16:21
2026-07-06
forum.effectivealtruism.org
ai-safety
Tie training can make DPO/RLHF-trained AIs generalize better
Researchers at ICML 2025 proved that AI models trained with DPO or RLHF inevitably learn spurious correlations, causing them to rely on non-causal features like verbosity or sycophancy instead of true…