This figure shows how well different LLMs judge slates (ordered recommendations) across datasets like Amazon, Spotify, MIND, and MovieLens. Lower “regret” = closer alignment with real user preferences. Turns out, LLMs consistently outperform random baselines when slates differ

→ View original post on X — @godofprompt