We are delighted to announce that two papers from our research lab have been accepted at the ECCV 2026 CDEL Workshop (Computer Vision for Data Curation, Efficiency & Learning).

We are delighted to announce that two papers from our research lab have been accepted at the ECCV 2026 CDEL Workshop (Computer Vision for Data Curation, Efficiency & Learning), held as part of the official ECCV 2026 program in Malmö, Sweden.

---

MedCurate-Bench: Auditing the Diagnostic Validity of Curated Medical Image Datasets

Authors: Sarthak Pandey, Shreshth Rai, Seifedine Kadry

MedCurate-Bench investigates an important but often overlooked question in medical dataset curation: does improving dataset accuracy necessarily preserve diagnostic validity?

The study demonstrates that curation strategies optimized solely for accuracy can inadvertently compromise clinically important properties such as calibration and rare-class sensitivity.

Key Highlights:

• Evaluates different curation strategies across five medical imaging modalities.
• Covers approximately 3,700 experimental runs.
• Identifies which pruning and distillation strategies preserve diagnostic validity rather than optimizing accuracy alone.

---

When Agreement Is Not Enough: A Selection Bottleneck in Non-Verifiable Reasoning

Authors: Shreshth Rai, Sarthak Pandey, Seifedine Kadry

This work examines the widely used practice of agreement-based reasoning trace curation, where multiple reasoning traces are sampled and those that agree with one another are preferentially selected.

Through extensive evaluation, the paper shows that this strategy can break down in open-ended multimodal reasoning, where multiple distinct explanations may be valid.

Key Highlights:

• Demonstrates limitations of agreement-based selection in open-ended multimodal reasoning.
• Shows the bottleneck is identifying valid traces, not just generating more of them.
• Proves that increasing the sampling budget alone does not eliminate this limitation.

---

Both papers will be presented at the ECCV 2026 CDEL Workshop in Malmö, Sweden, with members of our lab attending the conference in person.

The camera-ready versions and accompanying code will be released soon.

We are proud to contribute to ongoing research at the intersection of computer vision, medical AI, data curation, and efficient learning, and look forward to sharing our work with the research community at ECCV 2026.