Papers with calibration error

18 papers
Calibration of Pre-trained Transformers (2020.emnlp-main)

Copied to clipboard

Challenge: Pre-trained Transformers dominate benchmark tasks but use a large number of self-attention heads across many layers in a way that is difficult to unpack.
Approach: They analyze pre-trained Transformer models' posterior probabilities to determine whether they are calibrated for three tasks: natural language inference, paraphrase detection, and commonsense reasoning.
Outcome: The models are calibrated in-domain and out-of-domain, and their calibration error out-domain can be as much as 3.5x lower.
Predict the Next Word: <Humans exhibit uncertainty in this task and language models _____> (2024.eacl-short)

Copied to clipboard

Challenge: Language models (LMs) are statistical models trained to assign probability to human-generated text.
Approach: They evaluate language models' ability to reproduce variability that humans exhibit in the ‘next word prediction’ task.
Outcome: The language models are trained to assign probability to human-generated text . they exhibit low calibration to human uncertainty, and advise against it .
Calibration vs Decision Making: Revisiting the Reliability Paradox in Unlearned Language Models (2026.acl-srw)

Copied to clipboard

Challenge: Model calibration is commonly used as a proxy for reliability, but low calibration error does not necessarily imply reliable decision rules.
Approach: They investigate the calibration error gap in generative language models using the TOFU benchmark and attribution-based shortcut detection.
Outcome: The proposed model calibrations achieve low calibration error compared to pretrained models and retain low calibration despite reduced accuracy on the forget split.
Calibrating Zero-shot Cross-lingual (Un-)structured Predictions (2022.emnlp-main)

Copied to clipboard

Challenge: Existing need for model calibration when natural language models are deployed in critical tasks.
Approach: They compare model calibration methods in a context of zero-shot cross-lingual transfer with pre-trained language models.
Outcome: The proposed method fails to calibrate more complex confidence estimations in structured predictions compared to expressive alternatives like Gaussian Process Calibration.
On the Effects of Transformer Size on In- and Out-of-Domain Calibration (2021.findings-emnlp)

Copied to clipboard

Challenge: Large pre-trained transformer language models are notoriously expensive to train . prior work has developed smaller, more compact models to reduce training costs .
Approach: They propose to develop smaller, more compact transformer language models which can be calibrated in-domain . they show that smaller models can achieve competitive calibration compared to larger models .
Outcome: The proposed models achieve competitive calibration and better calibration than larger models on a wide range of tasks.
Posterior Calibrated Training on Sentence Classification Tasks (2020.acl-main)

Copied to clipboard

Challenge: Existing methods for posterior calibration have been used to correct poorly calibrated posterior probabilities.
Approach: They propose a posterior calibration procedure that optimizes posterior probability distributions while minimizing calibration errors.
Outcome: The proposed procedure reduces calibration error and improves performance on both objectives.
Platt-Bin: Efficient Posterior Calibrated Training for NLP Classifiers (2022.findings-acl)

Copied to clipboard

Challenge: Existing methods for posterior calibration return uncalibrated estimations of class posteriors, thus leading to poorer generalization.
Approach: They propose an end-to-end trained calibrator that directly optimizes the objective while minimizing the difference between predicted and empirical posterior probabilities.
Outcome: The proposed calibrator reduces calibration error and improves performance on benchmark NLP classification tasks.
A Data Cartography based MixUp for Pre-trained Language Models (2022.naacl-main)

Copied to clipboard

Challenge: Using MixUp, additional samples are generated during training by combining random pairs of training samples and their labels.
Approach: They propose a new MixUp strategy that leverages Training Dynamics and allows more informative samples to be combined for generating new data samples.
Outcome: The proposed method achieves competitive performance using a smaller subset of training data compared with strong baselines and yields lower expected calibration error on the pre-trained language model, BERT, on both in-domain and out-of-domain settings.
Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback (2023.emnlp-main)

Copied to clipboard

Challenge: Recent studies have shown that unsupervised pre-training produces large language models whose conditional probabilities are remarkably well-calibrated.
Approach: They propose to use verbalized confidences to extract confidence from large language models with reinforcement learning from human feedback to improve their accuracy.
Outcome: The proposed methods reduce the expected calibration error by 50% for RLHF-LMs such as ChatGPT, GPT-4, and Claude.
On the Calibration of Pre-trained Language Models using Mixup Guided by Area Under the Margin and Saliency (2022.acl-long)

