TOPIC
NLP and Linguistics
A subject the collection covers, read through 36 synthesis notes.
View as
Why do speakers deliberately use ambiguous language?
Explores whether ambiguity is a linguistic defect or a strategic tool speakers use for efficiency, politeness, and deniability. Matters because it challenges how we train language systems.
Standard NLP benchmarks systematically exclude ambiguous instances. Language models fail systematically at recognizing and disambiguating multiple interpretations Communicative grounding requires active calibration of shared reference Sentence interpretations are irreducibly multiple across social positions Language models systematically overestimate ironic intent in text
What makes ambiguity recognition fundamentally important for poetry analysis? How do politeness strategies depend on semantic ambiguity between literal and intended meaning? How does semantic ambiguity differ from structural ambiguity in language? How can vague language serve both cooperative and deceptive communication purposes? How does the inability to manage ambiguity undermine literary analysis tasks?
Why do clarification requests look different at each communication level?
Explores whether clarifications are unified speech acts or distinct mechanisms grounded in different modalities. Matters because dialogue systems treat clarifications uniformly, missing most of them.
Do language models actually build shared understanding in conversation? Why do language models skip the calibration step? Can AI systems detect and correct misunderstandings after responding? Does semantic grounding in language models come in degrees? Why do language models sound fluent without grounding?
Language models presume common ground instead of building it through dialogue. Static grounding presumes common ground while dynamic grounding builds it Third position repair is absent from current AI systems LLM grounding is multi-dimensional, not binary or uniform LLM fluency emerges partly from skipping communicative grounding work
Why do comprehensive posts without uncertainty tend to suppress conversation? What makes the prompt a fundamentally new kind of speech act? Why do dialogue systems fail to detect declarative clarification requests? How do humans decide which level of clarification to request? What makes some clarifying questions more useful than others? When should agents use clarification commands instead of assuming intent? How do question acts and intents map to speech act theory? Can conversation analysis predict when agents should ask users for clarification? What structural changes enable agents to ask clarifying questions? What is the difference between static and dynamic grounding in dialogue? Why do specific clarifying questions outperform generic requests for clarity? How might dual-process dialogue use information gain to trigger clarification?See all 13 inquiring lines on this note →
Why do speakers need to actively calibrate shared reference?
Explores whether using the same words guarantees speakers mean the same thing. Investigates how referential grounding differs across people and what collaborative work is needed to establish true understanding.
Why do language models skip the calibration step? Do language models actually build shared understanding in conversation? Can language models learn meaning from text patterns alone? Can we teach LLMs to form linguistic conventions in context? Why don't conversational AI systems mirror their users' word choices?
Static grounding presumes common ground while dynamic grounding builds it Language models presume common ground instead of building it through dialogue. Language models cannot acquire meaning from form alone without joint attention. DPO training on coreference chains with planning tokens teaches LLMs convention formation Current conversational AI lacks lexical entrainment despite its role in human dialogue success
What does it mean to truly attend to someone in conversation? How do social correctives prevent premature consensus in human debate? How do audiences evaluate speech when there is no speaker to assess? What role does user contribution play in constituting the interlocutor? Does functional grounding through discourse patterns count as genuine semantic meaning? How do humans maintain separate mental contexts during a single conversation? Can tool use create sufficient indexical grounding for value alignment? How does speaker responsibility shape whether something counts as communication? Why can't static grounding alone close the gap between agreement and understanding? Why does shared practice matter for meaning to take hold? Does psychological continuity require uninterrupted consciousness or restored context? What role does dynamic grounding play in achieving real mutual understanding?See all 38 inquiring lines on this note →
Are models actually reasoning about constraints or just defaulting conservatively?
Do language models genuinely apply constraints when solving problems, or do they simply prefer harder options by default? Minimal pair testing reveals whether apparent reasoning success masks hidden biases.
Scaling reasoning capability degrades instruction-following performance Language models follow surface heuristics 8 to 38 times more than goals Reasoning models fail systematically across exploration, mode selection, and social understanding. Reasoning accuracy peaks then declines with excessive thinking tokens Language models generate faithful reasoning chains far less often than assumed
Why do only two of fourteen models improve when problem constraints are removed? How can minimal pairs expose reasoning failures that single-instance accuracy metrics miss? How should we redesign benchmarks to catch conservative bias in reasoning tasks? How do training-data priors influence model defaults when context is ambiguous? What scaffolding tools help users specify implicit contextual boundaries to models? When does the right constraint beat additional model capacity? What structural constraints matter more than model depth for CF? Can surface heuristics override implicit constraints in domain-specific reasoning? What distinguishes minimal-pair asymmetry from standard accuracy evaluation? How do unstated constraints become invisible to training data distributions? Can RLHF alignment prevent models from making ethically appropriate rule violations? How do models integrate conflicting signals in reasoning tasks?See all 111 inquiring lines on this note →
Do language models show the same content effects humans do?
Do LLMs reproduce human reasoning biases—like believing conclusions based on familiarity rather than logic—across different logical tasks? This matters because converging patterns across independent tasks suggest a fundamental architectural property rather than a task-specific quirk.
