From the paper of Sunday, November 1, 2026
Five small models, working as a team, score 9.8 of 10 on LiveCodeBench Hard
Five Qwen3.8-27B models solved the first ten problems of LiveCodeBench Hard together, five passes each, averaging 9.8 of 10 hidden tests passed. 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. A script decides whether a run succeeded, not a model.
The workflow implements a method published as GVS5H (Lee, 2026), written as eight YAML files with no code and no model named. Whatever the run is pinned to answers, a local model included.
The same model run through the authors' own harness averaged 9.6 on the same problems. Fable 5, given one call per problem, averaged 8.0 in the authors' result files. The workflow's authors state the comparison's limits: ten problems and five passes, not the authors' hundred, and on a sample that small 9.8 and 9.6 are the same number. The claim is parity with the authors' harness, not a gain on it. Fable 5's column was not run in this harness. The contests date from spring 2025 and the model may have seen them in training; contamination cannot be ruled out for any column. The loop sees only the public samples; the score is the hidden tests.
A run took a median of 40 minutes and $0.39. The 50 runs cost $35.79. The model ran on OpenRouter, not a local machine.
The one problem the team lost was the manager inventing a wrong counterexample, after which 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.
Every agent prompt says "fresh context, read the files, do not rely on chat"; the paper behind the method credits that discipline for the gain. The workflow runs without an API key against Ollama, llama.cpp, vLLM or LM Studio. Ollama needs a 32K context window, not its 4K default.