many evaluation items survive in the filtered corpus after parsing/heuristic filters) nor a matched-deletion control for distribution shift — filtering out PP-modified subjects also deletes correlated material like preposition frequencies or subject lengths. Without those controls, contamination or a baseline shift remain live alternative explanations for the accuracy deltas. This is fixable, but it is a load-bearing assumption for the thesis's strongest claim. A second, smaller caveat: the provided file truncated after Chapter 5, so I could only assess later chapters (monotonicity, synthetic languages, feature interactions) from their overviews; my confidence in the full thesis is moderate, not high.\n\nWho this is for: anyone working on LM evaluation, psycholinguistic methods for LMs, or the memorization-vs-generalization debate. It deserves a serious referee — the methods are reusable and the reporting is straight — but a referee should require the FiCT leakage/control analysis before accepting the indirect-evidence conclusion at face value.","headline":"Careful psycholinguistic-style thesis with a strong structural priming chapter and an honest adjective-order study; the FiCT 'indirect evidence' claim is plausible but not yet proven without a leakage audit or matched-deletion control.","tokens_in":50275,"tokens_out":2929,"would_cite":true,"duration_ms":29763,"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":"Language models encode abstract grammar and can generalize it from indirect evidence, while adjective-order preferences largely track corpus frequencies.","keywords":["language models","structural priming","linguistic generalisation","filtered corpus training","adjective order","negative polarity items","next-word prediction","interpretability"],"falsifier":"Run a FiCT experiment in which every sentence containing a prepositional-phrase-modified subject is removed from the training corpus, verified with an independent parser and n-gram search of the training files, then test the trained model on Targeted Syntactic Evaluations for that construction; if the model scores at chance, or if the filtered corpus is found to contain any direct example of the construction, the claim that models generalise from indirect evidence would be refuted.","tokens_in":1486,"feed_emoji":"🧠","tokens_out":1971,"duration_ms":88514,"temperature":0.7,"pith_summary":"The thesis asks whether Transformer language models—neural networks trained only to predict the next word in a sentence—acquire abstract grammatical structure rather than memorised text. It argues that they do: models show structural priming, the psycholinguistic signature in which recent exposure to a structure makes the same structure more probable next, and the effect is moderated by the same inverse-frequency and lexical-overlap factors found in humans. Training models on corpora with specific constructions filtered out still leaves them able to judge those constructions grammatically, which the author reads as generalisation from indirect evidence. At the same time, adjective-order preferences, while highly accurate on natural orders, are strongly correlated with bigram frequencies in the training data, with only limited generalisation to unseen combinations. Together the results support a two-sided picture: next-word prediction yields genuine abstract structural knowledge, but part of that knowledge is collocational rather than rule-based.","feed_headline":"Filtered training shows language models learn grammar they never see","feed_subtitle":"The thesis shows abstract grammar survives even when a construction never appears in training data.","key_machinery":"The Priming Effect (PE) is the load-bearing metric: the difference in log probability of a target sentence conditioned on a congruent versus an incongruent prime, averaged over controlled prime-target corpora, which isolates structural persistence from lexical and semantic overlap. Filtered Corpus Training (FiCT) is the load-bearing method: train on a corpus with a linguistic construction removed, then evaluate with Targeted Syntactic Evaluations, with the accuracy delta between filtered and unfiltered training measuring generalisation from indirect evidence. The adjective-order analysis adds AOP-Δ, the log-probability difference between natural and swapped adjective order, correlated against relative n-gram counts extracted from the training corpus.","core_discovery":"On its own terms, the central claim is that language models do not merely memorise training text: they build representations of abstract grammatical structure that persist across sentences and can be acquired from indirect evidence. The structural-priming experiments show that congruent primes raise the probability of a target structure even when prime and target share no lexical content, and that the effect increases with repeated exposure and decreases with distance, mirroring human production patterns. The Filtered Corpus Training experiments show that models trained on corpora from which a construction was removed can still make correct grammaticality judgements about that construction, and that Transformers' better perplexity does not translate into better linguistic generalisation than LSTMs. The adjective-order experiments complicate the picture: high accuracy on