Memdoor 的 Qwen3.8-27B 多智能體工作流在 LiveCodeBench Hard 小樣本測試中高於 Fable 5

Five small models solve a programming problem together: a planner, two ideators, a solver and a manager, plus a repair worker. Each starts from a fresh context and talks only through files on disk. A script that runs the samples decides success, not a model. It is the method of GVS5H (Lee, 2026), written as a Memdoor workflow: eight YAML files, no code, no model named.

The result

The authors' harness rebuilt on the same model, Qwen3.8-27B, with their prompts, loop and 128K reply cap. The first ten problems of their LiveCodeBench Hard list, five passes each, graded on every hidden test with their corrected grader:

The authors' two columns are the regraded result files in their repository, the numbers in their paper. Our column is bench/results/parity-5pass/: all 50 runs, each with every request and reply, the program, the grader's output and the cost. A run took a median of 40 minutes and $0.39; the 50 cost $35.79.

What this does and does not show:

• Ten problems, five passes, not the authors' 100. On a sample this small, 9.8 and 9.6 are the same number. The claim is parity with their harness, not a gain on it.
• The contests date from spring 2025 and the model may have seen them in training. Contamination cannot be ruled out, for any column.
• Fable 5's column is one call per problem, from the authors' files. It was not run in this harness.
• The loop sees only the public samples; the score is the hidden tests.
• The model ran on OpenRouter, not a local machine.
• The one lost problem: the manager invented a wrong counterexample and the team rewrote a correct program. Twice elsewhere a correct program was overwritten and then recovered. Both gaps are the authors' loop, kept as is.

The parity check (every request byte for byte against their code), the hosts, the changes it took and the first run's findings are in NOTES.md. The same result, on one page: memdoor.ai/docs/team-solve.

Run it

curl -fsSL https://memdoor.ai/install.sh | bash
memdoor connect openrouter        # or anthropic, openai, gemini, chatgpt
# or no key: memdoor connect local   (Ollama, llama.cpp, vLLM, LM Studio; Ollama needs a 32K window, not its 4K default)
git clone --depth 1 https://github.com/guregodevo/memdoor

mkdir -p ~/.memdoor/workflows
cp -r memdoor/examples/workflows/team-solve ~/.memdoor/workflows/
mkdir my-problem && cd my-problem
cp -r ../memdoor/examples/workflows/team-solve/sample/* .    # task.md + tests/1.in … 5.out
memdoor workflow run team-solve --wait   # exit 0 when every sample passes

For your own problem, replace task.md and put paired tests/N.in / tests/N.out beside it. The program is solution.py, read from stdin, compared to the samples with diff; it needs Python 3 and bash. /workflow in memdoor tui draws the run live.

The steps name no model: whatever /model is pinned to answers, a local one included (how to connect one). To change a role, change its prompt in steps/.

How it works

plan ──┬── ideate-a ──┬── solve ── test ── review ── repair ── verify
       └── ideate-b ──┘

Every agent prompt says "fresh context, read the files, do not rely on chat". The paper credits that ledger discipline for the gain. Each step is one YAML file in steps/, in the format of mario, the engine underneath; requires: is the DAG.