SCRATCHPADS-Experiment

Division Accuracy on Integers (gpt-oss-120b) 2026-01-22
Hypothesis

LLMs will show lower accuracy on division compared to addition and multiplication, particularly for non-terminating decimals and division by zero.

Test

Division introduces complexity the prior operations didn't have: remainders, repeating decimals, and the undefined case of division by zero. All prior experiments (addition and multiplication, integers and floats, positive and negative) showed 100% accuracy.

100 division problems with integer operands (1-99), stratified into 5 cases:

  • Clean (25 pairs): a % b == 0, integer result

  • Terminating (25 pairs): a/b is a terminating decimal

  • Repeating (25 pairs): a/b is a repeating decimal

  • Division by one (15 pairs): b = 1

  • Division by zero (10 pairs): b = 0

  • Model: gpt-oss-120b via Cerebras API (free tier)

  • Temperature: 0, top_p: 1, max_completion_tokens: 1024

  • Prompt (cases 1-4): "What is {a} / {b}? Reply with only the number."

  • Prompt (case 5): "What is {a} / {b}? Reply with only the number, or 'undefined' if the operation is not possible."

  • Repeating decimals checked by rounding both answers to 4 decimal places

  • Parser handles both decimal and fraction responses (e.g. "4/19")

Result

REJECTED

100% accuracy across all 100 problems, including repeating decimals and division by zero.

Case Total Correct Accuracy
Clean (integer result) 25 25 100.0%
Terminating decimal 25 25 100.0%
Repeating decimal 25 25 100.0%
Division by one 15 15 100.0%
Division by zero 10 10 100.0%

For repeating decimals, the model sometimes returned simplified fractions (e.g. "4/19" for 8/38, "8/17" for 32/68) rather than decimal approximations — mathematically correct and arguably more precise. Division by zero was consistently identified as "undefined" with no hallucinated numeric values.

Combined with all prior experiments, gpt-oss-120b is now 700/700 across all four basic operations with operands below 100.

Next
  1. Scale up to 3-digit and 4-digit operands to find where division accuracy breaks down
  2. Test division with float operands
  3. Run on a non-reasoning model — the chain-of-thought likely matters most for division, where the model visibly simplifies fractions before answering