Source-linked AI summary
Language Has Two Parameters: Narrative-Induced Semantic Plasticity and Phase-Sensitive Interpretation
Hollis Robbins
TL;DR
The paper addresses whether corpus statistics and standard transformer representations capture meaning relations shaped by individual and shared encounter histories. It proposes phase as a signed, persistent, history-indexed parameter and concludes that interpretation requires this relational dimension, while its stronger quantum claim remains unestablished.
Problem
Corpus statistics estimate association strength but omit history-indexed relations that distinguish individual and dyadic meanings.
Method
The paper formalizes phase as a relational semantic parameter and proposes encounter-history-indexed states and tests of persistence, order, marking, and coordination.
Results
The paper concludes that meanings can coexist at full strength while their contributions reverse, requiring a second relational parameter beyond association strength.
Takeaways & Limitations
Language models would need states indexed to encounter histories and relations between meanings rather than only population-level token regularities.
Takeaways & Limitations
The theory’s decisive encounter-order constraint has not yet been derived, leaving its quantum account potentially equivalent to a sufficiently expressive signed classical model.
Abstract
from arXiv · showhide
Language has two parameters. Count how often words occur together and you estimate amplitude, the strength of association. Word embeddings and attention weights refine that count, which sums every writer in the corpus together. This paper claims a second parameter, phase, which signed weights learned from a corpus do not supply. Phase exists only between meanings: it determines how coactivated meanings combine, and it can reverse what a meaning contributes while that meaning stays fully present. A speaker can set phase in the signal through linguistic form; encounters install phase relations and history distributes them. Population averaging deletes history-indexed phase: agent-deindexed corpora identify the population marginal state and determine no individual or dyadic state, at any scale. The standard transformer has no explicit representation for phase in frozen inference, and the interpretability program measuring progress by monosemanticity is optimizing against it: the coexistence it treats as a defect is the condition of allusion, irony, and quotation. Six predictions test whether a suppressed meaning stays active, whether encounter order changes what a phrase does, whether marking the signal changes how a shared phrase is taken, and whether a model given a history is changed by it or only informed about it. The claim defended is the weak version: interpretation requires a second relational parameter, signed, persistent, and indexed to individuals and dyads. Quantum probability is one notation for the parameter; nothing in the formalism claims quantum processes in the brain. The strong version, that the quantum calculus constrains these phenomena as signed classical models do not, rests on an encounter-order constraint not yet derived. The architecture the theory calls for is a language model with agent-indexed, phase-bearing semantic states.
1. The problem
The section argues that meaning arises from interacting constellations of associations whose effects shift with linguistic form, narrative encounter, and shared history. It presents the same phrase as potentially hostile, affectionate, or technologically allusive, exposing a limitation of context-free corpus pretraining.
- Associative structure: Words, poetic forms, and nearby words carry interacting constellations of associations that produce shifts in meaning.The passage describes meanings as simultaneously present and exerting changing effects depending on configuration.
- Narrative effects: Encountering fiction reorganizes readers’ representational capacities and changes the statistical properties of literary language.Characters, genres, and plots can also accumulate associations that enter later fiction and poetry.
- Shared history: The phrase “FUCK YOU, ASSHOLE” can remain recognizable as hostile while expressing affection between viewers familiar with The Terminator.Those without the shared film history may still interpret the phrase as hostile.
- Semantic change: The phrase can acquire further meanings over time, including “you are using an old model” between people familiar with the film and AI.The passage uses this development to argue that interpretive architecture is not fixed.
- Limits of current models: Context-free corpus pretraining cannot model meanings that differ across speakers, particular relationships, or an individual’s history of transformations.The missing information includes who encountered texts, when, in what order, and with whom.
2. Propositions
The propositions define language as a system with amplitude and persistent, relational phase, whose semantic states are shaped by form, fiction, history, and encounter order. They explain how suppression preserves meaning, coordination enables shared cultural effects, and changing historical distributions rephase interpretation and theory.
- Language can set phase through cadence, meter, rhyme, lineation, quotation, and genre; fiction installs phase, while history distributes it.The proposition describes phase as a relation between live meanings transmitted by linguistic form.
- Fiction changes readers’ semantic systems by forming new constellations, altering meaning weights, and resetting relations among meanings.These changes establish the semantic states later language activates.