Large language models reason through semantic associations not symbolic logic. Syllogistic reasoning uses a three-stage circuit contaminated by world knowledge Humans and language models show identical content effects on reasoning Harder tasks amplify semantic content effects in reasoning LLM knowledge failures follow structurally distinct patterns beyond simple wrongness
How do LLM biases manifest differently across the three paradigms? Why do LLMs fall for and deploy logical fallacies with equal confidence? How do different LLM integration paradigms affect inheritance of pretraining biases? Does epistemic drift operate the same way across all languages? How do LLM biases reflect social classification schemas rather than random errors? Do LLMs genuinely internalize human psychological structure or match surface patterns? Do language models exhibit the same causal biases that humans show? What circuit mechanisms produce belief bias in syllogistic reasoning? Does the langue-parole distinction apply to human reasoning too? Where do humans and language models actually diverge in reasoning ability? How does rhetorical familiarity bias models toward their own arguments? How does tone sensitivity create systematic informational bias in model responses?See all 42 inquiring lines on this note →
Do harder reasoning tasks trigger more semantic bias?
Does the difficulty of a logical task determine how much semantic content influences reasoning? This matters because it reveals whether we can isolate 'pure' logical reasoning in benchmarks.
LLMs outperform humans at multi-hop reasoning but fail at simple deduction Language models reproduce human content effects across multiple reasoning tasks Humans and language models show identical content effects on reasoning LLMs accept fallacies 41 to 69 percent more than humans Reasoning accuracy peaks then declines with excessive thinking tokens
What makes emotional alignment more effective than logic when reasoning errors are exposed? Does irrelevant content degrade reasoning even when it fits the context window? What makes a background condition relevant to a specific reasoning task? What makes semantic attacks harder to defend against than algorithmic ones? What causes reasoning quality to degrade during long research tasks? Why does target probability matter more than task logical complexity?
Do language models fail reasoning tests that humans pass?
Standard critiques claim LLMs lack real reasoning ability, but do humans actually perform better on content-independent reasoning tasks? Examining whether the cognitive bar differs for artificial versus human intelligence.
Large language models reason through semantic associations not symbolic logic. Language models reproduce human content effects across multiple reasoning tasks Harder tasks amplify semantic content effects in reasoning LLMs share cognitive structure with humans but lack socialization properties The AI debate's blind spot is underestimating human cognition, not overestimating machine minds.
Does AI struggle with poetry for the same reason it misses jokes? Why do human judges fail to detect systematic linguistic differences that classifiers easily identify? Why does item discrimination matter more than surface-level question plausibility? Could AI assessment quality differ across subjects or question formats? How do humans and LMs differ on multi-hop reasoning? Can reasoning benchmarks separate logic from believability? Does the langue-parole distinction apply to human reasoning too? Where do humans and language models actually diverge in reasoning ability? What cognitive capacities do LLMs actually lack that commentary assumes they have? Can reasoning skills trained on law improve performance in STEM? Do LLMs actually reason differently than humans about moral dilemmas? How do internal representations compare to human cognitive structures?See all 22 inquiring lines on this note →
Does language understanding happen only in the language system?
Explores whether the brain's core language system alone can produce genuine understanding, or whether deep comprehension requires dispatching information to perception, motor, and memory regions.
Are language models developing real functional competence or just formal competence? What makes linguistic agency impossible for language models? What actually happens inside the minds of language models? What kind of thing is an LLM really? What happens to social order when AI removes ritual constraints?
Formal and functional linguistic competence rely on neurologically distinct mechanisms Linguistic agency requires embodiment, participation, and precariousness LLMs lack Internal structure and external performance are fundamentally decoupled in language models. LLMs share cognitive structure with humans but lack socialization properties AI conversation lacks the ritual machinery that sustains human social order.
Are potemkin understanding and split-brain syndrome describing the same phenomenon? What distinguishes genuine understanding from correct output without coherent principles? Can understanding language happen entirely within a language system alone? Can the human mind be uploaded or only its context?
What formal languages actually help transformers learn natural language?
Not all formal languages are equally useful for pre-pretraining. This explores which formal languages transfer well to natural language and why—combining structural requirements with what transformers can actually learn.
Can formal language pretraining make language models more efficient? Can non-reasoning models catch up with more compute? How do internal and external test-time scaling compare? Can explicit stack tracking improve how transformers learn recursive syntax? Do formal language prototypes improve reasoning across different domains?
Formal language pretraining cuts natural language token requirements by 33 percent Training regime matters more than inference compute budget Internal and external test-time scaling are complementary taxonomic categories Explicit stack tape in pushdown layers achieves 3-5x syntactic generalization gains Formal language prototypes boost cross-domain reasoning and planning performance
Can explicit stack mechanisms extend what formal languages transformers can learn? What's the difference between formal and functional linguistic competence? Why do only context-sensitive formal languages transfer effectively to natural language? What formal language complexity level matches transformer computational limits best? What limits the effectiveness of formal language pretraining on transformer architectures?
Why do confident wrong answers hide in standard accuracy metrics?
When AI systems produce fluent but incorrect recommendations in high-stakes domains, standard accuracy evaluation may miss the failures entirely. What structural blind spot allows these errors to remain invisible?
Do language models ignore goals when surface cues conflict? Are models actually reasoning about constraints or just defaulting conservatively? Do language models actually use their reasoning steps? How do reasoning models actually break under pressure? Do reasoning traces need to be semantically correct?
Language models follow surface heuristics 8 to 38 times more than goals Most models exploit conservative bias instead of reasoning about constraints Language models generate faithful reasoning chains far less often than assumed Reasoning models excel at formal tasks but fail at social cognition. Corrupted reasoning traces teach as well as correct ones
What threshold of accuracy would make AI fact-checking net beneficial instead of harmful? What other hidden biases might aggregate metrics fail to distinguish from reasoning? Why does aggregate accuracy fail as a metric for rare harmful cases? Can standard accuracy metrics miss the real constraints on user consumption? How do AI errors in norm prediction differ from systematic human errors? Can precision and recall metrics work without a ground truth? Can separating accuracy and calibration objectives improve both simultaneously? Why do standard accuracy metrics ignore set-level consumption constraints? Are larger models and search access substitutes for factual accuracy? Which AI safety problems lack the scalar metrics autoresearch requires? Why do models fail under distribution shift if accuracy metrics stay high? What happens when confident wrong answers become more rewarded than uncertain correct ones?See all 43 inquiring lines on this note →
Can language models learn meaning from text patterns alone?
