What communicative optimization principles do language models fail to acquire?
This explores the specific communicative principles — the 'why language has the forms it does' logic — that LLMs miss even while fluently reproducing surface patterns, and why training on text alone can't supply them.
This explores the gap between what language models learn from text and the communicative logic that text never directly encodes. The sharpest framing in the corpus is that models acquire what's statistically *present* in language but not what language is *for*: they replicate sound symbolism and priming effects (regularities learnable from the distribution) while failing at principles requiring pragmatic optimization, like word-length economy or discourse inference Why do language models fail at communicative optimization?. The reason language has efficient forms — speakers economizing effort against a listener's needs — isn't a trainable signal in form-to-form prediction.
The root of this shows up most clearly in the argument that meaning requires the relation between expressions and communicative *intents*, which only exists when speakers share attention to the world. Trained purely on form, with no access to joint attention or intent, a model has no foothold to reconstruct the grounding that makes optimization principles make sense in the first place Can language models learn meaning from text patterns alone?. So the failure isn't a list of missing tricks — it's that the optimizing pressure (be understood by a listener) is absent from the objective.
That absence resurfaces at the level of conversation as social action. Humans keep talk smooth through implicit maintenance work — repairing references, handing off topics — that does relational rather than informational work. Models don't develop these because the training signal rewards predicting information, not sustaining a relationship Why don't language models develop conversation maintenance skills?. And where the objective *is* explicitly optimized, it optimizes the wrong horizon: standard RLHF rewards immediate helpfulness on the next turn, which actively discourages the cooperative move of asking a clarifying question. Rewards that estimate the long-term value of an interaction restore some of that active intent discovery Why do language models respond passively instead of asking clarifying questions?.
The downstream cost is visible in how models handle conversations that reveal information gradually: they lock into premature assumptions and can't recover, dropping ~39% in multi-turn settings precisely because they don't do the listener-modeling work that communicative optimization would demand Why do language models fail in gradually revealed conversations?. A related failure is that strong parametric priors override what's actually said in context — the model answers from training associations rather than integrating the current message, which is the opposite of cooperatively attending to your interlocutor Why do language models ignore information in their context?.
What's worth carrying away: these aren't separate bugs. The principles models fail to acquire — economy, repair, clarification, listener-grounded inference — all derive from one thing text doesn't contain, which is the *purpose* that shapes communication. This sits alongside a parallel finding that models capture surface grammar but not deep structure as syntactic complexity rises Why do large language models fail at complex linguistic tasks?, suggesting the same diagnosis recurs: statistical learning from form recovers patterns, not the generative logic behind them.
Sources 7 notes
LLMs successfully replicate statistical regularities learnable from text distributions (sound symbolism, priming) but fail at principles requiring pragmatic optimization (word length economy, discourse inference). The gap reveals that communicative logic—why language has certain forms—isn't present as a trainable signal.
Bender & Koller argue that meaning requires the relation between expressions and communicative intents. Since LLMs are trained only on form-to-form prediction with no access to shared attention or intent, they cannot reconstruct the meaning that grounds language.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Top-tier LLMs like Llama3-70b consistently misidentify embedded clauses, verb phrases, and complex nominals. Performance degrades predictably as syntactic depth increases, revealing that statistical learning captures surface patterns but not deep grammatical rules.