Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Linguistics — the study of language itself — morphology, lexicalism, word segmentation, syntax and semantics as linguists pose them, not as models encode them.
25 of 1352 THEORY documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| THEORY-017 v6 | In a Hilbert-space model only unitarily invariant structure is intrinsic; a basis or tensor factorisation, and so any parts or features, is extra data supplied from outside the space, and the record’s Hilbert-space works differ in where they get it | Carroll (2021), LIT-123: a quantum theory is fixed up to unitary equivalence by the Hamiltonian’s spectrum, and subsystems come from a factorisation that the dynamics must select. Van Rijsbergen (LIT-262) computes only traces and inner products, and treats each observable’s eigenbasis as a “point of view”. DisCoCat’s ε/η composition (LIT-273) is orthogonally covariant, while its Frobenius sequel (LIT-272) stipulates a basis and composes by elementwise product in it. THEORY-004 is the same point for learned representations. This is a synthesis across the four, not a result any one of them states. It does not say that basis-dependent models are wrong; it says what they have assumed. | Active |
| THEORY-155 v2 | Iterated learning by Bayesian agents who sample from the posterior converges, in an ergodic chain, to the shared prior | Griffiths & Kalish (2007), LIT-769, read in NOTE-599: a chain of learners who share a prior and each sample a hypothesis from their posterior is a Gibbs sampler, so the probability a learner holds a hypothesis converges to its prior probability, whatever the amount of data passed on; with MAP learners the outcome centres on the prior’s mode but depends on data amount and noise, and can amplify weak biases. It is a theorem about an idealized chain. It does not say human chains reach the prior, and Kirby, Cornish & Smith’s laboratory chains, LIT-771, are consistent with it without testing it. Whether human chains do is QUESTION-137. | Active |
| THEORY-159 | In a language a term’s value is fixed by its relations to the coexisting terms of the same system, so terms that share a signification can differ in value | Saussure (1916), LIT-774, read in NOTE-594, Part Two ch. IV. Value, the horizontal relation among signs, is distinguished from signification, the link inside one sign; the concept side and the sound side of a term are each “purely differential”. It does not say that the sign as a whole is nothing but a difference: the book calls the sign a positive fact and the relation between signs opposition. The support is argument by example. | Proposed |
| THEORY-162 | What an utterance does to the context it is used in is part of its meaning, and its truth conditions do not fix it: sentences with the same truth conditions can differ in what they make available to later discourse, in whether they inform or only test, and in whether they describe a state of affairs or bring one about | Heim (1982), LIT-838; Veltman (1996), LIT-818; Krifka (2024), LIT-835. Three readings find, each in its own domain, sentences that agree in truth conditions or in the proposition they carry but differ in their effect on a conversational state. The effects are anaphoric potential (Heim), tests rather than information (Veltman), and changing indices rather than eliminating them (Krifka). This is the dynamic view: meaning as context-change potential. It does not say that truth conditions are irrelevant, since each framework recovers them, or that the three formalisms agree with each other. | Proposed |
| THEORY-163 v3 | In Zaslavsky, Kemp, Regier and Tishby’s analysis, the colour-naming systems of the World Color Survey languages and English lie near the information-bottleneck bound for compressing perceptual meanings into words, at a trade-off parameter just above 1 and nearer than hue-rotated variants of themselves, with the need distribution over meanings estimated from the naming data | Zaslavsky, Kemp, Regier & Tishby (2018), LIT-810, read in NOTE-631: treating a colour lexicon as an encoder from Gaussian perceptual meanings to words, the World Color Survey languages and English lie near the IB curve of LIT-338 at β ≈ 1.03, beat hue-rotated variants of themselves, and their soft, partly inconsistent naming is what IB optima look like. It does not show that languages are driven to the bound by any process, and the need distribution that makes the fit best is estimated from the naming data themselves. | Active |
