CANDOR Reveals Weakness in Frozen Foundation Encoders

Soroosh Tayebi Arasteh, Sven Nebelung, Daniel Truhn· July 22, 2026 View original

Summary

This research introduces CANDOR, a chance-calibrated discordance measure for frozen foundation encoders, revealing that many encoders are weaker than previously thought in separating features for specific findings. It shows that even high-performing models can misclassify a significant percentage of cases due to feature geometry issues, not just selection.

When evaluating frozen foundation encoders, the common practice is to assess how well a lightweight classification head can interpret their features. However, this approach can be misleading because it doesn't account for the inherent geometric separation of features within the encoder itself. This paper introduces CANDOR, a novel discordance measure designed to provide a chance-calibrated assessment of feature separation. CANDOR addresses a critical flaw in previous nearest-neighbor discordance methods by ensuring equal-sized banks for opposite labels, making the measure symmetric and fixing its chance level at exactly one half. This correction dramatically alters conclusions about encoder performance. Across a wide range of encoders, datasets, and domains, CANDOR reveals that while no encoder is truly "blind," many are significantly weaker than assumed. For instance, a chest model achieving 84.5 AUROC still places nearly one-fifth of its positive pneumothorax cases closer to an opposite-label film. This indicates that the problem lies in the encoder's feature geometry, not just the choice of classification head, and highlights a substantial deficit in information retention for certain findings.

Why it matters

Professionals relying on frozen foundation models for downstream tasks need to understand the true limitations of these encoders, especially regarding feature separation for specific, critical findings. This can impact model selection and fine-tuning strategies.

How to implement this in your domain

  1. 1Integrate CANDOR into your evaluation pipeline for foundation models to assess the intrinsic feature separation capabilities of frozen encoders.
  2. 2Prioritize fine-tuning or adapting foundation models for tasks where CANDOR scores indicate poor feature separation for critical classes.
  3. 3Use CANDOR to identify specific findings or classes that are poorly represented in the latent space of a chosen encoder, guiding data augmentation or model architecture choices.
  4. 4Develop strategies to improve the geometric separation of features in foundation models during pre-training or domain adaptation.

Who benefits

Healthcare AIComputer VisionAI/ML DevelopmentQuality Assurance

Key takeaways

  • CANDOR provides a more accurate, chance-calibrated measure of feature separation in frozen encoders.
  • Many foundation encoders are weaker than previously believed in distinguishing specific findings.
  • Poor performance can stem from encoder feature geometry, not just classifier selection.
  • Erasure retention is linked to performance collapse, highlighting data representation issues.

Original post by Soroosh Tayebi Arasteh, Sven Nebelung, Daniel Truhn

"arXiv:2607.18451v1 Announce Type: new Abstract: Frozen encoders are chosen by how well a lightweight head reads a finding from their features, not whether the geometry separates it. Nearest-neighbor discordance does, but with unequal banks the opposite-label neighbor wins on dens…"

View on X

Originally posted by Soroosh Tayebi Arasteh, Sven Nebelung, Daniel Truhn on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI Engineering & DevToolsAI Research

New Tool Generates Contamination-Resistant, Labeled Code Datasets for LLMs

Spaghetti Architect is a new open-source tool that generates controlled, multi-language code datasets, addressing issues of contamination and lack of semantic control in existing code corpora. It creates correct-by-construction programs with adjustable "messiness" and difficulty labels, making it ideal for training and evaluating code-generating LLMs.

Yuxiang JiJul 22, 2026
AI ResearchAI Engineering & DevTools

New Method Safely Gates Hazardous LLM Knowledge Without Deletion

Researchers introduce Token Inoculation, a method that allows large language models to retain sensitive "dual-use" knowledge while selectively refusing hazardous queries. This approach uses a special token to condition the model's behavior, improving safety without sacrificing benign domain performance.

Seunghyun Lee, Dongyoon Han, Sangdoo YunJul 22, 2026
AI ResearchAI Engineering & DevTools

GNNAS-TSP Selects Optimal Algorithms for Traveling Salesman Problem

Researchers introduce GNNAS-TSP, a Graph Neural Network (GNN)-based framework for automated algorithm selection (AS) for the Traveling Salesman Problem (TSP). GNNAS-TSP learns TSP instance representations directly from raw graph data, avoiding manual feature engineering, and formulates AS as a joint cost-prediction and ranking task to select the best solver from a portfolio under fixed computational budgets.

Zhaoxuan Li, Jiale Yang, Yifei Lu, Mustafa MisirJul 22, 2026