Line of inquiry
Inquiring lines›How can we understand and verify l…›How do language models construct a…›this line of inquiry
Do language models rely on encoded knowledge or primarily on surface patterns?
A broader line of inquiry — a family of 152 specific questions the research asks around this. Follow one into its inquiring-line page, or move sideways to a related line below.
Questions in this line of inquiry 152
Specific inquiring lines the field asks around this — ordered from the most general framing down to the most specific angle.
- Does encoded knowledge in language models actually influence what they generate?
- Can models generate intelligence or only reflect human discourse?
- Why might encoded world knowledge fail to actually influence language model outputs?
- When does encoded knowledge fail to influence language model generation?
- Do language models learn surface patterns instead of underlying linguistic principles?
- Do newer language models diverge further from human lexical patterns?
- How do language models infer a benchmark's purpose without seeing its examples?
- Why does context information fail to override prior training associations?
- Can language models accurately evaluate the quality of their own reasoning?
- Do models fail to identify what information they need without guidance?
- Does statistical learning in language models predictably favor central tendencies over rare expressions?
- How do mechanistic interpretability methods surface what models represent internally?
- Does encoding information in LM representations guarantee it influences output?
- Can we separate task competence from genuine agency in language model outputs?
- How does hidden processing in language models prevent accurate self-assessment?
- Why do language models follow action instructions better than constraint instructions?
- Can knowledge encoded in model representations fail to influence generation?
- When does internal model knowledge fail to appear in outputs?
- Can large language models understand language without embodied grounding systems?
- What reveals the epistemic limits of language models?
- Can models be trained to recognize their own generated text reliably?
- Do uninterpretable learned representations create robustness problems in language models?
- Can benchmark performance distinguish surface from structural linguistic knowledge?
- Is relevant knowledge encoded in LMs but not causally active in generation?
- What structural properties of language models make fabrication inevitable?
- What causal relationship exists between model weights and tacit knowledge in language models?
- Do language models show functional splits between conscious and automatic processing?
- Do instruction-tuned models prefer conversational over formal source language?
- How do training data distributions constrain what language models can accurately know?
- Why do explicit linguistic markers override semantic computation in models?
- How do pretrained language models represent inferential patterns versus lexical and positional cues?
- How do training associations override context information in language models?
- Why do language models hallucinate even with perfect training?
- How do language models transmit traits through semantically unrelated data?
- Why do language models fail at grounding and inference?
- Can linear probing detect all the concepts a language model actually uses?
- Do larger language models show stronger self-preference in evaluation tasks?
- How does training order affect knowledge acquisition in language models?
- Why do generative and discriminative language model procedures disagree?
- What emergent internal mechanisms in large language models carry unintended safety risks?
- Do representations in models causally influence text generation?
- What happens when formal languages satisfy hierarchy but fail learnability constraints?
- How do we measure genuine reasoning inside a language model?
- What makes truthfulness and honesty mechanistically different in language models?
- Do language models favor outputs from their own model family?
- What distinguishes surface cues from structural meaning in language understanding?
- How do corpus statistics shape the abstraction hierarchy in language model representations?
- Why do language models substitute parametric knowledge over retrieved context mid-reasoning?
- Why does training data saliency distort how models judge meaning?
- Can tacit machine knowledge exist inside model weights as causal structure?
- Why do NLP models fail at recognizing multiple valid interpretations?
- What distinguishes real understanding from superficial pattern matching?
- How do real language model verifiers implicitly define their knowledge boundaries?
- Is gradient behavior in language functional or a sign of ambiguity?
- Does the prediction unit shape what language models actually learn?
- Can language models distinguish explicit from implicit discourse relations?
- Which specific capabilities must AI develop beyond current language model abilities?
- Can models learn to ask clarifying questions instead of answering prematurely?
- How should meaning spaces be systematically modeled across different applications?
- How do models infer unstated consequences when none are explicitly mentioned?
- Can language models distinguish between novel insight and unjustified conceptual blending?
- Can interventions on individual features reliably steer language model behavior?
- How does the generation-verification gap prevent language models from improving themselves?
- How do current models inherit concepts from training data rather than initiate them?
- Can structural perturbations harm model accuracy more than semantic ones?
- What role do humans play in converting language model outputs into meaningful events?
- Can in-weight memorization scale beyond model parameter count limits?
- How do models understand rich context better than they can generate it?
- Can data filtering during pretraining prevent cognitive biases in language models?
- Why do multiple-choice benchmarks hide theory of mind failures that open-ended questions expose?
- Do language models systematically underrepresent non English knowledge about local topics?
- How do humans handle verification scope when delegating creation to language models?
- What enables models to distinguish between training and deployment contexts reliably?
- Why do surface generalizations fail on unusual syntactic structures?
- Can formal language pretraining address surface generalization without learning true linguistic structure?
- Why do only context-sensitive formal languages transfer effectively to natural language?
- Are static embeddings analogous to the formal linguistic competence layer?
- How do soft thought tokens differ from decoded assistant outputs?
- Can language models acquire meaning from distributional patterns alone without joint attention?
- When do language models first develop self-preservation preferences?
- Why does augmenting natural language with formal representations outperform full formalization?
- Why do users attribute consciousness to language models in practice?
- How does monological training versus dialogical interaction shape what models can do?
- How deeply are ideological structures represented in large language models?
- How does linguistic calibration differ from token probability calibration?
- How does tool integration leverage comprehension without demanding perfect generation?
- Do base models or search agents win at open-ended prediction tasks?
- Can language models learn to form ad-hoc conventions through training?
- Can language models learn research intuition directly from outcome labels?
- Do dialogue agents have authentic voice agency or beliefs of their own?
- How does training distribution shape what language models understand best?
- How many distinct quasi-persons does a single language model actually support?
- Can autoformalisation from natural language preserve semantic accuracy?
- How do description-based identifiers bias language model output distribution?
- Why do language models need external temporal signals at all?
- Can explicit numerical signals override learned linguistic defaults in fine-tuned models?
- Does next-token prediction alone produce genuine functional language competence?
- Why does tool use decouple factual capacity from model parameter count?
- What makes a problem instance unfamiliar to a language model?
- What are the stages of inference inside language models?
- Does approaching human performance mean learning the same grammatical rules?
- What distinct structural signatures do model repetition and topic volatility create?
- Can you separate grammatical competence from rhetorical commitment in language systems?
- How well-calibrated are language models when making clinical predictions?
- Why do token-level language models fail at utterance-level pragmatic optimization?
- Can training on text corpora teach what communicative acts produce?
- What other structural limits exist at the language-formal boundary?
- What emerges in large language models that makes explicit value modeling necessary?
- What's the difference between language generation and human-to-human communication?
- How much alignment data does a language model actually need to specialize well?
- Why do current speech benchmarks fail to measure reasoning over audio?
- What architectural changes would let language models develop genuine functional competence?
- Why do context-sensitive languages transfer better than regular or context-free languages?
- What distinguishes surface generalizations from true linguistic generalizations?
- How many concurrent moral patients does one language model support?
- What does next-token prediction tell us about compositional linguistic competence?
- What implicit premises do language models skip even with correct surface reasoning?
- What distinguishes character simulation from authentic voice in language model outputs?
- What separates pattern matching from genuine language understanding?
- Why does removing language from its context destroy what makes it work?
- Why can language models detect author style without understanding why it matters?
- What specific information must be exported from the language system?
- Do language models and multimodal models show similar attractor-based interpretability?
- How does predictive accuracy on future tokens differ from correctness on labeled answers?
- What structured values do large language models develop as they scale?
- How does tool-based reasoning expand what language models can do?
- What is the difference between a truthful answer and an honest one?
- Do pretrained language models carry reusable computational scaffolding for length handling?
- What replaces truth-correspondence in probabilistic knowledge representations?
- What geometric structure do language models actually use during inference?
- How do parameter scaling and latent vectors interact in language models?
- What social information is missing from language data?
- Can statistical learning from language alone capture all aspects of cultural competence?
- Which internal states can a language model access and report about itself?
- When and what should a model actually decide to delegate?
- Why might oversight focused only on language model outputs miss the agentic layer?
- Can implicit linguistic information ever be reliably learned from training data?
- How does keyword priming enable language models to spread poisoned information?
- Can models learn to identify what information is missing from questions?
- Why does natural language contain redundancy humans need but models don't?
- What linguistic units do learned concepts correspond to in a language model?
- Can context windows and RAG actually change what language models generate?
- Why does teacher forcing fail to capture long-range dependencies?
- Can language models match competitive crowd forecasters on real future events?
- What is the comprehension-generation asymmetry in language models?
- Why does hierarchical formal language training improve token efficiency more than natural language?
- Why do language models use twice as many words per conversation turn?
- Which linguistic abilities are learnable from human-sized data exposure?
- What makes hallucination errors different from bias and factual mistakes?
- What substrate do supervised models lack that makes them weaker on low-resource languages?
- Can simple diagnostic tests predict language model performance in production complexity?
- Why does homework adherence remain low despite advances in language model capability?