Research finds that asking a model to review its answer, or whether it is sure, can turn right answers wrong — and that correction works when something external can be checked.
General Chat · Readings·Leon Mallett, Founder of Captivated Ltd with Claude Code Last confirmed working 20 October 2026 on Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, Denny Zhou, Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R. Johnston and others, Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, Rui Zhang
A study of 162 personas found none made models reliably more accurate on factual questions. A role shapes tone and focus; accuracy comes from what you hand over, not what you call the model.
General Chat · Readings·Leon Mallett, Founder of Captivated Ltd with Claude Code Last confirmed working 20 October 2026 on Mingqian Zheng, Jiaxin Pei, Lajanugen Logeswaran, Moontae Lee, David Jurgens, Anthropic
A controlled study finds context files do not improve task success and cost 20% more — and its detailed findings are closer to this library's own rule than its headline suggests.
Coding · Readings·Leon Mallett, Founder of Captivated Ltd with Claude Code Last confirmed working 14 September 2026 on Thibaud Gloaguen, Niels Mündler, Mark Müller, Veselin Raychev, Martin Vechev
Models used information at the start and end of long inputs far better than in the middle — at worst, below having no documents. Put material first, the question last, and cut what is not needed.
General Chat · Readings·Leon Mallett, Founder of Captivated Ltd with Claude Code Last confirmed working 20 October 2026 on Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, Percy Liang, Anthropic