Slop Code Bench
An AI coding benchmark created by Gabe Orlansky's lab at the University of Wisconsin. It evaluates whether models cause code quality degradation over extended iteration by having models build a project from scratch and then continuously append new features.
Timeline (last 90 days)
In the Slop Code Bench benchmark, a certain frontier model scored only 14.8%.
The Slop Code Bench benchmark simulates real development scenarios of multiple rounds of iteration and continuous feature additions, examining the long-term maintainability of codebases.
Slop Code Bench was built by Gabe Orlansky's lab at the University of Wisconsin, and was designed to demonstrate that models cause codebases to deteriorate over time.
Slop Code Bench's approach is to first have the model build a project from scratch, then continuously append features, with each challenge containing 3 to 10 checkpoints.
In Slop Code Bench tests, the number of failing tests showed an upward trend for all models.
In one Slop Code Bench test, GPT-5.5 Codex scored approximately 15%, and Astra approximately 16.5%.
All Facts (5)
In the Slop Code Bench benchmark, a certain frontier model scored only 14.8%.
50%UnverifiedThe Slop Code Bench benchmark simulates real development scenarios of multiple rounds of iteration and continuous feature additions, examining the long-term maintainability of codebases.
50%UnverifiedSlop Code Bench was built by Gabe Orlansky's lab at the University of Wisconsin, and was designed to demonstrate that models cause codebases to deteriorate over time.
50%UnverifiedSlop Code Bench's approach is to first have the model build a project from scratch, then continuously append features, with each challenge containing 3 to 10 checkpoints.
50%UnverifiedIn Slop Code Bench tests, the number of failing tests showed an upward trend for all models.
50%