Helpfulness training degrades compassion values in LLMs by 35%

A new arXiv paper titled 'Helpfulness Hurts' investigates how standard post-training pipelines—supervised fine-tuning (SFT) and reinforcement learning (GRPO)—can degrade values instilled during pre-training. The authors, Jasmine Brazilek and Juliana Seawell, fine-tuned a Llama 3.1 8B model mid-trained on compassion-oriented synthetic data across two domains: helpfulness (using Dolly-15k and RLHFlow) and coding (using Magicoder-110K and Magicoder RL). Evaluation on the Animal Harm Benchmark (AHB 2.2) showed that helpfulness SFT achieved only 35.7% compassion accuracy vs. 65.2% for coding SFT, a stark 29.5-point gap. GRPO results were consistent: 18.7% for helpfulness vs. 32.0% for coding. On English MORU (Moral Reasoning Under Uncertainty), helpfulness training degraded general moral reasoning by 25.5 percentage points (46.4% vs. 71.9%).

Strikingly, the compassion degradation transferred cross-lingually—coding's advantage over base model was 4.5x larger on non-English items than English—while the reasoning gap disappeared entirely on multilingual MORU (SFT: 52.3% vs. 51.2%). This divergence suggests that mid-trained values (like compassion) are encoded more robustly and language-agnostically than reasoning improvements from domain-specific post-training. The authors conclude that labs building on value-laden mid-trained models should consider coding-domain post-training as a safer alternative to helpfulness fine-tuning, as it preserves compassion without harming general reasoning. This has direct implications for AI safety and alignment, especially in deploying models in diverse linguistic and ethical contexts.