Redwood Research finds GPT-6-Astra gains large non-reasoning boost from filler tokens
Original title:Astra is much better at reasoning with filler tokens than previous models
AI summary
Redwood Research measured GPT-6-Astra's performance after adding meaningless filler tokens to prompts, finding significant gains on tasks requiring extensive serial cognition: 4-hop natural fact reasoning rose from about 10-20% to about 50%, and old AIME problems from about 60-70% to about 90%.
Key Takeaways
- Redwood Research measured GPT-6-Astra's performance after adding meaningless filler tokens to…
- finding significant gains on tasks requiring extensive serial cognition: 4-hop natural fact…
- and old AIME problems from about 60-70% to about 90%.
Quick Facts
- Published
- Updated
- Source
- Redwood Research:Blog(RSS)
- Category
- Papers & Research
- AI score
- 73
原文给出填充 token 显著提升 GPT-6-Astra 无推理表现的实测数据,指向思维链监控失效这一安全影响。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Redwood Research finds GPT-6-Astra gains large non-reasoning boost from filler tokens. September 23, 2026. https://new.nat6.net/en/news/o0bzx5jxh2r108wvqlcyeevd4/