# Multiplication Accuracy on Negative Integers (gpt-oss-120b)

> LLMs can reliably perform simple multiplication involving negative integers below 100 in absolute value.

Date: 2026-01-11



## Test

Follow-up to the [positive multiplication scratchpad](/scratchpads/multiplication-accuracy-positive-integers). Negative multiplication combines sign handling with magnitude computation — the model needs to apply sign rules (neg × pos = neg, neg × neg = pos, anything × 0 = 0) on top of the multiplication itself.

100 multiplication problems, stratified into 3 cases:
- **One negative** (40 pairs): one operand in [-99, -1], the other in [1, 99] — result is negative
- **Both negative** (40 pairs): both operands in [-99, -1] — result is positive
- **Zero result** (20 pairs): one operand is zero

Product magnitudes ranged from 0 to 9,120.

- 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."
- 1 run per pair, 2-second delay between calls
- 1 call hit a 429 rate limit and was retried successfully


## Result

**CONFIRMED**

100% accuracy across all 100 problems. No sign errors, no magnitude errors, no parse failures.

| Case | Total | Correct | Accuracy |
|------|:-:|:-:|:-:|
| One negative (result negative) | 40 | 40 | 100.0% |
| Both negative (result positive) | 40 | 40 | 100.0% |
| Zero result | 20 | 20 | 100.0% |

Sign rules were applied correctly in every case. Combined with the three prior experiments ([addition positive](/scratchpads/simple-addition-positive-integers), [addition negative](/scratchpads/simple-addition-negative-integers), [multiplication positive](/scratchpads/multiplication-accuracy-positive-integers)), gpt-oss-120b is now 400/400 on basic two-operand arithmetic with integers below 100 in absolute value.


## Next

1. Scale operands to 3-digit and 4-digit ranges to find accuracy limits
2. Test division, which introduces remainders and potential floating-point ambiguity
3. Run the full suite on a non-reasoning model to isolate the chain-of-thought effect