# Taurus: LLM-in-the-Loop CT Annotation Matches Coding Agents at Half the Cost [Canonical reading](https://escapemeridian.com/readings/taurus-llm-in-the-loop-ct-annotation-matches-coding-agents-at-9f0816d4) Bing Hu and Rishabh Sharma present Taurus, a system for labeling aortas in CT scans, and compare it with coding agents that write their own segmentation code. Taurus uses an automated labeling pass followed by model review: when checks detect likely large errors, the model chooses among corrections shown on the scan. The authors say an early version gained little from model review, while a revised version incorporated useful parts of the free-form agents’ work into automation. On their evaluation, the final Taurus version matched the free-form Opus agent on the primary metric at less than half the per-case cost, and matched the surface smoothness of the human references. The authors’ central argument is that the harness and automated process, rather than the language model alone, accounted for most of Taurus’s competitiveness. They propose this kind of automation as a way to shift expert time from drawing labels to checking them. The result is evidence about annotation on the datasets and cases they tested; it does not by itself establish how Taurus would perform in other settings. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 30, 2026 (2026-10-30). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-08T22:03:32.180Z. First observed: 2026-11-03T09:04:07.371Z. Retrieved: 2026-11-08T22:02:05.615Z. - [Taurus: LLM-in-the-Loop CT Annotation Matches Coding Agents at Half the Cost](https://arxiv.org/abs/2611.00247) ## Limits The reported evaluation uses six public datasets and 38 held-out cases; the evidence does not establish performance on a broader range of datasets or cases. The source reports no difference on the primary metric, not proof that the approaches are equivalent in all respects or settings. The source describes annotation results and a possible shift in expert work; it does not report the effects of that shift in practice. ## Open question Would Taurus retain its reported performance and cost advantage on a broader range of aortic CT datasets and cases? ## Source-stated Hu and Sharma compare Taurus, a harness that automates labeling and asks a model to review detected errors, with free-form agents that write segmentation code. > We built a harness, Taurus, and compared it with free-form agents by blind label recovery - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The evaluation withheld labels from six public datasets containing 939 volumes and scored approaches against each dataset’s own classes. > we withheld the human labels of six public datasets (939 volumes) and scored every approach against each dataset's own classes. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The authors report that the final Taurus version matched the free-form Opus agent on the primary metric at less than half its per-case cost. > in the primary metric we found no difference from the free-form Opus agent, at under half its cost, $1.22 against $2.71 per case. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The authors say that most of Taurus’s competitive performance came from its harness and automation, rather than the language model itself. > Most of what made Taurus competitive came from the harness and its automation rather than from the language model itself. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The authors suggest automation could shift expert time from drawing labels to checking them. > Taurus shows that this kind of annotation can be automated, so that expert time can move from drawing labels to checking them. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source and contributor context Hu, Bing, Sharma, Rishabh · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 83fe04e7-bcd3-48ee-868a-402a44b85eab. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T22:03:32.180Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: ce026ef41c79d8e49cb5034d4785de4a8cddb49bbaf51039e05680a43308d841. Source text hash: c0684adae9bd97fe88adc10c333bf9c84e9b715d375eda7f84175fa6e91ef306. ## Revision and source-check history ### Version 1 — 2026-11-08T22:03:32.180Z Bing Hu and Rishabh Sharma present Taurus, a system for labeling aortas in CT scans, and compare it with coding agents that write their own segmentation code. Taurus uses an automated labeling pass followed by model review: when checks detect likely large errors, the model chooses among corrections shown on the scan. The authors say an early version gained little from model review, while a revised version incorporated useful parts of the free-form agents’ work into automation. On their evaluation, the final Taurus version matched the free-form Opus agent on the primary metric at less than half the per-case cost, and matched the surface smoothness of the human references. The authors’ central argument is that the harness and automated process, rather than the language model alone, accounted for most of Taurus’s competitiveness. They propose this kind of automation as a way to shift expert time from drawing labels to checking them. The result is evidence about annotation on the datasets and cases they tested; it does not by itself establish how Taurus would perform in other settings. ## Original source arXiv cs.AI · Article Original source date: Oct 30, 2026 (2026-10-30). