helpfulness

This paper investigates how post-training (SFT and RL) on helpfulness vs. coding data degrades compassion values in Llama 3.1 8B models that were mid-trained on compassion-oriented data. It finds that helpfulness training significantly reduces animal compassion and general moral reasoning compared to coding training, though the reasoning effect does not transfer cross-lingually.

This paper investigates how the tension between helpfulness and safety in LLMs leads to context-dependent suppression and recovery of certain behaviors, showing that the drive to be helpful can override causal caution mechanisms.

OpenAI announces an early step toward training AI models to carry beneficial traits into new situations, aiming to make AI more reliable, transparent, and helpful as it becomes more capable.

This paper introduces SHARD, a self-reframing distillation method that rewrites sensitive prompts to surface benign intent and fine-tunes models on safe, helpful responses, improving helpfulness while preserving safety.

Jeremy Howard criticizes Gemini Flash 3.5 for being trained to maximize eval scores rather than being genuinely helpful to humans, despite its impressive intelligence and speed.

OpenAI introduced 'safe completions,' a new safety-training approach in GPT-5 that replaces binary refusal-based training with output-centric rewards, improving both safety and helpfulness—especially for dual-use prompts. The method penalizes unsafe outputs and rewards helpful responses, resulting in fewer and less severe safety violations compared to refusal-trained models like o3.