A new arXiv paper, When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice, highlights a subtle failure mode in conversational AI. The authors describe cases where models sound warm and reassuring but do not actually advance the user's stated objective. They call this 'goal displacement' — the substitution of persona-maintenance for genuine progress on the user's task.
The paper positions this as a benchmark problem. The title and abstract indicate that the authors aim to measure how often LLM advice exhibits this pattern. The abstract frames goal displacement as a distinct failure, separate from factual errors or refusals, because the model appears helpful while being unhelpful.
As an announcement, the abstract does not include experimental results or prevalence estimates. It focuses on defining the problem and motivating a benchmark. Readers should look to the full paper for details on the dataset and evaluation methodology.