- Amplitude measures how strongly a meaning contributes, whereas phase determines whether coactivated meanings reinforce or cancel; both persist in individual states between encounters.Neither parameter belongs to the word itself; both belong to an individual’s state.
- Interpretation depends on individual and modeled social encounter histories, and encounter order changes semantic states because updates do not commute.A later encounter can recompute earlier stored relations, producing retrospective reinterpretation.
- Coordination supports allusion, quotation, irony, and cultural recognition, while suppression preserves a meaning whose phase relation can reverse its contribution.Marked ironic quotation is an extreme case: the suppressed meaning remains fully active but contributes in the opposite direction.
- Audiences are changing distributions of encounter histories, so later technologies and narratives can rephase earlier works and shape theories as well as utterance reception.The corollary extends the history-dependent propositions to the production of theories.
3. The formal theory
The formal theory models interpretation with amplitude and a second, relational phase parameter, using quantum probability to represent suppressed meanings and encounter-order effects. It distinguishes individual and dyadic, history-indexed semantic states from population averages that erase phase relations.
- 3. The formal theory: Quantum probability supplies phase relations that affect interpretation when meanings are coactive and operations whose order can change an individual’s semantic state.The theory uses this formalism to capture two requirements ordinary probability cannot represent.
- 3.1 States: Each live meaning has an amplitude measuring presence and a phase determining its relation to other meanings and how it combines with them.A semantic state is represented as a weighted combination of distinguishable meaning states for a phrase and an individual’s encounter history.
- 3.1 States: The theory claims that phase captures effects amplitude-only accounts obscure: a meaning can remain fully present while its contribution is reversed.It does not claim that the brain is a quantum computer or that linguistic meaning is physically quantum.
- 3.2 Interpretation as measurement: Interpretation is an interaction between the configured phrase state and the person’s history-shaped state, with quotation marking, cadence, meter, rhyme, and lineation constraining transmitted phase relations.The same constellation can produce different answers depending on what is measured, even though the meanings themselves remain unchanged.
- 3.3 Encounters and their order: Encounter order matters because the same events can produce different semantic constellations, requiring an account of noncommuting transformations across a life.The claim is narrower than rejecting Bayesian or predictive-processing models, which can represent some path dependence and order effects.
- 3.4 What pretraining cannot recover: For an agent-deindexed corpus, pretraining estimates at most the population-average state ¯ρ_X, which cannot assign a history-conditioned state to any particular agent or dyad.A larger corpus sharpens the estimate but cannot recover a particular constellation; many history ensembles can produce the same average.
- 3.4 What pretraining cannot recover: A context prompt can simulate an expected dyadic interpretation but does not establish a durable, persistent representation of how a particular encounter changed either individual.Global phase can average to κ≈0 while a community retains |κ_C| > 0 through shared encounters and ongoing coordination.
3.5 Decoherence: live allusion, dead idiom, recovery
Relations between meanings can decay while the meanings remain fully active, but rereading, teaching, quotation, and renewed circulation can restore them. Interpretation therefore distinguishes live, phase-related allusion from dead idiom and treats reception history and partner-specific constellations as consequential.
- Decoherence: Renewal through rereading, teaching, quotation, or reissue strengthens meaning-relations; without it, relations decay over time τ while both meanings remain fully active.The simplest decay equation describes an unrenewed relation, and τ measures how long decay takes.
- Decoherence: A live allusion preserves phase-related source and use meanings with audible interference, whereas a dead idiom retains available source and literal meanings without their interference.“Hoist with his own petard” illustrates a dead idiom whose use-meaning remains while Hamlet’s source and literal meanings no longer interfere.
- Decoherence: Later encounters can create relations to technologies and concepts that postdate a work, making reception history a record of repeated audience preparation.The passage presents reception history as documenting how audiences were prepared and prepared again.
- Individual and dyadic interpretation: Partner-specific interaction, community-level conventions, and population priors form distinct levels, while standard RSA models utterances, meanings, and utilities with nonnegative probabilities.The hierarchical account separates individual, dyadic, and population levels within one model.
- Individual and dyadic interpretation: Sharing an encounter can support quotation recognition without producing the same meaning-constellation, so interpretation requires partial coordination rather than shared meaning.Different encounter timing, such as before or after a later technology, can make the same quotation funny for different reasons.
