SCRATCHPADS-Experiment

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

LLMs will show lower accuracy on float addition involving negative numbers, combining decimal handling with sign handling challenges.

Test

Combines two challenges that individually produced no errors: decimal handling (positive floats) and sign handling (negative integers). Tests whether the combination introduces errors.

100 addition problems with negative floats (absolute values 0.01–99.99, up to 2 decimal places, seed=42), stratified into 4 cases:

  • One negative, positive result (30 pairs)

  • One negative, negative result (30 pairs)

  • One negative, zero result (10 pairs): b = -a

  • Both negative (30 pairs)

  • 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

100% accuracy across all 100 problems. The combination of decimal handling and sign handling produced zero errors.

Case Total Correct Accuracy
One negative, positive result 30 30 100.0%
One negative, negative result 30 30 100.0%
One negative, zero result 10 10 100.0%
Both negative 30 30 100.0%

Combined with all prior arithmetic experiments, gpt-oss-120b is now 600/600 on basic two-operand arithmetic (addition and multiplication, integers and floats, positive and negative) with operands below 100 in absolute value.

Next
  1. Test float subtraction and division with negative operands
  2. Scale up decimal places to find precision limits
  3. Run on a non-reasoning model to see if chain-of-thought is carrying the result