| THEORY-164 v3 | As Raviv, Meyer and Lev-Ari report it, in a laboratory game where 12 fully connected groups of eight and 12 of four built a new communication system with total interaction held fixed, the larger groups’ languages grew more systematic faster, ended more systematic and differed less from one another, while communicative success and convergence did not differ by size; input variability, higher in the larger groups, predicted the next round’s gain in structure, an association measured and not manipulated | Raviv, Meyer & Lev-Ari (2019), LIT-825, read in NOTE-628: fully connected groups of eight built more systematic languages than groups of four, faster and more uniformly, with total interaction and network structure fixed; input variability, larger in bigger groups, predicted the gains. One laboratory study, two group sizes, and a mechanism supported by lagged association, not manipulation. It says nothing yet about natural communities, where size travels with network density and second-language learners. | Active |
| THEORY-170 v2 | Languages differ in what their grammar obliges a speaker to convey rather than in what they can convey, so interlingual translation must add some information and leave some unexpressed, and repeated translation can erode a message | Jakobson (1959), LIT-786, after Boas: every cognitive content can be expressed in any language, but each grammar obliges its speakers to answer certain questions (aspect, number, gender, tense, definiteness). Translating between grammars that ask different questions forces the translator to supply answers the source did not give and to leave unexpressed answers it did, and a chain of translations of an isolated sentence can lose its content, less so the richer the context. Argued from examples, not tested; it concerns cognitive content, not pragmatic force. | Proposed |
| THEORY-172 v4 | In Frank and Goodman’s one-shot reference games, listeners’ mean bets matched a parameter-free listener inverting a speaker who chooses among true utterances by informativeness, with a separately measured salience prior, and not salience alone; and in Bergen, Goodman and Levy’s signalling games with novel symbols, listeners read a signal as excluding what the available alternative would have conveyed | Frank & Goodman (2012), LIT-821, read in NOTE-634: a parameter-free Bayesian listener inverting a size-principle speaker, with a separately measured salience prior, matches mean listener bets over seven context types at r = .99 (.87 without the 0 and 100 predictions), while salience alone correlates at r = .19. Bergen, Goodman & Levy (2012), LIT-809, read in NOTE-613: with novel signals and no conventions, listeners read the signal as the object its alternative did not name, and lexical uncertainty derives Horn’s mapping. Goodman & Frank (2016), LIT-816, states the framework. Shown for artificial games with given alternatives, by fits of condition means with no rival fitted; a test against named rivals on trial-level data is QUESTION-209. | Active |
| THEORY-204 v3 | On TinyStories and WikiText-103, the data-limited exponent of transformer test loss, fitted to the lower envelope of learning curves across context lengths at up to about 10^8 training tokens, overlaps γ/(2β), where β is the decay with lag of the top singular value of the token–token covariance and γ the decay of next-token conditional entropy with context length; under assumed per-lag scaling forms, a horizon n*(P) ~ P^(1/(2β)) gives the exponent min(δ, γ/(2β)), δ being the decay rate of the within-horizon excess losses | Cagnetta, Raventós, Ganguli and Wyart (2026), LIT-888, building on LIT-883’s effective context window. Under assumed scaling forms for the conditional entropy and for each lag’s learning curve, the exponent is min(δ, γ/(2β)) (re-derived 2026-10-20). Measured on TinyStories and WikiText-103, with transformers up to about 10^8 tokens and horizons of tens of tokens, the fitted exponent overlaps γ/(2β), best near the start of the envelope; γ is estimated from trained models’ own n-gram losses. Whether the horizon is why is open (QUESTION-233). | Active |
| THEORY-271 v2 | In a toy subject–verb–object corpus whose eight subjects carry every combination of three binary aspects, the subjects’ raw co-occurrence rows form a parallelepiped if and only if, for each object, its four sentences are equally probable; any other sentence distribution breaks it | Torii, Maeda and Hidaka (2024), LIT-977, Theorem 1 (§5.5), read in full in NOTE-781. Necessity is sound and short. The case analysis over permutations, and sufficiency, are loosely written. Exact shapes, raw counts, one design, and independent aspects that make the concept lattice Boolean. It says nothing about the logarithm or PMI, about noisy or approximate shapes, or about attributes that imply one another. | Active |
| THEORY-477 | None of the ten morphosyntactic criteria linguists use to identify words is necessary and sufficient for the orthographic word across languages, and the test batteries that combine them pick criteria opportunistically, so no cross-linguistically valid definition of the word is yet available | Haspelmath (2011), LIT-063, read in full in NOTE-073. Each of ten criteria (pauses, free occurrence, mobility, uninterruptibility, non-selectivity, non-coordinatability, anaphoric islandhood, nonextractability, morphophonological idiosyncrasy, biuniqueness) fails as necessary and sufficient against cross-linguistic counterexamples, and nine published batteries use no common set (Table 1). The target is orthographic practice by design. It does not say that words cannot be defined within one language, that sub-word units are privileged, or that morphology and syntax are shown to be one domain; whether units cluster on a boundness scale is open. | Active |
| THEORY-518 | If a word’s meaning in a distributional model is read differentially, as its relations to its nearest neighbours rather than its coordinates, then adding text moves every similarity value slightly but leaves the neighbourhoods of frequent, consistently used words intact and shifts those of rare words that the new text uses heavily, so the instability that holism implies is local and graded rather than total | Grindrod, Porter and Hansen (2024), LIT-208, read in full in NOTE-158. The conceptual reply to the instability objection is argued; the empirical half is four hand-picked words in two before-and-after demonstrations (a count model on Stein plus a Hemingway story; Word2vec on Wikipedia plus the SEP), top-10 lists, one run each, no metric. The abstract’s proportionality of resistance to frequency is not measured, the criterion used (a preserved neighbour list) is not the one defined (relative distances), and static embeddings only are tested. | Proposed |
| THEORY-542 | A word learner that infers speakers’ referential intentions jointly with the lexicon, and lets words be uttered non-referentially, learns object names from child-directed video more precisely than co-occurrence and translation models, and mutual exclusivity, word-guided object individuation and intention-guided mapping follow from it without principles built in for them | Frank, Goodman & Tenenbaum (2009), LIT-815, read in NOTE-611. On about twenty minutes of video from two dyads, annotated by the authors, the intentional model’s lexicon scores precision .67 and F .55 against .06–.15 and .10–.22 for five baselines. The behavioural fits are qualitative, the baselines also produce mutual exclusivity, and the Baldwin case supplies the intention rather than inferring it. The reading said no THEORY, because the findings are specific to a small corpus and the record held nothing they correct; this account is filed Proposed for that reason. | Proposed |
| THEORY-588 | Lemma-based models of language production build in a lexicalist assumption, and the failure is algorithmic as well as representational: as long as retrieval of stored word-sized units and building of syntactic structure are separate stages, richer stored units cannot handle polysynthetic words, idioms whose parts syntax separates, or suppletion conditioned by an unplanned object | Krauska and Lau (2023), LIT-031, read in full in NOTE-029. A Hypothesis and Theory article with no new data and no implementation. Its diagnosis of named models (Levelt, the Consensus Model, lexically sized increments, treelets) is strong as textual analysis; the case that they fail on Inuktitut, Vietnamese collocations and Hiaki suppletion is argued from a few examples each, and Eberhard et al. are conceded to handle Hiaki in part. The 8-stage replacement model and its aphasia predictions are a verbal sketch, and this account does not rest on them. | Proposed |
