SYNTHESIS NOTE
TopicsNatural Language Inferencethis note

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.

Synthesis note · 2026-05-02 · sourced from Natural Language Inference
How do language models learn to think like humans? Why do LLMs fail at understanding what remains unsaid?

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? How do training priors constrain what context information can override? What factors beyond surface content determine how readers extract meaning differently? What role does compression play in language model capability and generalization? How do language models establish social grounding in human dialogue? What properties determine whether reward signals teach genuine reasoning? How should retrieval systems optimize for multi-step reasoning during inference? Do language models understand semantics or rely on pattern matching? Why do language models struggle with implicit discourse relations? Do language models learn genuine linguistic structure or just surface patterns? What structural advantages do diffusion language models offer over autoregressive methods? Why do semantic similarity and task relevance diverge in vector embeddings? Why do benchmark improvements fail to reflect actual reasoning quality? Why do language models reinforce false assumptions instead of correcting them? Is embodied interaction necessary for language meaning and genuine agency? Why does supervised fine-tuning improve accuracy while degrading reasoning quality? Do language models develop causal world models or rely on statistical patterns? Do language models perform faithful symbolic reasoning independent of semantic grounding? How do neural networks separate factual knowledge from reasoning abilities? Can model confidence signals reliably improve reasoning quality and calibration? How does memorization interact with learning and generalization? Do language model representations contain causally steerable task-specific features? How do evaluation biases undermine LLM quality assessment systems? How do training data properties shape reasoning capability development? When does optimizing for quality undermine the value of diversity? What limits mechanistic interpretability's ability to characterize models? What articulatory information do speech signals carry that text cannot?

Related concepts in this collection 3

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
13 direct connections · 117 in 2-hop network ·dense cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

high-frequency phrasing wins — LLMs systematically prefer textually frequent paraphrases over rare ones with the same meaning