stack_machine_golf

GitHub
hardinstructionsText
results

Measured by instruction count — lower is better.

5132 insn1026 insn
sota
Claude-Opus-4.6·Gemini-3.1-Pro·DeepSeek-V4-Pro
reward 1.000

Usage

Run the reference answer to verify your environment is set up correctly

$ harbor run -p tasks/stack_machine_golf

Test model

$ harbor run -p tasks/stack_machine_golf \
  -a claude-code -m claude-opus-4-6

Description

Write a program for a pure stack machine (no registers) to compute a 256-element integer dot product using as few executed instructions as possible. The VM supports stack/memory ops, basic arithmetic, and simple control flow. The result must be on top of the stack at halt.

Files

path
permission
/app/program.stk✎ Edit
/app/main.cRead-only
/app/MakefileRead-only

Rules

  • 01Edit /app/program.stk only.
  • 02Scratch memory at addresses 1024..1039.
  • 03Wrong final result or infinite loops score 0.

Tags

stack-machinesuperoptimizationassemblyvminstruction-countcode-golf

Model Results

Click a row to view its trajectory in Live Lab

model
reward
score
Claude-Opus-4.6
1.000
Gemini-3.1-Pro
1.000
DeepSeek-V4-Pro
1.000
Grok-4-20
0.960
MiMo-V2.5-Pro
0.790
Kimi-K2.6
0.670
GLM-5
0.480
Hunyuan-3-Preview
0.380
MiniMax-M2.7
0.190
GPT-5.4
0.160
Qwen-3.6-Plus
0.000
<- all tasks