Post-training causes LLMs to produce novel but impractical language
TuhinChakr · x · 2026-08-20
Research indicates that larger, 'better' models tend to generate novel but non-pragmatic expressions. The author suggests that current post-training and benchmark maximization induce novel yet nonsensical language in LLMs. This aligns with user complaints about models using verbose metaphors and complex phrasing that hinder communication rather than helping it.
More from Models
- Agnost AI Launches Log-Fine-tuned Model: +22.9% Success, -94.5% Cost — ycombinator · 2026-08-21
- Efficiency Chart: Qwen-3.8-27B Hits Pareto Frontier, Outperforms Claude — MikePFrank · 2026-08-21
- Scoop: OpenAI to Launch Astra in Two Weeks; Rivals Fable 5.1 Ready — haider1 · 2026-08-21
- DataCamp CEO on Choosing Open vs Frontier Models in Production for 19M Learners — kimmonismus · 2026-08-21
- Claude accurately predicted Qwen 3.8 27B performance benchmarks — OneMoreName1 · 2026-08-21
- GLM 5.3 Scores 47.1% on SlopCodeBench, Ties with Fable 5 — corruptbytes · 2026-08-21