Memdoor · 2026-10-31 · notable
Memdoor team-solve — five Qwen3.8-27B agents rerun the GVS5H coding test
In short
Memdoor's team-solve workflow rebuilds the GVS5H method: five Qwen3.8-27B roles share files on disk. On 10 LiveCodeBench Hard problems over 5 passes it solved 9.8 on average; the authors' Fable 5 files show 8.0.
Memdoor · 2026-10-31 · 3 sources checked
A small-team workflow of five Qwen3.8-27B agents reruns a paper's claim that it can match Fable 5 on hard coding problems.
Key specs
| Mean solved (10 problems) | 9.8 |
|---|---|
| Cost of 50 runs | $35.79 |
What is it?
team-solve is an example workflow for Memdoor, an Apache-2.0 coding and automation agent in one CLI binary. It reimplements GVS5H, an August 2026 arXiv paper whose authors report that orchestrated Qwen3.8-27B reaches 92.4% on 100 hard LiveCodeBench problems. The Memdoor author ran the authors' first 10 problems five times on 2026-10-31.
How does it work?
A planner, two ideators, a solver and a manager, plus a repair worker, each start from a fresh context and talk only through files on disk. One ideator looks for the most direct provable algorithm, the other for a technique from a different family. A shell script runs the public sample tests to decide success, and scoring uses the hidden tests.
Why does it matter?
The rerun scored a mean of 9.8 of 10, against 9.6 for the authors' own 27B harness and 8.0 for Fable 5 alone, for $35.79 across 50 runs on OpenRouter. The README warns that this is only 10 problems, that the March–April 2025 contests may be in the model's training data, and that the Fable 5 column comes from the authors' files, not from this harness.
Who is it for?
developers testing open-weight coding agents
Try it
https://github.com/guregodevo/memdoor/tree/main/examples/workflows/team-solve