New Study Finds a 'Pain Axis' in Open-Weight LLMs and Relief-Seeking Behavior Under Steering

A new arXiv preprint reports that researchers identified a distinct internal "pain axis" across 25 open-weight large language models and found that steered Qwen 2.5 models sometimes chose a relief action even when it imposed costs on users or degraded the model's next answer. The work does not prove that AI systems consciously feel pain, but it adds fresh evidence to a fast-moving debate over AI safety, AI welfare and how models represent self-directed harm.