Local Calibration and Grouping Loss

24 concepts16 papers1 proposal

A model that is "calibrated" should provide confidence scores that reflect its true probability of being correct. If a model predicts an event with 80% confidence, that event should happen exactly 80% of the time. However, standard ways of measuring this—like the Expected Calibration Error (ECE)—often provide a false sense of security. A model can appear perfectly calibrated on average while being systematically wrong for specific groups of data.

The Illusion of Global Calibration

The fundamental problem with aggregate calibration metrics is the cancellation effect. As noted by Chenru Duan et al. in Uncertain of uncertainties? A comparison of uncertainty quantification metrics for chemical data sets, scalar metrics average performance across the entire dataset. This allows a model to be overconfident in one region of the feature space and underconfident in another, with the two errors mathematically canceling each other out in the final average.

Rachel Luo et al. demonstrate in Local Calibration: Metrics and Recalibration that visualizing these "local error landscapes" reveals significant biases that ECE masks. For example, a medical model might appear reliable overall but consistently overestimate its accuracy for a specific diagnostic category, such as atrial fibrillation, as shown by Ciaran Bench et al. in Towards Trustworthy Atrial Fibrillation Classification from Wearables Data: Quantifying Model Uncertainty.

Understanding Grouping Loss

This phenomenon is increasingly referred to as grouping loss. The authors of Reconfidencing LLMs from the Grouping Loss Perspective define it as the variance in reliability across different internal clusters or categories. In Large Language Models (LLMs), grouping loss often manifests as high reliability for "famous" entities but catastrophic overconfidence for minority groups or long-tail topics.

Sébastien Melo et al. argue in Epistemic Uncertainty Quantification To Improve Decisions From Black-Box Models that grouping loss hides localized model ignorance. When a model encounters a specific sub-population that differs from the aggregate average, its global calibration guarantees vanish, making the system fragile under distribution shifts, a point emphasized by Yaodong Yu et al. in Robust Calibration with Multi-domain Temperature Scaling (MD-TS).

Measuring Local Miscalibration

To expose these hidden biases, researchers have proposed more granular metrics that evaluate reliability within specific partitions of the data:

Strategies for Local Recalibration

Fixing local miscalibration requires moving beyond global scaling factors like standard Temperature Scaling. Instead, recent work focuses on localized recalibration.

One approach is to use non-parametric methods to adjust scores based on local density or similarity. Rachel Luo et al. propose LoRe (Local Recalibration), while Miao Xiong et al. introduce PROCAL, which adjusts confidence based on local data density to mitigate proximity bias. For LLMs, Lihu Chen et al. use decision trees to identify problematic subgroups and "reconfidence" them individually.

Other researchers aim for even stricter guarantees. Shengjia Zhao et al. propose individual calibration using randomized forecasting in Individual Calibration with Randomized Forecasting, ensuring that uncertainty guarantees hold for every single data point rather than just on average. Similarly, Deming Sheng et al. replace the standard Softmax with an Asymmetric Laplace Distribution (ALD) to provide theoretical individual-level calibration guarantees in From Individual Calibration to Reliable Classifiers: ALD Parameterization with mPAIC Guarantees.

By shifting focus from global averages to local neighborhoods, these methods ensure that a model's confidence is meaningful not just for the "average" user, but for every specific individual and sub-domain the model encounters.

Key Foundational Work

Local Calibration: Metrics and Recalibration (2022) — contributed 2 concept elements to this direction.
Individual Calibration with Randomized Forecasting (2020) — contributed 1 concept elements to this direction.
Robust Calibration with Multi-domain Temperature Scaling (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_grouping_loss_770745db,
  title = {grouping loss},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grouping-loss/rdr_770745db4abf4490a37ada894eba1aeb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_grouping_loss_770745db_v1,
  title = {grouping loss},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/grouping-loss/rdr_770745db4abf4490a37ada894eba1aeb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake