Do language models really understand meaning or just surface frequency?
Explores whether LLMs comprehend semantic meaning independently of textual frequency, or whether high-frequency paraphrases systematically outperform rare ones even when meaning is identical across math, translation, and reasoning tasks.
Adam's Law (TFL) generalizes a previously local finding into a global property of LLM computation. The earlier NLI work showed predicates in entailment hypotheses skew higher-frequency than premises, and that fine-tuning amplifies rather than dilutes this bias — see Does fine-tuning on NLI teach inference or amplify shortcuts?. Adam's Law extends this across four task families: math reasoning, machine translation across hundreds of language pairs, commonsense reasoning, and agentic tool calling. The constant: when meaning is held fixed and only surface form varies, the higher-frequency paraphrase outperforms the lower-frequency one.
The mechanism is straightforward but uncomfortable. Higher-frequency text occurred more often during pre-training, so it sits in a denser, better-modeled region of the distribution. The model's "comprehension" is therefore not meaning-recognition first and surface-decoding second — it is statistical-mass recognition first, with meaning emerging downstream of that recognition. This converges with Can models pass tests while missing the actual grammar?: correct outputs do not certify that meaning is what the model is tracking.
The pattern matters because paraphrase invariance is a load-bearing assumption almost everywhere LLMs are deployed. We assume the same prompt, said two ways, will yield the same answer. Adam's Law says no: it will yield the frequency-weighted answer, and the surface form is a covariate of accuracy, not a transparent vehicle for the request. This also shadows the output side. Do different AI models actually produce diverse outputs? documents convergence in what models say; Adam's Law documents the same convergence in how models comprehend what is said to them. Both endpoints of the prompt-response loop pull toward the corpus mean. Frequency is not noise around meaning. Frequency is a substantial fraction of what comprehension means inside a transformer.
Inquiring lines that read this note 82
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
What makes AI persuasion effective and how can we counter it? Why can't humans reliably detect AI-generated text despite measurable linguistic signatures?- Can AI detect sense-of-nonsense the way human readers do?
- Can adversarial paraphrasing defeat feature-based detection of LLM text?
- Why does training data saliency distort how models judge meaning?
- How does training distribution shape what language models understand best?
- What makes ambiguity recognition fundamentally important for poetry analysis?
- What semantic classifier design avoids lexical variation without genuine conceptual distinctness?
- Can adding more words to a passage actually interfere with meaning?
- Why do different readers extract different meanings from identical text?
- Can readers detect meaning through resonance patterns alone without knowing authorial intent?
- Why does statistical compression destroy literary connotation and meaning?
- How does modeling capability relate to lossless compression in language models?
- Does functional grounding through discourse patterns count as genuine semantic meaning?
- Can language models ground clarifications without vision and kinesthetic modalities?
- How does semantic grounding differ between human minds and language models?
- How does implicit meaning processing limit LLM pragmatic reasoning?
- Can LLMs infer implicit meaning without surface linguistic markers?
- Can LLMs identify implicit metaphoric mappings that require pragmatic inference?
- How does syntactic encoding relate to semantic feature representation?
- How does semantic ambiguity differ from structural ambiguity in language?
- Is interpretive multiplicity a bug in language or a feature?
- Why do language models fail when semantic content is stripped away?
- How do rare linguistic registers differ from conceptually complex examples?
- Can LLMs improve at metaphor if they handle decoupled semantics better?
- Can frame semantics explain why context matters more than word similarity?
- Do LLMs struggle more with semantic accuracy than syntactic correctness across domains?
- Why do LLMs fail at semantic generalization despite grammatical accuracy?
- Do metaphors work by decoupling meaning from linguistic associations?
- Why do LLMs choose surface-order quantifier scope over contextually correct readings?
- How much semantic meaning survives when LLMs paraphrase poetry and literary text?
- Why do NLP models fail at recognizing multiple valid interpretations?
- How does bidirectional entailment distinguish semantic equivalence from token similarity?
- How do LLMs compress literary language without losing essential nuance?
- Why does cross-text analogical reasoning fail when semantics decouple from symbols?
- What percentage of natural language relies on plausible deniability through ambiguous phrasing?
- What other semantic relations benefit from explicit surface markers in text?
- Can presupposition projection strength vary by context in embeddings?
- How should meaning spaces be systematically modeled across different applications?
- Does generalization frequency explain why models favor upward semantic movement?
- What distinguishes surface cues from structural meaning in language understanding?
- Can encoder models match human conceptual structure better than larger language models?
- Do LLMs learn linguistic generalizations or just surface-level frequency patterns?
- What distinguishes real understanding from superficial pattern matching?
- What is the comprehension-generation asymmetry in language models?
- Why do embeddings measure semantic association instead of task relevance?
- What makes vector embeddings fail on single-hop semantic relevance queries?
- Why do semantic similarity and task relevance diverge in vector search results?
- Can vector embeddings measure task relevance instead of semantic similarity?
- How well does semantic similarity preserve survey response nuance?
- What paraphrase and conceptual matching tasks favor dense over exact-match retrieval?
- Why does homework adherence remain low despite advances in language model capability?
- Can correct model outputs prove that semantic meaning rather than surface patterns drove the response?
- Can benchmark performance distinguish surface from structural linguistic knowledge?
- Why do readability and style metrics plateau while reasoning improves with scale?
- Why do speech benchmarks still measure transcription instead of comprehension?
- Do LLMs compute scalar implicature differently across conversational contexts?
- Is paraphrase invariance a reliable assumption when deploying language models in production?
- Can language models acquire meaning from distributional patterns alone without joint attention?
- Can understanding language happen entirely within a language system alone?
- Why does joint attention matter for acquiring linguistic meaning?
- How do static embeddings and contextualized representations divide semantic labor?
- Does language convey meaning purely through relational structure without external grounding?
- Where does the meaning actually originate in reader-detected resonance across language?
- Can language models develop world models that ground meaning in causal reality?
- Can LLM semantic representations exist without causally influencing their generation output?
- How does the distance between natural language and formal notation affect translation accuracy?
- Can autoformalisation from natural language preserve semantic accuracy?
- Can language models translate theorems faithfully without semantic loss?
Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Does fine-tuning on NLI teach inference or amplify shortcuts?
When LLMs are fine-tuned on natural language inference datasets, do they learn genuine reasoning abilities or become better at exploiting statistical patterns in the training data? Understanding this distinction matters for assessing model capabilities.
local finding that Adam's Law generalizes
-
Can models pass tests while missing the actual grammar?
Do language models succeed on grammatical benchmarks by learning surface patterns rather than structural rules? This matters because correct outputs may hide reliance on shallow heuristics that fail on novel structures.
mechanism: surface, not semantics
-
Do different AI models actually produce diverse outputs?
Explores whether using multiple different language models together creates genuine diversity or whether shared training and alignment cause them to converge on similar answers despite independence.
output-side counterpart of the same dynamic
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Adam's Law: Textual Frequency Law on Large Language Models
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- Word Meanings in Transformer Language Models
- Simple Linguistic Inferences of Large Language Models (LLMs): Blind Spots and Blinds
- Bigger is not always better: The importance of human-scale language modeling for psycholinguistics
- Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs
- Language models show human-like content effects on reasoning tasks
- Large Linguistic Models: Investigating LLMs' metalinguistic abilities
Original note title
high-frequency phrasing wins — LLMs systematically prefer textually frequent paraphrases over rare ones with the same meaning