| THEORY-623 | In the two tests of contextuality in language data that the record has read, the contextuality is put there by the experimenter or not secured: meaning-selection models of ambiguous phrases estimated from corpora are all signalling and their two Contextuality-by-Default verdicts are plausibly noncontextual by their own bootstrap, while BERT’s 52 million anaphora models have a support fixed by the sentence schema, so whether each counts as contextual depends only on how uncertain the model’s three referent choices are | Lo, Sadrzadeh & Mansfield (2024), LIT-844, read in NOTE-648, and Wang, Sadrzadeh, Abramsky & Cervantes (2021), LIT-842, read in NOTE-645. Both readings recomputed the criteria. That 71.1% CbD-contextual measures BERT’s uncertainty is NOTE-648’s derivation from the stipulated support (CNT1 = 2 − Δ). The only possibilistically contextual phrase model has hand-set supports. Neither study uses human judgements, so neither bears on whether people’s interpretations are contextual. | Active |
| THEORY-643 | The restrictions on syntax reaching into words that a separate word-formation component was credited with are restrictions on heads versus phrases and on what individual affixes select, and phrases feed word formation, so a single combinatorial component can build both words and phrases | Bruening (2018), LIT-071, read in full in NOTE-065. Phrases feed English word formation productively, ellipsis and focus reach sub-word units, and claimed differences of principle (locality, adjuncts, headedness, idiosyncrasy) fail; extraction, coordination and the no-phrase constraint are restated over X0 vs XP and affix selection, with no mention of “word”. The English evidence of §§2–3 is the strong part. Several sub-arguments are provisional by the author’s own account, and Müller’s reply in the same volume disputes extraction, coordination and the resultative argument. | Proposed |
| THEORY-670 | Ordinary dispositions can be modelled as restricted possibilities individuated by their manifestation alone: ‘fragile’ as manifestation in a sufficient proportion of cases over maximally varied external circumstances, ‘breakable’ as plain possibility; grading by closeness to an ideal world fails for any choice of ideal | Vetter (2014), LIT-109, read in full from the penultimate draft in NOTE-125. A proposal, which the paper says it does not show superior to the conditional conception (pp. 6, 26). Its strongest part is negative: for any ideal world, an Achilles’-heel block that breaks only there counts as fragile, so closeness grading fails. The positive proportion measure is not defined, and the finite, maximally varied set of cases it counts over is admitted to be non-unique. The evidence is ‘-able’ semantics and a COCA count in which none of 226 relevant uses of “disposed to” ascribes a standing tendency to an inanimate object. Finks and masks are left unsolved, and dispositions of fundamental science are set aside. | Proposed |
| THEORY-693 | In conversation, evaluating an object, positioning oneself and aligning with another speaker are three aspects of one stance act rather than three kinds of stance, because a stance takes a shared object and answers a prior stance, which are usually supplied by earlier turns; a resonant follow must mark its relation to the lead, with too or either in English | Du Bois (2007), LIT-779, read in full in NOTE-590. Argued from short excerpts of American English conversation (Santa Barbara and Longman corpora), with no counts. The co-occurrence of the three aspects is shown in chosen examples, not tested for separability; the too/either result rests on a separate study (Du Bois 2004) whose figures are not given. It does not treat a speaker who animates someone else’s stance, which it sets aside for Goffman. | Proposed |
| THEORY-784 | Language models trained on limited data show a test loss falling as P^(−γ/(2β)) because P training tokens resolve token correlations only out to a horizon growing as P^(1/(2β)), so that more data mainly lengthens the context a model can use, when the model learns within that horizon faster than it grows, as Cagnetta, Raventós, Ganguli and Wyart propose | Cagnetta, Raventós, Ganguli and Wyart (2026), LIT-888, read in NOTE-689, building on Cagnetta and Wyart’s effective context window (LIT-883). The paper presents the horizon as the reason the data-limited exponent is γ/(2β). THEORY-204 holds what it showed: the derivation under assumed scaling forms, and the measured overlap on TinyStories and WikiText-103 up to about 10^8 tokens. Whether the horizon is why is the test QUESTION-233 asks. | Proposed |
