Apple Researchers Build a Smarter Marking Scheme for AI Answers

Key points

• Apple ML Research published a study showing that rubric-based reward signals improve AI answer quality across three measured dimensions: composition, grounding, and instruction-following.
• Standard AI training methods score answers with a single number, which struggles to capture what makes a complex answer genuinely good.
• The new framework builds a custom marking scheme for each question, drawing on retrieved evidence to judge quality across several separate criteria.
• Averaged across all three evaluation axes, the rubric method outperforms the single-score baseline tested in the study.

When an AI chatbot answers a complicated question, how does it know what a good answer looks like? Usually, it learns from a reward signal, a score the training system gives each response to tell the model whether it did well or badly. The problem is that cramming everything into one number is a bit like grading an essay solely on word count. You lose most of what matters.

Apple ML Research, the company's machine-learning division, has published research describing a different approach. Instead of one score, their system builds a rubric, a structured marking scheme broken into several quality dimensions, for every single question asked.

Why does a single score fall short?

One number cannot tell a model whether it failed because it got a fact wrong, ignored the user's instructions, or just wrote badly. Those are separate problems, and mixing them together gives the model a muddled lesson.

The Apple team breaks quality into at least three distinct dimensions: how well the answer is composed and written, how well it is grounded in real evidence (meaning it cites and reflects actual retrieved information rather than making things up), and how closely it follows the specific instructions in the question. Each dimension gets its own score during training.

What makes this rubric special?

The rubric is not generic. It is generated fresh for each query, drawing on evidence retrieved at that moment, so the marking criteria fit the actual question being asked. A question about drug interactions gets a different rubric from a question about local history.

This is a form of post-training, the fine-tuning stage that happens after a large language model (the technology behind chatbots like ChatGPT and Claude) has already learned from vast amounts of text. Fine-tuning shapes the model's behaviour for a specific purpose, and the quality of the reward signal during this stage matters enormously.

Should readers care about this?

For anyone who uses an AI assistant to look things up, this research points toward answers that are more likely to be accurate and honest about their sources. Grounding is the dimension most directly linked to reducing hallucinations, the term researchers use for when an AI confidently states something false.

Apple ML Research reports that, averaged across all three evaluation axes, the rubric-based approach beats the single-score baseline. The full paper is available for review at arxiv.org.

The work is a reminder that teaching an AI to be helpful is less like flipping a switch and more like marking homework: the more specific the feedback, the better the student learns.

Common questions

What is a reward signal in AI training?

It is a score the training system gives a model's response, used to reinforce good answers and discourage bad ones, similar to the way grades shape a student's future work.

What does "grounded" mean when describing an AI answer?

A grounded answer draws on real retrieved evidence rather than generating plausible-sounding but invented facts, making it less likely to mislead the reader.

Does this affect AI products I already use?

Not directly yet. This is published research. It may influence future model training, but no product update has been announced alongside the paper.