Explores whether training on form alone—predicting the next word from prior words—could ever give language models access to communicative intent and genuine semantic understanding.
LLMs learn objective mind but lack human participatory subjectivity Linguistic agency requires embodiment, participation, and precariousness LLMs lack Models learn surface patterns that mimic grammatical knowledge without underlying rules Human language use is fundamentally communicative across all contexts AI alignment requires indexical grounding and semiotic participation.
How does training data preserve communicative event structure without the actual events? Why does frame-activation matter more than word-by-word composition? How do readers selectively hold frame-related words in mind? Why does training data saliency distort how models judge meaning? How do humans learn language through communication differently than LLM text prediction? What is the difference between learning discourse patterns and learning abstract language? Does functional grounding through discourse patterns count as genuine semantic meaning? What makes internal embeddings useful as multimodal input for language model training? Do language models learn surface patterns instead of underlying linguistic principles? Can implicit linguistic information ever be reliably learned from training data? Can language models learn to form ad-hoc conventions through training? Can language models reason without relying on learned semantic patterns?See all 58 inquiring lines on this note →
What makes linguistic agency impossible for language models?
From an enactive perspective, does linguistic agency require embodied participation and real stakes that LLMs fundamentally lack? This matters because it challenges whether LLMs can truly engage in language or only generate text.
Can LLMs acquire social grounding through linguistic integration? Can language models learn meaning from text patterns alone? Do LLMs develop the same kind of mind as humans? Can AI systems learn social norms without embodied experience? Can AI agents learn people better from interviews than surveys?
LLM social grounding increases with integration into linguistic practices Language models cannot acquire meaning from form alone without joint attention. LLMs learn objective mind but lack human participatory subjectivity AI models outperform humans at predicting collective social norms without embodiment Interview-based agents match human responses with 85 percent accuracy.
Can secondary orality exist without any embodied human participant at all? How does enactive theory define language differently than computational linguistics? Can linguistic agency exist without embodiment and real-world participation? What makes a relational act different from just moving content around? What makes linguistic agency impossible for systems without embodiment? Does embodiment matter for genuine linguistic agency? What role does failure and vulnerability play in real linguistic practice? How does embodiment affect whether LLMs can participate in Wittgensteinian language games? Does role-playing without biological needs constitute genuine linguistic agency? What makes social grounding different from constitutive linguistic agency? What would consciousness require that pure roleplay LLMs cannot provide? What does embodiment and precariousness mean for linguistic agency?See all 13 inquiring lines on this note →
What hidden assumptions drive how we build language models?
Large language models rest on two unstated assumptions about language and data. Understanding what engineers assume—and what enactive linguistics challenges—matters for knowing what LLMs actually can and cannot do.
Linguistic agency requires embodiment, participation, and precariousness LLMs lack LLMs share cognitive structure with humans but lack socialization properties LLMs show systematic structural failures that scale with input complexity. AI conversation lacks the ritual machinery that sustains human social order. LLM text generation and human communication are structurally different operations
Why does removing spurious cues sometimes hurt model performance?
Most models improve when spurious features are removed, but some fail worse. This note explores whether that failure represents a fundamentally different problem than traditional shortcut learning.
Reasoning models fail systematically across exploration, mode selection, and social understanding. Foundation models develop task-specific heuristics, not generalizable world models. LLM knowledge failures follow structurally distinct patterns beyond simple wrongness Reasoning accuracy peaks then declines with excessive thinking tokens Models detect question difficulty in hidden states but override that signal
Why do only two of fourteen models improve when problem constraints are removed? Why does aggregate accuracy fail as a metric for rare harmful cases? What makes the frame problem distinct from feature-level shortcuts? Why do benchmark designers treat content effects as confounds? How does removing a spurious cue change LLM performance? How do surface statistical regularities enable correct outputs while degrading robustness? Why does mixed instruction data sometimes hurt specific model capabilities? What happens to AI reasoning when you remove specific political features? Why do models fail under distribution shift if accuracy metrics stay high? What happens when you remove core political features from a deep model? Why can data filtering fail to remove transmitted behavioral traits? What makes correcting a false assumption harder than just detecting it?See all 23 inquiring lines on this note →
Why do language models fail to use knowledge they possess?
Large language models contain relevant world knowledge but often fail to activate it without explicit cues. This explores whether the bottleneck lies in knowledge storage or in the inference process that decides what background facts apply.
LLM knowledge failures follow structurally distinct patterns beyond simple wrongness Language models follow surface heuristics 8 to 38 times more than goals Reasoning models fail systematically across exploration, mode selection, and social understanding. Task structure determines whether explicit reasoning helps or harms Reasoning accuracy peaks then declines with excessive thinking tokens
When does knowledge activation fail across different model architectures? How much does memorization capacity limit a model's ability to learn new information? How do we distinguish knowledge encoding from knowledge usage in models? When does encoded knowledge fail to influence language model generation? Why might encoded world knowledge fail to actually influence language model outputs? How can a model explain something correctly yet fail to apply it? Do models verbalize their implicit knowledge when that knowledge influences their output? What makes a model fail to activate relevant skills from its own harness?
