Why LLMs still make math errors.

Modern language models can write competent code, translate between 40 languages, and produce a working three-act play. They also occasionally get 234 × 87 wrong. Here's why.

The core reason

Language models predict the next token based on training patterns. They don't have an internal calculator. When you ask "what's 234 × 87," the model produces plausible-looking digits based on similarity to problems it saw during training — not by actually computing the multiplication.

For simple arithmetic (single-digit, common multiplication tables), the model has enough training examples that it can produce correct answers by pattern-matching. For less-common combinations, especially multi-digit multiplication with carries, the model can drift.

The failure mode is often subtle. The model might get the first few digits right and slip on the middle ones. It might round unexpectedly. It might invent a plausible-looking number that's off by a factor of 10.

Where models get math right anyway

- **Well-known constants.** Pi to a few decimals, e, common conversion factors. Enough training coverage that recall works. - **Simple arithmetic.** Addition and subtraction of small numbers, single-digit multiplication. - **Explaining math.** Talking about how to solve a problem, describing an algorithm, walking through a proof. These are language tasks the model handles well. - **Setting up the equation.** Given a word problem, the model translates it correctly to symbolic form.

Where they trip

- **Multi-digit multiplication with carries.** The kind of computation where you'd naturally reach for a calculator. - **Long chains of arithmetic.** Errors compound. - **Statistics on real datasets.** Confusing means and medians, drifting on standard deviations. - **Financial calculations with compounding.** Compound interest over N periods with rate changes. - **Unit conversions with unusual units.** Sometimes right, sometimes off by 10× or 1000×.

What extended thinking does for math

Turning on Deep + Extended thinking meaningfully improves accuracy on multi-step problems. The model uses the reasoning budget to do intermediate work more carefully — including breaking multiplication into digit-by-digit steps.

Still not perfect. Still worth verifying for anything material.

The right pattern

- **Use the model to set up the problem.** Translate the word problem to symbols. Identify the right formula. - **Actually compute with a calculator.** Or ask the model to write Python code you can run. Or (LADLE roadmap) use the code-execution tool. - **Ask the model to explain what the answer means.** Once you have a correct number, the model is great at contextualizing.

Why this is fixable but not fixed

Anthropic and other labs have shown improvements in multi-step math with each generation. Techniques like extended thinking, tool use (giving the model a real calculator), and code execution close the gap materially. But at the model-only level, arithmetic is a category the pattern-matching architecture will always be shaky at.

The compensating principle: for anything numerical and stakes-material, verify with a calculator or code execution.