Copied to clipboard

Challenge: Existing studies have shown that mixing up can improve model calibration on image classification tasks, but little is known about using it on natural language understanding (NLU) tasks.
Approach: They propose a mixup strategy for pre-trained language models that improves model calibration further by using the AUM statistic and saliency map.
Outcome: The proposed mixup improves model calibration on natural language understanding tasks while maintaining competitive accuracy.
Making Heads and Tails of Models with Marginal Calibration for Sparse Tagsets (2021.findings-emnlp)

Copied to clipboard

Challenge: despite high accuracy, modern neural networks can still suffer from severe miscalibration.
Approach: They propose to use tag frequency grouping to measure calibration error in different frequency bands to reduce error.
Outcome: The proposed techniques reduce calibration error across the marginal distribution for two existing sequence taggers.
Uncertainty Quantification with Pre-trained Language Models: A Large-Scale Empirical Analysis (2022.findings-emnlp)

Copied to clipboard

Challenge: Pre-trained language models (PLMs) have gained increasing popularity due to compelling prediction performance in diverse natural language processing tasks.
Approach: They compare three popular options for encoding and Temp Scaling for PLMs . they recommend using Temp Loss as uncertainty quantifier and Focal Loss for fine-tuning .
Outcome: Using pre-trained language models, we compare three options on NLP classification tasks and domain shift.
On Diversified Preferences of Large Language Model Alignment (2024.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) can be fine tuned with human feedback, but human preferences can be diversified due to annotators’ different tastes, which hinders the effectiveness of LLM alignment methods.
Approach: They propose a calibration error metric to evaluate large language models (LLMs) and a multi-objective reward learning method to enhance the calibration performance of RMs on shared preferences.
Outcome: The proposed model can be adopted as a key calibration error and MORE can achieve superior alignment performance.
InternalInspector I2: Robust Confidence Estimation in LLMs through Internal States (2024.findings-emnlp)

Copied to clipboard

Challenge: Large Language Models (LLMs) often struggle with generating reliable outputs, often producing high-confidence inaccuracies known as hallucinations.
Approach: They propose a framework that leverages contrastive learning on internal states including attention states, feed-forward states, and activation states of all layers to enhance confidence estimation in LLMs.
Outcome: The framework outperforms existing methods in the hallucination detection benchmark HaluEval and the previous methods at the same time.
KARL: Knowledge-Aware Retrieval and Representations aid Retention and Learning in Students (2024.emnlp-main)

Copied to clipboard

Challenge: Existing student models use study data like student's past responses to predict the probability a student can recall a flashcard.
Approach: They propose to use student models to predict recall of flashcards to build a content-aware student model that uses deep knowledge tracing, retrieval, and BERT to predict student recall.
Outcome: The proposed content-aware student model outperforms existing student models in AUC and calibration error and is more efficient than SOTA.
Fill In The Gaps: Model Calibration and Generalization with Synthetic Data (2024.emnlp-main)

Copied to clipboard

Challenge: Existing calibration methods negatively impact model accuracy due to the lack of diversity of validation data.
Approach: They propose a calibration method that incorporates synthetic data without compromising accuracy.
Outcome: The proposed method improves model accuracy on real data and reduces calibration error by 34% on four different tasks.
Calibrated Seq2seq Models for Efficient and Generalizable Ultra-fine Entity Typing (2023.findings-emnlp)

Copied to clipboard

Challenge: CASENT predicts ultra-fine entities mentioned in text into types with calibrated confidence scores.
Approach: They propose a model that predicts ultra-fine entities with calibrated confidence scores for entity typing.
Outcome: The proposed model outperforms existing models in terms of F1 score and calibration error while achieving 50 times faster inference speed.
Quantized Can Still Be Calibrated: A Unified Framework to Calibration in Quantized Large Language Models (2025.acl-long)

Copied to clipboard

Challenge: Existing methods to quantify uncertainty of large language models (LLMs) but their influence on uncertainty calibration remains unexplored.
Approach: They propose an analytic method to estimate the upper bound of calibration error (UBCE) for quantized LLMs and propose a method to recover calibration errors through soft-prompt tuning.
Outcome: The proposed method improves the calibration accuracy of quantized models on multiple datasets and LLMs.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations