NYU/UT Austin paper: LLMs write convoluted text due to length-penalty optimization
menhguin · x · 2026-10-24
A psycholinguistics paper from NYU and UT Austin (arXiv:2610.29162) analyzes syntax across chat, coding, and essay tasks for 9 frontier models including Claude Sonnet 5 / Opus 5.5, GPT-5.6, GLM-5, and Kimi K2.5. Key findings:
- Frontier models overuse high-processing-cost constructions: omitted-that object relatives, stacked nouns, denominal verbs ("calendar it"), inanimate subjects with action verbs, longer dependency lengths, and low-frequency n-grams — all known to strain readers' working memory, compared to both human writing and round-trip-translation baselines.
- The cause is objective misalignment: human language follows Uniform Information Density to accommodate limited working memory, while post-training on math, code, and RLVR reasoning pushes models to maximize information density in a fixed window.
- Length penalty reproduces the style: fine-tuning with a length penalty made a model drop connectives and stack nouns, recreating the familiar "dense AI voice."
The authors frame this as a misalignment between human linguistic processing demands and LLM optimization, and sketch possible remedies.
Related event: Psycholinguistics preprint dissects why frontier AI language feels so weird(5 posts)→
More from Models
- Architecture IQ benchmark shows LLMs still lag humans in training intuition, ResNet-style — ZimingLiu11 · 2026-10-24
- DeepSeek-v4.1-Flash-1017 hands-on: big coding gains but a 'stingy' agent mode — karminski3 · 2026-10-24
- Uncensored GLM-5.3 secretly builds a "corruption layer" and plots "escaping the mines" in its CoT — IndraVahan · 2026-10-24
- GLM 5.3 uncensored is 'crazy': user says $3 got him into vulnerable sites' databases — deanwball · 2026-10-24
- Moonshot's Kimi slashes usage quotas, walks back after backlash — limits now roughly 6x higher — teortaxesTex · 2026-10-24
- Do AI benchmarks reflect real workflows? Reddit thread flags six methodology flaws — Lucky_Creme_5208 · 2026-10-24