naturally occurring orders is largely predicted by simple n-gram counts, though context and unseen pairs reveal some abstraction beyond counts.","pith_inferences":["Beyond the paper, FiCT could be pointed at adjective order: remove every double-adjective phrase from training and test whether preferences persist; if they do, the signal must come from adjective-noun statistics or semantic classes rather than memorised collocations.","The cumulative priming effect reported here, where more congruent primes monotonically increase the effect, resembles a minimal form of in-context learning; a direct test would compare priming curves to few-shot demonstrations on a held-out structure with matched lexical content.","The strong bigram correlation in adjective order implies that interpretability claims about LM linguistic rules should routinely be checked against corpus-frequency baselines; otherwise human-like accuracy may be misread as human-like representation."],"forward_implications":["If language models genuinely prime, structural information is active at inference time and can be studied through sentence-level probability contrasts rather than probes or fine-tuning.","If FiCT results are correct, a language model can acquire a construction it never saw, so the language-modelling objective itself can support poverty-of-stimulus-style generalisation.","Perplexity and linguistic generalisation dissociate: a Transformer's perplexity advantage does not imply better structural learning, so architecture comparisons should separate the two.","Adjective-order results imply that claims about LM linguistic rules must be checked against training-data frequency baselines, since bigram counts alone reach around 90 percent accuracy on natural orders.","Inverse-frequency and lexical-overlap effects in priming suggest LM predictions incorporate abstract expectations that mirror human production patterns."],"supporting_citations":[{"why":"Supplies the structural priming paradigm and the claim that structural persistence evidences abstract structural representation in human production.","marker":"Bock, 1986"},{"why":"Provides the review of structural priming in humans that motivates the Priming Effect metric and the factor manipulations.","marker":"Pickering and Ferreira, 2008"},{"why":"Gives the surprisal and inverse-frequency account of human priming that Chapter 3 tests in language models.","marker":"Jaeger and Snider, 2013"},{"why":"Meta-analysis establishing lexical overlap as the most consistent moderator of human priming, used as the human comparison in the regression.","marker":"Mahowald et al., 2016"},{"why":"Defines Targeted Syntactic Evaluations, the evaluation paradigm used throughout the FiCT experiments.","marker":"Marvin and Linzen, 2018"},{"why":"Provides the BLiMP phenomenon list from which FiCT filters and evaluations are drawn.","marker":"Warstadt et al., 2020"},{"why":"Earlier priming-in-LMs work that modelled priming as fine-tuning adaptation; the Priming Effect metric is positioned against it.","marker":"Prasad et al., 2019"},{"why":"Ablates rare constructions from training data to study acquisition; FiCT extends this intervention to a broad phenomenon set.","marker":"Misra and Mahowald, 2024"},{"why":"WIMBD API providing exact n-gram counts of the training corpus used to test whether adjective order is driven by data statistics.","marker":"Elazar et al., 2024"}],"fun_headline_variants":["Grammar learned even when never seen in training","Language models build grammar they never encounter","Unseen grammar: models infer rules from structure","Filtered training: models still learn missing grammar","Abstract grammar emerges without direct examples"],"cache_read_input_tokens":52480,"weakest_assumption_plain":"The conclusion rests on the assumption that the filtering step removes every direct example of the target construction from training while leaving the rest of the distribution intact; if some examples leak through or the filtering changes correlated cues, the apparent generalisation could come from direct evidence.","fun_headline_variants_meta":{"raw":{"variants":["Grammar learned even when never seen in training","Language models build grammar they never encounter","Unseen grammar: models infer rules from structure","Filtered training: models still learn missing grammar","Abstract grammar emerges without direct examples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000477,"raw_usage":{"total_tokens":2369,"prompt_tokens":954,"completion_tokens":1415,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1350}},"tokens_in":570,"tokens_out":1415,"duration_ms":11464,"temperature":1.0,"reasoning_tokens":1350,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:06:42.983083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a FiCT experiment in which every sentence containing a prepositional-phrase-modified subject is removed from the training corpus, verified with an independent parser and n-gram search of the training files, then test the trained model on Targeted Syntactic Evaluations for that construction; if the model scores at chance, or if the filtered corpus is found to contain any direct example of the construction, the claim that models generalise from indirect evidence would be refuted.","supporting_citations":[],"review_version":1}