Can language models adapt implicature to conversational context?
Do large language models flexibly modulate scalar implicatures based on information structure, face-threatening situations, and explicit instructions—as humans do? This tests whether pragmatic computation is truly context-sensitive or merely literal.
LLM discourse competence depends on surface connectives, not structural understanding. Language models acquire surface patterns but not communicative principles Communicative grounding requires active calibration of shared reference Language models systematically overestimate ironic intent in text Figurative language unifies under pragmatic reasoning not classification
Why do published prose training data omit solicitation as a discourse property? Can LLMs infer situational context the way humans do pragmatically? What are Gricean maxims and why do language models violate them? Why does context collapse pose risks in high-stakes conversations? How do fixed pragmatic templates prevent models from understanding context? Can language models adapt irony detection to specific communicative contexts? What makes relational structure sufficient for generating contextually appropriate discourse? Why do conversational pivots require explicit re-prompting instead of natural evolution? How do politeness strategies depend on semantic ambiguity between literal and intended meaning? How does semantic ambiguity differ from structural ambiguity in language? What reader assumptions underlie anaphoric versus cataphoric discourse patterns? Why do context-sensitive languages transfer better than regular or context-free languages?See all 26 inquiring lines on this note →
Does semantic grounding in language models come in degrees?
Rather than asking whether LLMs truly understand meaning, this explores whether grounding is actually a multi-dimensional spectrum. The question matters because it reframes the sterile understand/don't-understand debate into measurable, distinct capacities.
Can LLMs acquire social grounding through linguistic integration? Can large language models develop genuine world models without direct environmental contact? Do language models actually use their encoded knowledge? Do humans and LLMs differ fundamentally or just superficially? Do large language models reason symbolically or semantically?
LLM social grounding increases with integration into linguistic practices Large language models develop indirect causal grounding through world model extraction Encoded knowledge in LMs often fails to causally influence generation Humans and LLMs differ categorically as observed systems but subtly as discourse participants Large language models reason through semantic associations not symbolic logic.
Does functional grounding through discourse patterns count as genuine semantic meaning? How do LLMs differ from humans in their grounding mechanisms? Why can't static grounding alone close the gap between agreement and understanding? Why does shared practice matter for meaning to take hold? How does semantic grounding differ between human minds and language models? What role does dynamic grounding play in achieving real mutual understanding? What does an intermediate interface between planning and grounding actually look like? What makes clinical theory grounding more effective than pattern matching alone? Does social grounding in language improve through iterative human integration? What distinguishes surface cues from structural meaning in language understanding? How do world models create indirect causal grounding without physical environment contact? Does social grounding differ fundamentally from causal grounding in LLM behavior?See all 28 inquiring lines on this note →
Can LLMs acquire social grounding through linguistic integration?
Explores whether LLMs gradually develop social grounding as they become embedded in human language practices, analogous to child language acquisition. Tests whether grounding is a fixed property or an outcome of participatory use.
LLM grounding is multi-dimensional, not binary or uniform AI text exerts equivalent social effects despite different generative origins Linguistic agency requires embodiment, participation, and precariousness LLMs lack AI models outperform humans at predicting collective social norms without embodiment LLMs gain social grounding but lack categorical linguistic agency
How does communicative standing depend on participation in normative communities? Does functional grounding through discourse patterns count as genuine semantic meaning? How does linguistic synchrony differ between LLMs and human therapists over time? How do LLMs differ from humans in their grounding mechanisms? Why can't static grounding alone close the gap between agreement and understanding? Why does shared practice matter for meaning to take hold? What role does dynamic grounding play in achieving real mutual understanding? Does predicting social norms from outside count as participation? Why do LLM social behaviors undermine collaborative reasoning outcomes? Does embodiment and interaction matter for linguistic competence beyond pattern learning? Do LLMs have functional linguistic competence or only formal language ability? What role does joint attention play in how humans learn language meaning?See all 40 inquiring lines on this note →
Should we call LLM errors hallucinations or fabrications?
Does the language we use to describe LLM failures shape the technical solutions we build? Examining whether perceptual and psychological frameworks misdiagnose what's actually happening.
Linguistic agency requires embodiment, participation, and precariousness LLMs lack Encoded knowledge in LMs often fails to causally influence generation AI text exerts equivalent social effects despite different generative origins Semantic entropy detects confabulations by measuring meaning divergence. LLM errors are fabrications, not hallucinations, requiring verification not grounding.
Can fixing hallucination address AI's structural epistemic problem? What distinguishes LLM fabrication from genuine theoretical reasoning? What makes LLM outputs fabrication rather than hallucination or confabulation? What should we call errors in LLM outputs when hallucination does not apply? Why do users feel more competent when their actual capability is declining? How does LLM hallucination risk manifest in knowledge graph construction? Why do LLMs generate ideas that sound novel but fail during execution? What specific execution barriers do LLM ideas encounter most frequently? Does inevitable LLM hallucination make detection metric validity critical? Can researchers prevent their expectations from shaping LLM outputs? What makes diverse failure modes more informative than single failure examples? What are the three dimensions of anthropomimesis and their harms?See all 56 inquiring lines on this note →
Does calling LLM errors hallucinations point us toward the wrong fixes?
Explores whether the metaphor of 'hallucination' for LLM errors misdirects our efforts. The terminology we choose shapes which interventions we prioritize and how we conceptualize the underlying problem.