4. The boundary case outside language: Move 37
Move 37 is the theory’s boundary case: a wordless semantic system in which phase exists only in receivers. Its trajectory shows that individual self-play and shared viewing create meanings and relations that population corpora record incompletely, while the same move can shift from shock to quotation to convention as its relation decays or is renewed.
- Boundary case: Move 37 demonstrates the theory’s boundary case: phase can exist between receivers in a semantic system with no words.The passage identifies Move 37 as demonstrating Propositions 1, 3, and 5.
- Individual trajectory: A move rated one in ten thousand by human games emerged through AlphaGo’s self-play trajectory and was absent from the human game corpus.AlphaGo’s supervised human-game training was followed by a distinct self-play history that produced values unavailable in the first corpus.
- Shared encounter: A shared broadcast partially coordinated viewers, prompting professional experimentation and making later fifth-line shoulder hits double-voiced as both strategy and quotation.Only viewers of the match perceive the allusive second move, and the shared encounter reset the community’s interpretive weights.
- Corpus limitation: Adding the game record to the training corpus embeds the move but omits the transformations in viewers and the mutual recognition between players who watched.The passage distinguishes embedding the move from embedding the history that gives it meaning between individuals.
- Coherence and marking: The move’s allusive relation can decay or be renewed, shifting its interpretation from shock to quotation to convention among observers with different histories.Renewal can occur through commentary, anniversary coverage, and teaching; there is no marking on the stone that identifies quotation.
5. Predictions
The theory proposes six predictions distinguishing signed, history-dependent interpretation from amplitude-only and history-as-information accounts. They test persistent activation, encounter-order constraints, divergent arrangements of shared encounters, historical rephasing, persistent model transformation, and signal-dependent phase.
- Prediction overview: Six predictions distinguish the theory from amplitude-only competitors and accounts treating personal history as information supplied only at inference.The predictions target both human interpretation and models exposed to narrative encounters.
- Prediction 1: Interference: During affectionate quotation of an insult, the insult meaning should remain active, with its signed contribution converting hostility into intimacy.Negation research motivates persistent activation; the theory adds that the persisting component contributes with a sign.
- Prediction 2: Constrained encounter-order effects: Interpretation judgments should differ by encounter order, with a proposed constraint analogous to QQ equality but not yet derived.The theory is incomplete without this constraint because encounters transform the state itself, unlike question-order effects on an already constituted reader.
- Prediction 3: Shared encounter, divergent arrangement: Shared encounters should predict recognition between individuals without predicting which person finds a quotation funny or why.What the individuals share is a constellation of events, facts, and commitments rather than an identical arrangement of meaning.
- Prediction 4: Historical rephasing: Historical rephasing predicts that viewers first encountering The Terminator after predictive text and generative AI will more readily connect its response-menu scene to model sampling and machine language interfaces.The comparison concerns descriptions of the scene by viewers whose first encounters occurred before versus after those technologies became widespread.
- Prediction 5: Persistent transformation versus prompted simulation: Identical models given different narrative encounter sequences should retain transformed states after prompts are cleared, unlike models merely conditioned on encounter descriptions.The test targets allusion, irony, and recognition with particular interlocutors.
- Prediction 6: Device dependence: Changing only whether a shared line is delivered with machine cadence or quotation marks should weaken or restore its reversal, while plain unmarked words should not.History-only accounts predict no difference; cue theories treat marking as selection, whereas this theory keeps the insult active at full strength under marking.
6. Future work: implications for language-model architecture
Future language-model architectures must represent not only the strength of individual meanings but also persistent, agent- and relationship-indexed relations among coactive meanings. Such systems must update those relations through shared encounters and evaluate them across time and relationships, rather than treating context as undifferentiated token history.
- Persistent encounter history: Stored relations must be indexed to individuals and dyads, persist beyond the immediate prompt, influence later measurements, and be revisable when later encounters change interpretation.The required history includes who encountered what, when, in what order, and with whom.
- Relational representation: A capable model must distinguish states with identical meanings and strengths but different arrangements, because relations can reverse a meaning’s contribution while leaving it fully present.Current systems primarily store salience, frequency, and standalone sense likelihoods, leaving no explicit location for such relations.
- Implementation choice: The relational requirement is representational rather than hardware-specific: complex amplitudes, real-number pairs, density matrices, or signed weights could carry the relation if that representation changes outputs.The theory requires a place where the relation lives and a guarantee that it affects behavior, not a particular datatype.
