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?
"Do Large Language Models Resemble Humans in Language Use?" (Yiu et al. 2023) evaluates LLMs on a wide range of human linguistic regularities — not just grammaticality but psycholinguistic phenomena. The results show a consistent pattern of success and failure that tracks a specific distinction.
LLMs succeed on:
- Sound symbolism (maluma/takete roundness-spikiness associations)
- Sound-gender associations
- Structural priming (using the same syntactic structure as prime)
- Semantic priming (accessing recently primed word meanings)
- Dialect sensitivity (accessing dialect-appropriate vocabulary based on stated interlocutor identity)
These regularities are learnable from distributional patterns in text — they appear consistently across large corpora and can be acquired through form-to-form prediction.
LLMs fail on:
- Word length economy (choosing shorter forms when context is predictive — the Zipfian efficiency principle)
- Syntactic ambiguity resolution (selecting the contextually appropriate reading of ambiguous syntax)
- Semantic illusions (detecting incongruent words in otherwise coherent sentences)
- Drawing discourse inferences (bridging two pieces of information)
These regularities require something beyond distributional pattern matching. They involve principles of why language works for communication — efficiency under communicative pressure, contextual interpretation that goes beyond local statistics, integration across discourse.
The discriminating principle: statistical regularities that appear as consistent patterns in training data transfer. Regularities that emerge from communicative optimization — the pragmatic logic of why language has the forms it does — do not transfer, because they are not present in surface form as trainable signals.
Inquiring lines that read this note 16
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.
Can language models build genuine grounding through interaction?- How do humans learn language through communication differently than LLM text prediction?
- Why do LLMs lack the communicative scaffold that humans learn?
- What is the difference between learning discourse patterns and learning abstract language?
- Do LLMs have functional linguistic competence or only formal language ability?
- Why do LLMs fail at implicit elements in literary and poetic text?
- Why does LLM knowledge fail to influence their actual outputs?
- Why do LLMs fail at iterative numerical computation in latent space?
- What communicative optimization principles do language models fail to acquire?
- Why do LLMs understand efficient language but fail to produce it?
- Why do language models fail at iterative numerical optimization despite scale?
- What makes human language fundamentally different from what language models produce?
Related concepts in this collection 5
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
-
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.
this paper provides the specific taxonomy of which generalizations do and don't transfer
-
Why does ChatGPT fail at implicit discourse relations?
ChatGPT excels when discourse connectives are present but drops to 24% accuracy without them. What does this gap reveal about how LLMs actually process meaning and logical relationships?
same pattern: surface-explicit success, contextual-implicit failure
-
Why do LLMs handle causal reasoning better than temporal reasoning?
Exploring whether language models perform asymmetrically on different discourse relations and what training data patterns might explain the gap between causal and temporal reasoning abilities.
parallel: what's in the training distribution determines what's learned
-
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.
ambiguity management is one of the failed communicative optimization principles
-
Why don't LLMs shorten messages like humans do?
Humans naturally develop shorter, efficient language during conversations. Do multimodal LLMs exhibit this same spontaneous adaptation, or do they lack this communicative behavior?
ICCA finding: convention formation (becoming more efficient through interaction) is precisely a communicative optimization principle that LLMs fail to acquire; they understand efficient language but don't produce it
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Computational structuralism: Toward a formal theory of meaning in the age of digital intelligence
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- Word Meanings in Transformer Language Models
- Large Linguistic Models: Investigating LLMs' metalinguistic abilities
- Large Models of What? Mistaking Engineering Achievements for Human Linguistic Agency
- The Missing Layer of AGI: From Pattern Alchemy to Coordination Physics
- Probing Structured Semantics Understanding and Generation of Language Models via Question Answering
- Large Language Model Programs
Original note title
llms replicate local statistical regularities in language but fail to acquire communicative optimization principles