← Back to feed

Intology published results from Locus, its automated AI research system, showing it post-trained a Qwen3-1.7B base model to scores that exceed Alibaba’s official human-tuned Qwen3-1.7B release. On PostTrainBench+, the expanded evaluation framework Intology built for the test, Locus achieved 51.6% against the human-tuned baseline’s 49.4%.

The 2.2-point gap is narrow but the mechanism matters: Locus did the work autonomously, without human researchers directing the training runs.

What PostTrainBench Measures

PostTrainBench is an external benchmark designed to evaluate the capability of AI agents at post-training other models. Each agent receives a fixed budget — one H100 GPU for 10 hours — and is scored on the quality of the model it produces.

Intology concluded the standard 10-hour budget introduced too much noise to distinguish methods. It extended the compute budget to 100 hours per run and raised the total allocation to 4,500 H100 hours, calling the result PostTrainBench+. At that scale, the performance separation between methods becomes statistically reliable.

The capability Locus is exercising: allocating compute across parallel training jobs, reading evaluation results as they arrive, abandoning underperforming branches, and scaling up the ones that show promise — the core loop of a human AI researcher compressed into an automated system.

Kaggle as a Real-World Test

Intology also ran Locus on all live Kaggle competitions carrying prize money and public leaderboards. After 16 days of continuous operation, Locus held the 4th highest average rank among all participants across every active competition.

Kaggle competitions are not controlled benchmarks. They involve heterogeneous tasks, real-time competition from thousands of human teams, and leaderboard dynamics that reward adaptation over time. A 4th-place average rank across all active competitions is a more direct indicator of general research automation capability than any single benchmark score.

Production Deployment

Intology states that models post-trained by Locus are already serving production users. The exact scale is described as “millions” without further breakdown, suggesting the system is not in a research-only phase.

The practical implication: automated post-training systems are not a future capability being benchmarked in controlled settings. They are shipping products using AI-optimized weights on AI-sized workloads.

The Broader Pattern

Locus joins a cluster of systems announced in 2026 — Sakana AI’s Darwin Gödel Machine, Recursive’s Automated Research System, and the OpenAI Astra internal run on 10 open math problems — all pointing at the same structural shift: AI systems are increasingly the ones doing the work of improving AI systems.

The limiting variable in each case is not the quality of the automation; it is the compute budget and the diversity of the evaluation signal. Intology’s decision to raise PostTrainBench from 70 H100 hours to 4,500 to get clean signal is a preview of how those evaluation frameworks will need to evolve as automated research systems mature.