GLM-5.2 Ranks Top in Fine-Tuning Leaderboard Sparking Doubt, Benchmark Author Clarifies: Claude Was Not Distilled

According to Watchful AI monitoring, the open-source model GLM-5.2 reached the top of the self-tuning benchmark PostTrainBench. However, it has been criticized by skeptic scaling01 for lacking practical value. He pointed out that the model's rapid rise from 22nd place to first within a few months is highly unusual. He also mentioned that the testing, due to a lack of a hidden set, is prone to inducing the AI agent to engage in gaming behavior for targeted optimization, resulting in a model that is difficult to deploy in the real world.

However, supporters countered that under the constraint of a single H100 GPU limited to 10 hours, expecting the AI agent to complete a universal fine-tuning is not realistic, and targeted optimization is a norm in machine learning. The public logs demonstrate that GLM-5.2 follows a clear experimental logic, automatically collects data under different sampling assumptions, autonomously plans a complete pipeline encompassing establishing performance baselines, fine-tuning, and utilizing rejection sampling to filter data, while attempting to mitigate overfitting in the thought chain.

The greater significance of this incident lies in the fact that the originally intended public run trajectory to assess fine-tuning capability unexpectedly debunked the industry rumor of the domestic large model "heavy distillation Claude." Benchmark author Maksym Andriushchenko, after reviewing the GLM-5.2 logs, pointed out that the model exhibits fundamental differences in data collection, strategy combination, and decision paths compared to Claude, showing no signs of imitation or distillation. The third-party benchmark's transparency has, in turn, become the most direct window into the open-source large model's self-proven original research and development capabilities.