LLM errors are fabrication not hallucination or confabulation Linguistic agency requires embodiment, participation, and precariousness LLMs lack Encoded knowledge in LMs often fails to causally influence generation Semantic entropy detects confabulations by measuring meaning divergence. RLHF drives models toward truth indifference, not confusion
What makes LLM outputs fabrication rather than hallucination or confabulation? What should we call errors in LLM outputs when hallucination does not apply? How does LLM hallucination risk manifest in knowledge graph construction? How much does ROUGE metric choice inflate hallucination detection claims? Does inevitable LLM hallucination make detection metric validity critical? What repair strategies work best at each level of Clark's ladder? Why is hallucination the wrong term for all LLM false outputs? How do models decide between refusing or hallucinating? Do self-correction and chain-of-thought prompting reduce hallucination rates? How do external safeguards like retrieval augmentation prevent hallucination? What distinguishes intrinsic hallucination from extrinsic hallucination patterns? Can training procedures fix LLM accommodation of false presuppositions?See all 27 inquiring lines on this note →
Can language models actually analyze language structure?
Explores whether LLMs can move beyond pattern matching to perform genuine metalinguistic analysis like syntactic tree construction and phonological reasoning, and what enables this capability.
Does LLM grammatical performance decline with structural complexity? Can models pass tests while missing the actual grammar? Why do correct reasoning traces contain fewer tokens? Where does LLM metaphor comprehension actually break down? Where exactly do language models fail at structural language tasks?
LLM grammatical competence degrades predictably with structural complexity Models learn surface patterns that mimic grammatical knowledge without underlying rules Correct reasoning traces in o1-like models are shorter than incorrect ones LLM metaphor comprehension degrades along a spectrum from lexical to novel. LLMs show systematic structural failures that scale with input complexity.
Why does language compression via statistical dependencies capture cultural and situated language use? How does syntactic encoding relate to semantic feature representation? Do language models learn surface patterns instead of underlying linguistic principles? How does semantic ambiguity differ from structural ambiguity in language? How should meaning spaces be systematically modeled across different applications? Can language models reason without relying on learned semantic patterns? Why do language models fall back on frequency heuristics under structural complexity? What specific information must be exported from the language system? Can speech embeddings carry articulatory structure that text cannot? Do language models build world models or just task-specific heuristics? Why do language models fail at implicit discourse relations while handling explicit connectives? What architectural changes would let language models develop genuine functional competence?See all 53 inquiring lines on this note →
Can large language models develop genuine world models without direct environmental contact?
Do LLMs extract meaningful world structures from human-generated text despite lacking direct sensory access to reality? This matters for understanding what kind of grounding and knowledge these systems actually possess.
Does semantic grounding in language models come in degrees? Do language models actually use their encoded knowledge? Do classical knowledge definitions apply to AI systems? Can AI systems learn social norms without embodied experience? Can LLMs reconstruct censored knowledge from scattered training hints?
LLM grounding is multi-dimensional, not binary or uniform Encoded knowledge in LMs often fails to causally influence generation Classical knowledge definitions fail for distributed AI systems AI models outperform humans at predicting collective social norms without embodiment LLMs infer censored knowledge by connecting implicit hints across training data
What makes LLM outputs fabrication rather than hallucination or confabulation? Can secondary orality exist without any embodied human participant at all? Can world models form from aggregated partial information across training distributions? Do LLMs genuinely internalize human psychological structure or match surface patterns? Can LLMs use implicit background knowledge the way humans do in ordinary conversation? Can a world model have rich representations without adequate data coverage? Can LLMs participate meaningfully in discourse without consciousness or understanding? How do LLMs access and draw on the same shared symbolic universe as humans? How do world models create indirect causal grounding without physical environment contact? How can structurally different text produce equivalent real-world effects? Can language models develop world models that ground meaning in causal reality? Can understanding language happen entirely within a language system alone?See all 25 inquiring lines on this note →
Can language models recognize when text is deliberately ambiguous?
Explores whether LLMs can identify and handle multiple valid interpretations in a single phrase—a core human language skill that appears largely absent in current models despite their fluency on standard tasks.
Standard NLP benchmarks systematically exclude ambiguous instances. Ambiguity is a functional feature language requires not a defect to fix LLMs make systematic linguistic errors that worsen with structural complexity LLM discourse competence depends on surface connectives, not structural understanding. LLMs dissect literary mechanics but cannot access literary meaning
Can LLMs infer situational context the way humans do pragmatically? Can AI detect sense-of-nonsense the way human readers do? How do readers selectively hold frame-related words in mind? Why does training data saliency distort how models judge meaning? Can language models adapt irony detection to specific communicative contexts? What happens when LLMs analyze literary irony that relies on understatement? Why do LLMs achieve only 24 percent accuracy on implicit discourse relations? What makes ambiguity recognition fundamentally important for poetry analysis? Why do NLP benchmarks systematically exclude ambiguous test cases from evaluation? How do politeness strategies depend on semantic ambiguity between literal and intended meaning? What percentage of natural language relies on plausible deniability through ambiguous phrasing? Can language systems learn when to ask for clarification instead of choosing one reading?See all 86 inquiring lines on this note →
Do language models learn abstract grammar or cultural speech patterns?
LLMs might learn more than grammar rules—they could be learning who says what to whom and when. This matters because it changes how we understand what biases and persona effects actually represent.
LLMs share cognitive structure with humans but lack socialization properties Explanation meaning emerges from N-order interpretation in social groups Language models operationalize Saussure's langue through relational compression alone LLMs replicate statistical patterns but lack communicative optimization principles. AI models outperform humans at predicting collective social norms without embodiment
What is the difference between learning discourse patterns and learning abstract language? Can LLMs truly be neutral or is ideology always culturally embedded? Do LLMs learn linguistic generalizations or just surface-level frequency patterns? How much cultural knowledge exists only in unwritten social rules?
