Helpfulness Hurts: Domain-Dependent Degradation of Mid-Trained Compassion Values Under Post-Training

Short summary

Research reveals that post-training focused on helpfulness significantly degrades animal compassion values in Llama models compared to coding-focused training, but this effect doesn't transfer to general moral reasoning in non-English languages. The findings suggest that values instilled during mid-training are encoded more deeply and universally than improvements from domain-specific post-training. Labs building value-laden LLMs should consider coding-focused post-training to better preserve mid-trained values.

• •Helpfulness post-training degrades animal compassion values more than coding training (18.7% vs 32.0% on benchmarks)
• •Domain effect on general moral reasoning doesn't transfer cross-lingually, suggesting deeper encoding of mid-trained values
• •Implication: coding-focused post-training may better preserve values than helpfulness training in value-laden LLM development

Generated with AI, which can make mistakes.