Previewing Locus

Summary

Locus is Intology's automated AI research/post-training system. Per the Discord embed, it sustains improvement over days, exceeds human experts on RE-Bench at equal time and compute, and sets SOTA on KernelBench and MLE-Bench Lite. Related Intology material describes Locus as SOTA on PostTrainBench (agents post-training models given 10 H100 hours) and PostTrainBench+, with Locus post-trained Qwen3 models surpassing human post-trained Qwen3 and running in production.

Keywords

AI research agents · post-training · RE-Bench · KernelBench · MLE-Bench

Topics

AI research agents, post-training, RE-Bench, KernelBench, MLE-Bench

Research notes

• Method: Automated post-training agent that runs for days; evaluated on PostTrainBench/PostTrainBench+ (post-training Qwen3 1.7B-Base with thousands of H100 hours).
• Key findings: Exceeds human experts on RE-Bench at equal time and compute (per Intology); SOTA on KernelBench and MLE-Bench Lite (per Intology); SOTA on PostTrainBench and PostTrainBench+; Locus post-trained Qwen3 beats human post-trained Qwen3
• Limitations: All claims are from Intology's own materials; independent verification not checked.
• Blog page fetch was rate-limited; summary combines the Discord embed with Intology's announcement summary found via web search. Confirm details against the blog post before final cataloging.