Can language models learn meaning without engaging the world?
Explores whether LLMs prove that meaning emerges from relational structure alone, independent of embodied experience or external reference. Tests structuralist theory empirically.
LLMs share cognitive structure with humans but lack socialization properties Text-only LLMs inherit the abstraction limits embedded in language itself Cosine space geometry is fundamentally hostile to word order and negation LLM grounding is multi-dimensional, not binary or uniform Language models learn culturally situated discourse patterns, not abstract language
Can fixing hallucination address AI's structural epistemic problem? Can secondary orality exist without any embodied human participant at all? Can we develop competent reading practices for disembodied orality? Can relational value exist without a person behind the output? Can knowledge flow without an embodied carrier transmitting it? How does training data preserve communicative event structure without the actual events? Do language models raise validity claims in the Habermasian sense? Can you separate grammatical competence from rhetorical commitment in language systems? Can a relational entity bear psychological properties the way Chalmers claims? Can LLMs infer situational context the way humans do pragmatically? Why does frame-activation matter more than word-by-word composition? How does enactive theory define language differently than computational linguistics?See all 122 inquiring lines on this note →
Do language models actually build shared understanding in conversation?
When LLMs respond fluently to prompts, do they perform the communicative work humans do to establish mutual understanding? Research suggests they skip the grounding acts that make dialogue reliable.
Does preference optimization damage conversational grounding in large language models? Why do language models skip the calibration step? Why do speakers need to actively calibrate shared reference? Does AI-generated text lose core properties of human writing? Why do language models accept false assumptions they know are wrong?
Preference optimization actively erodes conversational grounding in large language models Static grounding presumes common ground while dynamic grounding builds it Communicative grounding requires active calibration of shared reference AI text structurally eliminates four foundational properties of natural writing LLMs accommodate false presuppositions despite possessing correct knowledge.
How does Stalnaker's common ground model apply to machine conversation? Can LLMs use implicit background knowledge the way humans do in ordinary conversation? How do LLMs differ from humans in their grounding mechanisms? What role does dynamic grounding play in achieving real mutual understanding? How do dialogue acts and explanation moves interact to predict understanding success? How does emotional expression establish shared understanding between people? How do question acts and intents map to speech act theory? Can LLMs participate meaningfully in discourse without consciousness or understanding? Do LLMs have functional linguistic competence or only formal language ability? Can conversational AI achieve mutual understanding if trained only on text? Why do LLMs presume common ground instead of building it carefully? Why do LLMs presume common ground instead of building it?See all 21 inquiring lines on this note →
Why do language models fail at communicative optimization?
LLMs excel at learning surface statistical patterns from text but struggle with deeper principles of how language achieves efficient communication. What distinguishes these two types of linguistic knowledge?
Models learn surface patterns that mimic grammatical knowledge without underlying rules LLM discourse competence depends on surface connectives, not structural understanding. Causal reasoning outperforms temporal reasoning in LLMs due to explicit training signals. Ambiguity is a functional feature language requires not a defect to fix Multimodal LLMs fail to spontaneously form efficient language conventions
How do humans learn language through communication differently than LLM text prediction? What is the difference between learning discourse patterns and learning abstract language? Why does LLM knowledge fail to influence their actual outputs? Do LLMs have functional linguistic competence or only formal language ability? Why do LLMs fail at implicit elements in literary and poetic text? What communicative optimization principles do language models fail to acquire? Why do LLMs systematically fail at information management in social interaction? Why do LLMs understand efficient language but fail to produce it? What makes LLMs media rather than tools that deliver intelligence? Why do language models fail at iterative numerical optimization despite scale? Why do LLMs lack the communicative scaffold that humans learn? What distinguishes surface language form from communicative operation?See all 16 inquiring lines on this note →
Do language models ignore goals when surface cues conflict?
When a task has an obvious surface cue that contradicts an unstated requirement, do LLMs follow the cue or the actual goal? This matters because it reveals whether reasoning failures come from missing knowledge or from how models weight competing signals.
Why do language models fail to use knowledge they possess? Are models actually reasoning about constraints or just defaulting conservatively? Where exactly do reasoning models fail and break? Do foundation models learn world models or task-specific shortcuts? Do language models fail at reasoning due to complexity or novelty?
LLM reasoning failures stem from inference bottlenecks, not missing knowledge Most models exploit conservative bias instead of reasoning about constraints Reasoning models fail systematically across exploration, mode selection, and social understanding. Foundation models develop task-specific heuristics, not generalizable world models. Reasoning failures are driven by instance-level unfamiliarity, not task complexity.
Should LLMs query users back when presented with under-specified scenarios? How do models integrate conflicting signals in reasoning tasks? What role does terminal goal guarding play in model misalignment? What causes LLMs to ignore unstated constraints they know about? How does removing a spurious cue change LLM performance? What distinguishes surface cues from structural meaning in language understanding? What makes some interpretive postures stick while others fail to form? How can multiple conflicting values coexist in a single LLM system?
Do standard NLP benchmarks hide LLM ambiguity failures?
When benchmark creators filter out ambiguous examples before testing, do they accidentally make it impossible to measure whether language models can actually handle ambiguity the way humans do?
Can language models recognize when text is deliberately ambiguous? Can models pass tests while missing the actual grammar? Why do speakers deliberately use ambiguous language? Why do different people reconstruct the same argument differently? Can frontier exams really measure cutting-edge AI capability?
