Solving math word problems

OpenAI has a new system for cracking grade school math word problems, and the headline number is simple: it nearly doubles the accuracy of a fine-tuned GPT-3 model on the same tests. That's not a marginal tweak. That's the kind of jump that makes you wonder what changed under the hood, even when the announcement itself stays fairly terse about the mechanics.

The comparison point that actually lands, though, isn't another model. It's kids. OpenAI ran a small sample of 9 to 12 year olds through a test built from its own dataset and watched them score 60%. The new system landed at 55% on those identical problems. So the machine is now solving roughly nine out of every ten problems that a human fourth or fifth grader can solve, on questions written for that exact age group.

Grade school math problems sound trivial until you remember why they've been a persistent headache for language models: they require multi-step reasoning, not just pattern matching on text. A model can be fluent and still trip over "if Sarah has three times as many apples as Tom, and Tom has four fewer than Jane..." That's arithmetic wrapped in a small logic puzzle, and it's exactly the kind of task where earlier GPT-3 setups stumbled without heavy fine-tuning.

Closing half the gap to human performance on that specific skill, using the same underlying model family, says more about training approach than about raw scale. OpenAI doesn't frame this as solved — 55% still leaves plenty of room for error — but the direction is the story: a system built to check its own math is getting a lot closer to how a ten-year-old actually thinks through a problem, one step at a time.