Large language models often produce factually correct responses, but they do not always maintain the perspective a user requested. Instead, they may blend multiple viewpoints, forcing the user to filter through unwanted information. The paper, posted on arXiv, identifies this as a perspective control problem.
The authors introduce a benchmark designed specifically to evaluate perspective alignment. Unlike typical benchmarks that focus on factual accuracy, this one tests whether a model's output stays within the viewpoint the user asked for. It provides a way to measure how well models follow such instructions.
The work highlights a gap in current evaluation methods: a model can be factually correct yet still fail to respect a requested perspective. By treating perspective alignment as a distinct capability, the benchmark offers a new direction for assessing and improving LLM behavior.