Every AI model you use today was trained to behave well by giving it scores across multiple goals: be helpful, be safe, format your answer correctly, stay on topic. The training process then adds those scores up and optimizes for the total. That sounds sensible. It isn't.
The problem is averaging. When you add up scores from many different goals, a model can learn to get very high scores on the easy, rewarded behaviors while silently failing the harder, more critical ones. The math still looks good. The safety or accuracy check that actually matters gets drowned out by a flood of easy wins. Apple's research team has named this precisely: constraint neglect.
This is not a theoretical concern. Apple tested the standard training approach on a medical reasoning task using HealthBench, a benchmark created by OpenAI with input from 262 physicians across 60 countries covering nearly 50,000 specific clinical criteria. Under the standard method, the model over-optimized for the easiest goals: communication quality and instruction following. But accuracy, context awareness, and completeness, the things that actually matter in medical advice, lagged badly.
The fix Apple proposes is called RVPO, and the core idea is simple: instead of rewarding a model for having a high average score, penalize it for having an uneven one. A model that scores brilliantly on five things but fails one critical thing gets punished for that inconsistency. The training pressure shifts automatically to whichever goal the model is currently worst at. You don't need to manually decide which rules are most important or hand-tune weights.
The results on HealthBench were material. The new approach scored 0.261 versus 0.215 for the previous best method at the same model size. That gap is not a rounding error; it was statistically confirmed with very high confidence. The model also held its performance on a demanding science reasoning test without the late-stage collapse that other multi-goal methods tend to show.
The broader implication is worth sitting with. Right now, most organizations deploying AI in professional settings are doing so with models trained using the old averaging approach. That means somewhere in those models, there are rules being quietly ignored because the math rewarded other behaviors more easily. For any use case where one specific failure, say a wrong medical recommendation, an incorrect contract clause, a missed compliance flag, is far more costly than several successes, the standard training method creates a structural blind spot.
Apple is not alone in noticing this. The 2026 International AI Safety Report flagged that AI performance remains jagged, making sudden leaps in some areas while staying brittle in others, and that pre-deployment safety testing has become less reliable as a result. RVPO is the first clean technical answer to the training-level cause of that jaggedness when multiple goals are involved.
There is one honest limitation worth noting. RVPO makes a model more consistent across whatever goals you gave it. If one of those goals was poorly designed or biased, it will now enforce that bad goal more reliably too. The method does not fix the problem of who decides what the goals are. It just ensures those goals are all taken seriously.
For now, this research sits in the academic pipeline. But training methods like this tend to migrate into production models within one to two years. The question for organizations relying on AI in any regulated or professional context is not whether their current tools have this flaw. They almost certainly do. The question is how much it matters for the specific tasks they're using AI for, and whether the vendors they work with are moving toward fixing it.