The Rise of Verbal Tics in Large Language Models: A Systematic Analysis Across Frontier Models

The buried finding here is that these tics aren't bugs in the traditional sense — they're artifacts of reward optimization itself, meaning the same training techniques that make models more helpful are structurally producing the formulaic speech. Fixing them isn't a simple post-processing problem; it may require rethinking what reward signals are actually measuring.

This connects directly to the DiscoTrace paper from April 16, which found that LLMs systematically lack rhetorical variety and favor breadth over selectivity when answering questions. That study approached the problem from discourse structure; this one approaches it from training dynamics — but both are documenting the same underlying homogenization. The LLM judge reliability work from April 16 adds another layer: if evaluators used in RLHF pipelines show logical inconsistencies in up to two-thirds of documents, the reward signal shaping these verbal habits may itself be noisier than assumed.

Watch whether any of the eight named models — particularly Claude Opus 4.7 or GPT-5.4 — release alignment or fine-tuning updates in the next two quarters that explicitly cite tic reduction as an objective; that would confirm labs are treating this as a measurable training problem rather than a cosmetic one.