AI

Taurus: LLM-in-the-Loop CT Annotation Matches Coding Agents at Half the Cost

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 03/11/2026, 09:03:25 UTC.

Briefings

A new system called Taurus uses an LLM to automatically annotate CT scans of the aorta, matching the quality of human experts at a fraction of the cost. The study compared a free-form agent against a human-in-the-loop harness and found the automated approach performed similarly on most datasets while being significantly cheaper. The final version included automated checks to fix common errors, resulting in no difference in primary metrics compared to the free-form agent but at under half the cost.

Key points

• The Taurus system uses a free-form LLM agent to write custom segmentation code for CT scans of the aorta, matching human expert performance on most datasets.
• The automated approach achieved similar primary metrics to the human-in-the-loop harness while costing under half the price per case.

Context

This research addresses the bottleneck of slow, inconsistent labeling of dense aortic CT labels, which currently requires significant expert time. The study compares a free-form LLM agent against a human-in-the-loop harness to determine which method is more efficient and accurate for public datasets.

Limitations

The study utilized six public datasets with 939 volumes, limiting generalizability to this specific scope. The initial version showed the harness raised metrics on only two datasets, and the final version matched human references but did not replicate the specific gains of the harness. The cost savings were measured per case, and the study did not account for long-term operational costs or scalability beyond the initial six datasets.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/11/2026, 10:50:49 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibipp2qglge6g2iligtdzomdgt2epzjbdi7gba65fvslb5d6wreem