| THEORY-785 | Listeners interpret an utterance by inverting a model of a speaker who chooses among alternatives by informativeness, so what a fixed form conveys depends on what else the speaker could have said, as the rational speech act account of Frank, Goodman and coauthors proposes | Frank and Goodman (2012), LIT-821, and Goodman and Frank’s review (2016), LIT-816, present listener inference as the inversion of an informative speaker, the rational speech act account. The record holds as shown, in THEORY-172, the fits of that model to mean bets in one-shot reference games and Bergen, Goodman and Levy’s signalling-game result. This is the mechanism the fits were offered for, filed under ADR-047 v4 when the narrowing dropped it. | Proposed |
| THEORY-791 | The colour-naming systems of the world’s languages lie near the information-bottleneck bound for compressing perceptual meanings into words, and differences in one trade-off parameter place each language along it, as Zaslavsky, Kemp, Regier and Tishby propose | Zaslavsky, Kemp, Regier & Tishby (2018), LIT-810, read in NOTE-631, using the bound of LIT-338. Their fit covers the World Color Survey languages and English, with the need distribution over meanings estimated from the same naming data, and that measured part is the Active THEORY-163. The authors present the result as a property of the world’s languages. This THEORY holds that general claim. | Proposed |
| THEORY-1004 | In Pythia-1B, OLMo-1B and BLOOM-1B, the crosscoder features that most affect subject–verb agreement change over pretraining: at early checkpoints they detect particular tokens and subtokens, at later ones grammatical categories; in BLOOM, language-specific features give way to features shared across English, French, Spanish and Portuguese while Hindi and Arabic keep language-specific ones; and which features matter keeps shifting after task accuracy has stopped improving | Bayazit, Mueller and Bosselut (2025, ACL 2026), LIT-1555, read in NOTE-1186. Crosscoders trained jointly on the middle layer of two or three checkpoints, with features scored by their relative indirect effect on agreement minimal pairs at each. The reconstruction quality (C1) and RelIE’s agreement with single-feature ablations (C2) are measured; the trends in this title are read by the authors from annotated features. It describes; it does not predict later checkpoints, and it cannot tell a feature carried forward from one re-learned in the same direction. | Proposed |
| THEORY-1131 | On native-authored pragmatic items in Hindi, Punjabi, Tamil and Malayalam, current multilingual LLMs pick the right reading from four options more often than they label it unaided, and when wrong they lean toward the literal reading | Singh et al. (2026), LIT-1817, read in NOTE-1345 from the Findings of EMNLP 2026 PDF. MCQ beat NLI in 20 of 20 model–language pairs (recomputed). 59.1% of wrong MCQ answers took the literal distractor, against 33.3% by chance, and 53.8% of Contradiction items were labelled Entailment. The gold answers are each language’s one item author’s reading, cells hold 5–16 items, and the MCQ and NLI sets are different items. So “recognizing and reasoning about pragmatic meaning are distinct competencies” is the authors’ gloss, not a finding. | Proposed |
| THEORY-1144 | A multilingual classifier trained only on English politeness ratings separates the most from the least polite requests in nine other languages at about 81% accuracy, without translation, yet stays about 10 points below one native judge’s agreement with two others in most of them | Srinivasan and Choi (2022), LIT-1799, read in NOTE-1339. XLM-R fine-tuned on 1,926 English requests reaches 0.808 mean accuracy on the balanced top/bottom-quartile sets (250 per language), 0.801 on Google Translate input; GPT-3 zero-shot 0.706 (Table 3). Leave-one-out human accuracy is around 90% (Fig. 3). The judges were screened for agreeing with this classifier, and the extreme items were chosen on a mean that includes the held-out judge, so the gap and the transfer are both measured on favourable ground. | Proposed |
| THEORY-1171 | In groups that build a new communication system, larger groups make it more systematic, and the greater variety of input they face drives the effect | The wide form of Raviv, Meyer and Lev-Ari (2019), LIT-825, kept here when the fifteenth status review narrowed THEORY-164 to what their experiment showed and promoted it (ADR-047, kind 1). The paper asserts that larger groups make an emerging language more systematic and that the greater variety of input they face drives the effect. One laboratory study with two group sizes found the size effect, and the mechanism rests on lagged association within its data; whether the effect replicates, and whether variability causes it, is open. | Proposed |