reasoningfactneutralRefinement methods in reasoning models can sometimes make answers worse rather than betterKirk Borne02 Aug 2026https://x.com/KirkDBorne/status/2083786496041988401