10:51
2026-10-07
github.com
large-language-models
We tried to reproduce LLM judge position bias
Viridian's Technical Note 001, a controlled replication study dated 5 October 2026, found that reversing the presentation order of two frozen candidate responses preserved the underlying preferred res…