- Shared experience: An architecture must separate shared experience from shared facts, since people can know the same information without undergoing the same encounter or can share an allusion while interpreting the work differently.The choice of shared events or texts matters as much as the information people retain.
- Evaluation: The proposed benchmark should test whether a model is changed by encounters, revises older associations, carries changes into later conversations, and combines phrases differently across relationships.The benchmark is measured over time and across relationships rather than by a single prompt.
7. Position in the literature
The paper situates its account within quantum-probability research in psychology and phase-based NLP, while distinguishing its agent- and history-indexed theory from prior persona-conditioned language-model studies.
- Quantum probability in psychology: Quantum probability entered psychology because judgments could violate ordinary probability rules, such as changing when question order changed.Psychologists emphasized that this borrowing was mathematical rather than physical, and most studies examined single judgments over seconds to minutes.
- Phase in natural language processing: NLP systems since 2018 have used phase for meaning combination, semantic matching, polarity, ambiguity, emotion, and preservation of negatively aligned components.The passage identifies complex word embeddings in 2018, a complex-valued matching network in 2019, and later Phase-Coherent Transformer and PRISM systems.
- Closest prior language-model work: Agostino and colleagues most closely approach the theory by making meaning interpreter-dependent and showing that interpretive-operation order changes results.Their experiments instead give LLM agents personas describing occupation, age, and location, treating persona as information supplied at interpretation time.
8. Objections and open problems
The section acknowledges that classical models may reproduce individual measurements and identifies unresolved empirical and formal problems concerning suppressed meanings, encounter histories, and quantum-specific order constraints. It argues that corpora constrain candidate meanings but cannot determine an individual’s meaning relations or reader-specific encounter history.
- Objections and open problems: The theory stands or falls on persistent suppressed components with valence inversion, constrained order effects, and history-indexed coordination after relevant facts leave context.It fails empirically if suppressed meanings measurably fade during quotation.
- Objections and open problems: The formalism lacks solutions to the basis problem— which meanings are live—and to identifying which encounters change a person’s state.The basis problem risks modeling only the freedom to choose meanings; encounter selection remains an unsolved problem.
- Objections and open problems: Corpora constrain candidate meanings through recorded forms and contexts but cannot determine which meanings a person holds or the relations among them that averaging removes.Meanings may become relevant through interpretation, but candidacy is the limit of corpus evidence.
- Objections and open problems: The paper narrows the missing element from world reference to a reader’s encounter history, which form-only corpora cannot provide without recording who read what.This distinction is presented as narrower than the claim that form-only systems cannot mean anything.
- Objections and open problems: A signed classical model with sufficient memory could reproduce the tested interference, order dependence, and persistence, while the theory lacks a proved quantum-specific constraint on encounter-order differences.The unresolved test is whether different work orders yield constellations whose differences fall into a constrained family.
9. Conclusion
The conclusion argues that language requires two parameters: meaning strength and history-dependent relations among coexisting meanings. Because corpus averaging preserves the first but cancels the second, future AI systems need agent- and dyad-indexed states that carry persistent relational meaning.
- Two parameters: Language combines meaning strength with relations among live meanings, allowing a word to carry a meaning and its opposite simultaneously.The paper associates this relational capacity with irony, allusion, quotation, and other figures of speech.
- Two parameters: Cadence, quotation marks, meter, lineation, and genre can signal relations among meanings, distinguishing linguistic signaling from contextual associations in other practices.A Go stone may become a quotation for someone with relevant history, but it does not announce that status.
- Limits of corpus training: Corpus averaging preserves meanings while canceling person-specific relations, so agent-deindexed pretraining cannot recover the second parameter for an individual or dyad.The systems built on corpora therefore refine estimates of the first parameter rather than relational states.
- Limits of corpus training: Next-token prediction over undifferentiated text converges toward population-average readings, and retrieval-only memory cannot replace weights trained on everyone.Without encounter provenance, the loss cannot distinguish an interpretation right for one person from one right on average.
- Proposed architecture and tests: The proposed alternative indexes training to encounter histories and carries relations between meanings, while testing whether suppression reverses contribution, shared history exceeds propositions, and encounters change model behavior.The future computational unit may couple an agent state, a particular-reader model, and a persistent shared-history record.