Locus Automated Research System Surpasses Human Post-Training on Qwen3 1.7B

Locus achieves SOTA on PostTrainBench by automating the post-training process for Qwen3 base models. Using an expanded PostTrainBench+ compute budget, Locus-trained 1.7B models outperformed the official human-post-trained Qwen3 version and secured a top-5 average rank across active Kaggle competitions.