Why AI Gets Stuck Halfway Through a Hard Task (and a New Fix From Apple)

Key points

• Apple ML Research published a study showing that current large language models fail to complete long, multi-step reasoning tasks reliably, even when given a clear plan to follow.
• The failure is not random: a small number of "hard" steps cause almost all the errors, and once an AI slips on one of them, it rarely recovers.
• Splitting a task into too many tiny steps makes the problem worse, because there is no room to correct a mistake before the next step locks it in.
• A new technique called LEAD (Lookahead-Enhanced Atomic Decomposition) adds a short forward check after each step to catch errors before they become permanent.
• The research was tested on controlled puzzle tasks, so real-world results will need further study.

Imagine asking a very capable assistant to assemble flat-pack furniture by following a 40-step instruction sheet. They are sharp, motivated, and they have the full plan. But there are three steps that are genuinely tricky. Miss one of them and every later step builds on a crooked frame. By step 35 the whole thing is unsalvageable.

That is, roughly, the problem Apple ML Research has put under the microscope.

What is actually going wrong?

Large language models, the AI systems behind chatbots like ChatGPT and Claude, struggle to complete long chains of reasoning without slipping up. Giving them a high-level strategy helps, but it does not solve the problem entirely.

The Apple team found something specific. Errors do not spread evenly across a task. A handful of steps are responsible for nearly all the failures. Miss one of those "hard" steps and the AI has no way back.

The instinct among researchers has been to slice tasks into smaller and smaller pieces so the AI only has to think about one tiny thing at a time. Fair enough. But the study shows that going too far creates what the researchers call a "no-recovery bottleneck": each piece is so small and so rigidly linked to the next that a single wrong move is immediately cemented into the chain.

What does LEAD actually do?

The fix is called LEAD, short for Lookahead-Enhanced Atomic Decomposition. It keeps the small-step approach but adds a short forward check after each step.

Before committing to a move, the system peeks a few steps ahead to ask: does this path still lead somewhere sensible? If not, it can adjust before the error becomes permanent. Think of it as the AI briefly scanning the next few squares on the board before placing its piece.

The researchers tested LEAD on algorithmic puzzles, controlled tasks with clear right-and-wrong answers, which made it easier to measure exactly where things broke down.

Should anyone act on this right now?

Not yet. This is a research paper, not a product update. The tests used structured puzzles, and puzzle performance does not always translate cleanly to the messy, open-ended tasks that real users put to AI assistants.

What it does tell us is that the next wave of AI improvements may come not from making models bigger, but from making them better at checking their own work mid-task. That matters for anyone who asks an AI to handle something with many steps: writing a long report, debugging code, or planning a project.

Watch for whether this technique (or something like it) appears in future model updates from the big labs.

Common questions

Does this affect the AI tools I use today?

Not directly. LEAD is a research technique, not yet a shipped feature in any consumer product. Current tools still suffer from the bottleneck the paper describes, especially on long or complex tasks.

Why test on puzzles instead of real tasks?

Puzzles have definite right answers, which lets researchers measure exactly where and how often the AI fails. Real-world tasks are harder to score precisely, so puzzles are a useful first proving ground.