We asked @TencentHunyuan authors why generating a harder coding task keeps…

AlphaSignal says it asked authors on Tencent Hy's team why generating harder synthetic coding tasks for agent training kept failing, per its post. Their answer: a usable terminal task depends on four things staying consistent, the public instructions a model sees, the workspace it operates in, the official grading script, and hidden tests it does not see, and a mismatch between any of them, like a script that fails inside its own container, invalidates the task.

Per AlphaSignal, the researchers' method grows a verified child task from an existing parent, and in doing so official grading scripts grew from an average 67 lines to 374 lines, 5.6 times longer, while public instructions barely grew, 85 words to 122. A model frozen at its original weights, DeepSeek-V4-Pro, saw its pass rate on these tasks fall from 90% to 2.5% as later, harder rounds were generated, while the process still accepted roughly 500 tasks per 1,000 attempts.

This account describes a Tencent Hunyuan research paper that was not located and read directly for this piece, so these specific figures should be read as AlphaSignal's paraphrase of the authors' claims rather than independently confirmed numbers from the paper itself.