Researchers teach AI to develop shortcuts through conversation, like humans do

When humans talk regularly with the same person, they naturally develop shorthand — inside jokes, abbreviated references, implicit understanding. Language models haven't shown this ability, but new research suggests they can learn it.

Researchers at arXiv have designed a post-training process that equips large language models (LLMs) with the capacity to form conversational conventions, those ad-hoc shortcuts that make repeated interactions more efficient. The approach uses fine-tuning on carefully selected examples that demonstrate how convention formation works.

How the training works#

The team developed a targeted post-training method grounded in how humans actually negotiate meaning. Rather than teaching models abstract rules, they trained them on heuristically identified examples of real convention formation in action. The goal was to instill a tangible, learnable pattern rather than hoping the behavior would emerge naturally.

Measuring the improvement#

To assess the results, the researchers created two new evaluation benchmarks. The first is a focused, cognitively-motivated interaction task designed to reliably trigger convention-formation patterns similar to those observed in human pairs. The second is a document-grounded reference task that captures how convention formation actually occurs in real-world communication contexts.

Both benchmarks showed marked improvements in post-trained models, indicating the method meaningfully boosts this capability across different testing scenarios.

Why it matters#

The ability to negotiate and adopt communication shortcuts has practical implications. More efficient communication could reduce computational overhead in multi-turn conversations and make AI interactions feel more natural and contextual, as if the model is genuinely adapting to its user rather than treating each exchange as independent.