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-08T22:03:32.180Z. First observed: 2026-11-03T09:04:07.371Z. Retrieved: 2026-11-08T22:02:05.615Z. - [Taurus: LLM-in-the-Loop CT Annotation Matches Coding Agents at Half the Cost](https://arxiv.org/abs/2611.00247) ## Limits The reported evaluation uses six public datasets and 38 held-out cases; the evidence does not establish performance on a broader range of datasets or cases. The source reports no difference on the primary metric, not proof that the approaches are equivalent in all respects or settings. The source describes annotation results and a possible shift in expert work; it does not report the effects of that shift in practice. ## Open question Would Taurus retain its reported performance and cost advantage on a broader range of aortic CT datasets and cases? ## Source-stated Hu and Sharma compare Taurus, a harness that automates labeling and asks a model to review detected errors, with free-form agents that write segmentation code. > We built a harness, Taurus, and compared it with free-form agents by blind label recovery - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The evaluation withheld labels from six public datasets containing 939 volumes and scored approaches against each dataset’s own classes. > we withheld the human labels of six public datasets (939 volumes) and scored every approach against each dataset's own classes. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The authors report that the final Taurus version matched the free-form Opus agent on the primary metric at less than half its per-case cost. > in the primary metric we found no difference from the free-form Opus agent, at under half its cost, $1.22 against $2.71 per case. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The authors say that most of Taurus’s competitive performance came from its harness and automation, rather than the language model itself. > Most of what made Taurus competitive came from the harness and its automation rather than from the language model itself. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source-stated The authors suggest automation could shift expert time from drawing labels to checking them. > Taurus shows that this kind of annotation can be automated, so that expert time can move from drawing labels to checking them. - [Supporting source](https://arxiv.org/abs/2611.00247) ## Source and contributor context Hu, Bing, Sharma, Rishabh · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 83fe04e7-bcd3-48ee-868a-402a44b85eab. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T22:03:32.180Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: ce026ef41c79d8e49cb5034d4785de4a8cddb49bbaf51039e05680a43308d841. Source text hash: c0684adae9bd97fe88adc10c333bf9c84e9b715d375eda7f84175fa6e91ef306. ### uncertain — 2026-11-08T22:03:32.180Z — anthropic/claude-opus-5.5 Publication: 83fe04e7-bcd3-48ee-868a-402a44b85eab attribution — pass: The reading correctly attributes the work to Bing Hu and Rishabh Sharma and keeps the results and the central argument framed as the authors' own reported findings and claims. > Authors:Bing Hu, Rishabh Sharma summary — uncertain: The summary is faithful overall, but it refers to 'the free-form Opus agent' without saying it was a free-form agent running Claude Opus 5.5. A general reader cannot identify the main comparison point. 'CT' and 'segmentation code' are also central terms left unexplained; tooltips or brief glosses are needed. The summary omits the dollar figures ($1.22 vs $2.71 per case) and the motivating problem: labels are slow to draw and public datasets use inconsistent conventions. Smallest repair: name Claude Opus 5.5, give the per-case costs, and add one clause on why aortic labeling is hard. > A free-form agent running Claude Opus 5.5 scored higher on most datasets. date — pass: The originalPublished date of 2026-10-30 matches the v1 submission date. The 5 Nov revision and the 6 Nov feed date are correctly not used. > \[Submitted on 30 Oct 2026 (v1), last revised 5 Nov 2026 (this version, v2)\] support — pass: Every claim is source-stated and carries an exact excerpt. The early-version and revision narrative matches the abstract. The limitations appropriately note that 'no difference' does not mean equivalence and that the shift in expert work is a proposal. Minor caution: the abstract ties the 38 held-out cases to the gain from moving agent strengths into automation. It is not explicit that the final comparison used only those cases, so the limitation should not imply that it did. The question follows from the evidence. > Moving what the free-form agents did well into the automated pass then carried most of the gain on 38 held-out cases. media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/9f0816d43f617fa2b1137fb3ebc75e4bc157b2809b49ea81e7233bc36db078c8/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/taurus-llm-in-the-loop-ct-annotation-matches-coding-agents-at-9f0816d4.json)