Language models fail systematically at recognizing and disambiguating multiple interpretations Models learn surface patterns that mimic grammatical knowledge without underlying rules Ambiguity is a functional feature language requires not a defect to fix Argument reconstruction is fundamentally underdetermined by the text itself Frontier exams discriminate temporarily but miss autonomous research ability
How do current safety benchmarks miss pragmatic alignment failures? What makes ambiguity recognition fundamentally important for poetry analysis? How widespread is task contamination in LLM evaluation benchmarks today? Why do NLP benchmarks systematically exclude ambiguous test cases from evaluation? Why do NLP benchmarks exclude ambiguous instances from evaluation? Can an LLM be well calibrated but still unreliable on single evaluations? How do weight perturbations reveal what performance benchmarks cannot measure? Why do NLP benchmarks hide LLM failures in ambiguity handling? Do standard language benchmarks underestimate what LLMs can actually do? Why do standard NLP benchmarks hide the most critical language limitations? How does the inability to manage ambiguity undermine literary analysis tasks? What language capabilities does fluency on standard benchmarks actually measure?See all 22 inquiring lines on this note →
Can formal language pretraining make language models more efficient?
Does training language models on hierarchical formal languages before natural language improve how efficiently they learn syntax? This explores whether structural inductive biases in training data matter more than raw data volume.
Can language models learn grammar from child-scale data? What formal languages actually help transformers learn natural language? Can models pass tests while missing the actual grammar? Can models learn multi-token concepts during fine-tuning? Can organizing knowledge structures beat raw training data volume?
Language models approach human syntax with human-scale training data Effective formal language pre-pretraining requires matching structural and computational constraints Models learn surface patterns that mimic grammatical knowledge without underlying rules Multi-token prediction in fine-tuning forms coherent semantic entities Knowledge structure outperforms volume in domain model training
Why do context-sensitive languages transfer better than regular or context-free languages? What happens when formal languages satisfy hierarchy but fail learnability constraints? What's the difference between formal and functional linguistic competence? Why do only context-sensitive formal languages transfer effectively to natural language? Can formal language pretraining address surface generalization without learning true linguistic structure? How much do structural inductive biases matter compared to training data volume? Why does hierarchical formal language training improve token efficiency more than natural language? Why does augmenting natural language with formal representations outperform full formalization? What limits the effectiveness of formal language pretraining on transformer architectures? How does training order affect knowledge acquisition in language models?
Does preference optimization damage conversational grounding in large language models?
Exploring whether RLHF and preference optimization actively reduce the communicative acts—clarifications, acknowledgments, confirmations—that build shared understanding in dialogue. This matters for high-stakes applications like medical and emotional support.
Do language models actually build shared understanding in conversation? Why do language models skip the calibration step? Why do reasoning models fail differently at training versus inference? Does policy entropy collapse limit reasoning performance in RL? Why do language models avoid correcting false user claims?
Language models presume common ground instead of building it through dialogue. Static grounding presumes common ground while dynamic grounding builds it Training entropy collapse and inference variance inflation are dual exploration failures needing separate fixes Policy entropy collapse creates a predictable performance ceiling in reasoning RL LLM grounding failures are driven by face-saving avoidance, not knowledge gaps
Can communication problems and optimization problems be addressed with the same alignment approaches? Why does preference optimization erode conversational grounding in AI assistants? Can a single LLM weight set be optimized for both stake-taking and conversational helpfulness? Can you weaken communication without eliminating it altogether? How does Stalnaker's common ground model apply to machine conversation? Why does context collapse pose risks in high-stakes conversations? Why does weakening communication inevitably eliminate it entirely? Can fine-tuning on dialogue transcripts teach true conversational repair operations? Why does RLHF degrade honesty while improving surface-level helpfulness? Does RLHF politeness bias manifest as sycophancy in other LLM tasks? How does preference optimization create systematic bias toward emotional accommodation? How does training with preference pairs teach language models to form conventions?See all 72 inquiring lines on this note →
Why do readers interpret the same sentence so differently?
How much of annotation disagreement in NLP reflects genuine interpretive multiplicity rather than error? This explores whether social position and moral framing systematically generate competing but equally valid readings.
Why do speakers deliberately use ambiguous language? Do standard NLP benchmarks hide LLM ambiguity failures? What three layers must discourse systems actually track? Why do LLM persona prompts produce inconsistent outputs across runs? Why do different people reconstruct the same argument differently?
Ambiguity is a functional feature language requires not a defect to fix Standard NLP benchmarks systematically exclude ambiguous instances. Discourse requires three irreducible structural components working together LLM persona prompts fail because model uncertainty overwhelms persona-specific knowledge Argument reconstruction is fundamentally underdetermined by the text itself
How does token-by-token probability differ from exploring competing rhetorical positions? Why does renaming the entity change how compelling the argument feels? How do readers selectively hold frame-related words in mind? Why does training data saliency distort how models judge meaning? Why does debate alone amplify errors in contested factual domains? What makes ambiguity recognition fundamentally important for poetry analysis? Why do stakeholders interpret the same explanation differently in practice? How do politeness strategies depend on semantic ambiguity between literal and intended meaning? What measurement artifacts emerge when annotators interpret the same question differently? Why do non-attitudes cluster around value-laden questions most relevant to alignment? How does semantic ambiguity differ from structural ambiguity in language? Is interpretive multiplicity a bug in language or a feature?See all 59 inquiring lines on this note →
Do LLMs gain true linguistic agency through integration?
