Why AI Models Keep Failing the Same Tool-Use Tests (and a Fix That Learns From Mistakes)

Key points

• On a standard tool-use benchmark called τ²-bench, 57% of trials run by the "teacher" AI end in failure, meaning more than half the training data gets discarded.
• Two-thirds of those failures are near-misses: the AI got most steps right but stumbled at the finish line.
• Current training pipelines repeat the same expensive process daily or weekly, yet keep hitting the same hard problems because failures provide no learning signal.
• PROOF-Gen, a method from Apple ML Research, recovers value from those near-misses instead of discarding them.
• The approach could cut the cost and repetition of training tool-calling AI agents without sacrificing quality.

Every time you ask an AI assistant to book a flight, check your calendar or pull up a file, it is using what researchers call "tool-calling": the ability to reach out and operate software on your behalf. Teaching a smaller, cheaper AI model to do this reliably is harder than it looks.

What is the problem, exactly?

Right now, the standard method is a bit like hiring an expensive tutor to write model answers, then only keeping the perfect ones. A powerful "teacher" AI runs through thousands of practice scenarios. Anything it gets wrong gets binned. The student model only learns from the clean successes.

The catch: on τ²-bench, a widely used test for tool-calling AI, 57% of the teacher's attempts fail outright. Two-thirds of those failures are near-misses where nearly every step was correct except the last one or two. All that almost-right work vanishes.

Worse, because the failures get discarded, the same hard scenarios keep tripping the system up in the next training cycle, and the one after that. Companies running these pipelines pay the frontier-teacher cost, the expensive compute bill for running a top-tier AI, every single week, with little to show for the repeated effort on the toughest tasks.

What does PROOF-Gen actually do?

PROOF-Gen, described by Apple ML Research, treats near-misses as raw material rather than rubbish. Instead of a simple pass-or-fail filter, it looks at a failed trajectory (the sequence of steps the AI took) and works out which parts were correct. It then constructs a partial but valid training example from those good steps, so the student model can still learn from them.

Think of it like a driving instructor who, after a student stalls on a hill start, still gives credit for the mirror checks and smooth steering that came before. The lesson is not lost just because the car stopped.

A teacher preparing a week of school dinners might draft a menu, realise one dish needs an ingredient she does not have, and scrap the whole plan. PROOF-Gen is the equivalent of salvaging four of the five perfectly good meals and only replacing the one that did not work.

What does this mean for everyday AI tools?

For ordinary users, the benefit is indirect but real. Tool-calling agents that schedule meetings, draft emails or manage your shopping list are only as reliable as the training that shaped them. A more efficient training process means developers can build more capable assistants at lower cost, and fix stubborn failure patterns that currently just keep recurring.

Privacy note: this research is about how AI models are trained in the lab, not about data collected from your personal use. No user data is involved in the PROOF-Gen method itself.

Common questions

Do I need to do anything differently as a user?

No. This is a behind-the-scenes training improvement. If it is adopted, your AI assistant simply gets better at multi-step tasks over time.

Is PROOF-Gen available in any product yet?

Apple ML Research has published the method as a research paper. It has not been announced as part of any shipping product at the time of writing.

Why does it matter that failures are near-misses?

Because a near-miss contains mostly correct information. Discarding it wastes the majority of a good answer. Recovering it means the student model learns faster and handles tricky scenarios it would otherwise keep failing.