Why don't LLMs shorten messages like humans do?

Humans spontaneously develop increasingly efficient language during interactions. A patient might start with "the medicine for my back pain in a small blue medicine bottle" and within a week say just "my back meds." This is lexical convention formation — a fundamental property of human communication documented extensively in reference game studies.

The ICCA framework evaluates whether multimodal LLMs exhibit this behavior. The results reveal an asymmetry: as listeners, models like GPT-4 show adaptation trends close to humans, improving accuracy as interactions progress. But as speakers, all models fail to spontaneously improve communication efficiency. Only with fairly heavy-handed instruction — explicitly telling the model to reduce message length and maintain lexical consistency — do GPT-4, Gemini, and Claude show partial adaptation.

Four prompting variants reveal the gradient:

• Standard: no mention of efficiency → no adaptation
• Gricean instruction: Grice's quantity maxim ("don't be more informative than necessary") → minimal effect
• Explicit instruction: "gradually condense your messages" → some length reduction
• Explicit + consistency: "extract salient tokens from previous messages" → closest to human behavior

The Word Novelty Rate (WNR) metric captures this precisely — counting word insertions and substitutions while ignoring deletions, since deletions reflect natural convention formation while additions indicate cognitive-load-increasing changes.

This finding is a concrete instantiation of a broader pattern. Since Why do language models fail at communicative optimization?, communicative efficiency through convention formation is precisely the optimization principle that training on form alone cannot produce. Convention formation requires modeling the listener's cognitive state and adjusting accordingly — a functional competence that next-token prediction does not select for.

A training-time solution has now been demonstrated. Since Can we teach LLMs to form linguistic conventions in context?, the convention formation gap is addressable through targeted post-training rather than architectural redesign. The approach — heuristically extracting coreference chains from TV scripts, constructing DPO preference pairs, and adding a [remention] planning token — produces general in-context convention formation behavior. The model spontaneously shortens references as interaction progresses, precisely the capability the ICCA framework found missing.