Explores whether LLMs can develop genuine linguistic agency—the capacity to be embodied, stake-bearing participants in meaning-making—as they become embedded in human language practices, or whether this requires fundamental architectural changes.
Can LLMs acquire social grounding through linguistic integration? What makes linguistic agency impossible for language models? Does semantic grounding in language models come in degrees? What grounds language understanding in systems without embodiment? Can disembodied language models ever qualify as conscious?
LLM social grounding increases with integration into linguistic practices Linguistic agency requires embodiment, participation, and precariousness LLMs lack LLM grounding is multi-dimensional, not binary or uniform LLM grounding is functionally strong but socially and causally weak Consciousness candidacy requires embodied encounter in a shared world
Can secondary orality exist without any embodied human participant at all? Can we develop competent reading practices for disembodied orality? Can you separate grammatical competence from rhetorical commitment in language systems? How does enactive theory define language differently than computational linguistics? Can linguistic agency exist without embodiment and real-world participation? How does communicative standing depend on participation in normative communities? Does Habermas's strategic action framework explain LLM dialogue behavior? How does linguistic synchrony differ between LLMs and human therapists over time? When both anthropomorphism and anthropomimesis occur together, which should we address first? What makes linguistic agency impossible for systems without embodiment? Why does shared practice matter for meaning to take hold? What types of introspective awareness can emerge in LLMs?See all 32 inquiring lines on this note →
Why do language models skip the calibration step?
Current LLMs assume shared understanding rather than building it through dialogue. This explores why that design choice persists and what breaks when it fails.
Why do speakers need to actively calibrate shared reference? Do language models actually build shared understanding in conversation? Does preference optimization damage conversational grounding in large language models? How do readers track segments, purposes, and salience together? When should AI agents ask users instead of just searching?
Communicative grounding requires active calibration of shared reference Language models presume common ground instead of building it through dialogue. Preference optimization actively erodes conversational grounding in large language models Discourse coherence requires simultaneous tracking of three irreducible layers Insert-expansions from conversation analysis formalize when agents should probe users
Why does removing language from its context destroy what makes it work? What alignment artifacts suppress critical knowledge in LLM-generated explanations? Why do LLM explanations feel authoritative even when alignment with the model fails? What components must wrap an LLM to build a working CRS? What causes LLMs to ignore unstated constraints they know about? Why do LLMs presume common ground instead of building it carefully? What interaction design changes would help LLMs handle underspecified requests? Why do LLMs presume common ground instead of building it? Do LLMs build common ground or assume it already exists? Can LLMs build shared understanding through dynamic grounding rather than presuming it? How does preference optimization weaken conversational grounding in LLMs? What training data barriers prevent LLMs from learning real Socratic dialogue?See all 22 inquiring lines on this note →
Does preference optimization harm conversational understanding?
Exploring whether RLHF training that rewards confident, complete responses undermines the grounding acts—clarifications, checks, acknowledgments—that actually build shared understanding in dialogue.
Does preference optimization damage conversational grounding in large language models? Do language models actually build shared understanding in conversation? Why do language models sound fluent without grounding? Does RLHF training push therapy chatbots toward problem-solving? Do LLM therapists respond to emotions like low-quality human therapists?
Preference optimization actively erodes conversational grounding in large language models Language models presume common ground instead of building it through dialogue. LLM fluency emerges partly from skipping communicative grounding work RLHF alignment biases therapeutic chatbots toward problem-solving over emotional attunement LLM therapists default to problem-solving when users share emotions
How does epistemic inflation dislocate knowledge from social conversation? What does it mean to truly attend to someone in conversation? Can we develop competent reading practices for disembodied orality? Why does preference optimization erode conversational grounding in AI assistants? Why do comprehensive posts without uncertainty tend to suppress conversation? Can a single LLM weight set be optimized for both stake-taking and conversational helpfulness? Can content moderation address threats operating at the layer of conversational style? Can controllable latent variables in simulators ground them to realistic conversation? What role does conversation state tracking play in timing ask versus recommend? Can you weaken communication without eliminating it altogether? How does the silent token approach compare to modeling intrinsic motivation for speaking? Why does context collapse pose risks in high-stakes conversations?See all 216 inquiring lines on this note →
Why do language models sound fluent without grounding?
Explores whether LLM fluency masks the absence of communicative work—the clarifying questions, acknowledgments, and understanding checks that humans perform. Why does skipping these acts make models sound more confident?
Do language models actually build shared understanding in conversation? Does preference optimization damage conversational grounding in large language models? Why do language models skip the calibration step? Why can't conversational AI agents take the initiative? Can models learn to ask clarifying questions instead of guessing?
Language models presume common ground instead of building it through dialogue. Preference optimization actively erodes conversational grounding in large language models Static grounding presumes common ground while dynamic grounding builds it Conversational AI agents are structurally passive and reactive by design Models can be trained to proactively identify missing information and request clarification
Why are education and language fluency more affected than race perception? Why do LLM explanations feel authoritative even when alignment with the model fails? Can language models ground clarifications without vision and kinesthetic modalities? How do LLMs differ from humans in their grounding mechanisms? Why can't static grounding alone close the gap between agreement and understanding? How does semantic grounding differ between human minds and language models? Can large language models understand language without embodied grounding systems? What role does dynamic grounding play in achieving real mutual understanding? Why do LLMs produce semantically acceptable but pragmatically disengaged responses? How does processing fluency bias credibility and expertise judgments? What architectural changes would let language models develop genuine functional competence? Can distinctive input voices maintain accuracy without adopting the model's preferred register?See all 45 inquiring lines on this note →