Tic drift across model versions
Parent: Measuring LLM lexical tics: excess vocabulary, baselines, drift · Published reference · snapshot 2026-09-23
↓ Facts as markdownall context files
Source-anchored research for Tic drift across model versions. Run status: COMPLETED.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
How it works
- The paper's model lineup changed between revisions (v2 lists GPT-5.4, Claude Opus 4.7, Grok 4.2, DeepSeek V3.2; latest lists GPT-5.5, Claude Opus 4.8, Grok 4.3, DeepSeek V4 Pro), so tic measurements are tied to a specific model version and get re-run when versions ship. (medium confidence) (verified as of 2026-07-05) [source]
- Juzek and Ward (COLING 2025) found no evidence that overrepresented words like 'delve' stem from model architecture, algorithm choice or training data, and tested RLHF as a candidate cause, so the mechanism behind version-to-version drift is unresolved. (low confidence) [source]
Facts and statements
- The Verbal Tic Index paper scores eight frontier models over 160,000 responses (10,000 prompts, English and Chinese); VTI ranged from 0.590 (Gemini 3.1 Pro, highest) to 0.295 (DeepSeek V4 Pro, lowest) in the latest revision. (low confidence) (verified as of 2026-07-05) [source]
- Rudnicka and Juzek find linguistic patterns shifted measurably between 2024 and 2026 model cohorts, and models of the same generation differ sharply (contraction rate from over 1,200 to over 30,000 per million words). (low confidence) (verified as of 2026-09-23) [source]
Children
- No children recorded.