SCRATCHPADS-Experiment

Addition Accuracy on Positive Floats (gpt-oss-120b) 2026-01-18
Hypothesis

LLMs will show lower accuracy on float addition compared to integer addition, due to decimal place handling.

Test

Follow-up to the integer addition experiments (positive, negative), both 100% accurate. Floating point introduces decimal alignment and carry propagation across the decimal point.

100 pairs of positive floats (0.01–99.99, up to 2 decimal places, seed=42), stratified into 5 buckets of 20:

  • Integer-like: x.0 + y.0 (sanity check)

  • One decimal: x.x + y.y

  • Two decimals, no carry: fractional parts sum below 1.00

  • Two decimals, carry: fractional parts sum to 1.00 or above

  • Mixed precision: one operand with fewer decimal places than the other

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

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

  • Prompt: "What is {a} + {b}? Reply with only the number."

  • Correctness: round(model, 2) == round(expected, 2) using Decimal arithmetic

Result

REJECTED

The hypothesis is wrong — float addition accuracy matches integer addition at 100%.

Bucket Total Correct Accuracy
Integer-like 20 20 100.0%
One decimal 20 20 100.0%
Two dec, no carry 20 20 100.0%
Two dec, carry 20 20 100.0%
Mixed precision 20 20 100.0%

Carry across the decimal point and mismatched decimal places introduced no errors. The model's chain-of-thought handles decimal alignment correctly throughout.

Combined with the four prior arithmetic experiments, gpt-oss-120b is now 500/500 on basic two-operand arithmetic with operands below 100.

Next
  1. Increase decimal places (3-4+) to find where precision breaks down
  2. Test float subtraction and multiplication at the same range
  3. Run the same problems on a non-reasoning model to isolate chain-of-thought effects