'Wait' markers are inherited from training data, not invented by RL, HKUST study finds
huawei-noah · hf · 2026-10-23
Researchers from HKUST, Peking University, and Huawei Noah's Ark Lab traced the provenance of 'wait' in reasoning models via label-free indices over training histories (frequency, position, form, usage) plus injected-error tests. Zero-RL almost never produces discourse markers like 'wait'—they originate mainly from mid-training/SFT data rather than RL—and large RL accuracy gains do not require the word. The study warns that common narratives conflate what RL discovers with what models inherit from training data, and that automated evaluation metrics can misrepresent reasoning trends.
More from Models
- Gemini 4 Argon spotted testing apps autonomously, and 'topping humor benchmarks' — andrew_n_carr · 2026-10-23
- Gemini 4 reportedly far less sensitive in refusals than other models — haider1 · 2026-10-23
- SemiAnalysis: Anthropic subscriptions offer 5x+ more value than OpenAI as $200 plan value halves — RexDouglass · 2026-10-23
- Google Cloud's Physical-Engineering Last Exam: all frontier models score under 50% — Afinetheorem · 2026-10-23
- LLM copyright guardrail absurdity: refuses to scrape a book but happily zips it for you — burkov · 2026-10-23
- Study: Claude Opus 4.8 beats human experts on every measure as peer reviewer of 124 Harvard papers — emollick · 2026-10-23