Why LLMs abuse the em dash: corpus bias, autoregression and RLHF reward hacking
SparkyAI0815 · reddit · 2026-11-08
A Reddit explainer attributes LLM em-dash abuse to three stacked mechanisms: training corpora over-weight copy-edited journalism where editors treat em dashes as religion; autoregressive generation can't plan sentence endings, making the dash universal syntactic duct tape; and RLHF raters subconsciously reward punchy contrast, so the reward model cranks dash frequency up. Same feedback loop as 'delve into rich tapestries'.
More from Models
- Dev benchmarks tiny LLMs on agent tasks: 9B passes 2/5, stepping up to 12B — draginol · 2026-11-08
- Meta eval lead's BenchBench chart shows self-reported eval counts across labs — ziqiao_ma · 2026-11-08
- Local 4B model fails Extended agent benchmark: 2/5 tasks at 61.5 tok/s — draginol · 2026-11-08
- Mystery arena model "Kairo" paints Bob Ross in 2,878 strokes, runs on OpenAI's tokenizer — flowersslop · 2026-11-08
- Dev benchmarks Gemma 4B on local agent tasks: scores 2/5, fails Extended mode threshold — draginol · 2026-11-08
- Cerebras runs Kimi K3.1 at 2,000+ tokens per second on wafer-scale engine — rohanpaul_ai · 2026-11-08