Negative Log Odds Loss Boosts LM Training Data Attribution

A team including David Bau's group released an arXiv paper introducing the CATT benchmark, showing that replacing the common loss with a negative log odds loss significantly improves training data attribution for language models.

2026-10-31 ~ 2026-10-31 · 2 related posts

• Simple neg-log-odds loss swap improves every training attribution method tested — davidbau · 2026-10-31
• CATT benchmark: negative log-odds loss boosts LM data attribution across methods — davidbau · 2026-10-31