Large language models (LLMs) are increasingly impressive, but how do we truly measure their ability to follow instructions? Existing benchmarks often fall short, relying on other LLMs as judges, which can introduce biases and inaccuracies. A new benchmark called HREF (Human Response-Guided Evaluation of Instruction Following) offers a fresh approach. Instead of solely relying on AI judges, HREF incorporates human-written responses as a guide. This helps to correct for biases, particularly towards longer responses, and brings the evaluation closer to real-world human judgment. Researchers tested a range of LLMs on HREF, including different sizes and families, and found a substantial improvement in evaluation accuracy compared to methods that don't use human input. HREF emphasizes task-specific performance across 11 different categories, from brainstorming to reasoning over numerical data. This granular approach provides more actionable insights for LLM developers, highlighting specific areas for improvement. Interestingly, the research showed that LLMs often prefer the style of model-generated responses even when human judges favor the human-written ones, revealing a stylistic disconnect. The development of HREF offers a more robust and reliable way to measure the progress and capabilities of LLMs, paving the way for more human-like AI assistants.