Tool-using agents that act in the world need to be both useful and safe. Well-calibrated model confidences can be used to weigh the risk versus reward of potential actions, but prior work shows that many models are poorly calibrated. Inspired by interpretability literature exploring the internals of models, we propose a novel class of model-internal confidence estimators (MICE) to better assess confidence when calling tools. MICE first decodes from each intermediate layer of the language model using logitLens and then computes similarity scores between each layer's generation and the final output. These features are fed into a learned probabilistic classifier to assess confidence in the decoded output. On the simulated trial and error (STE) tool-calling dataset using Llama3 models, we find that MICE beats or matches the baselines on smoothed expected calibration error. Using MICE confidences to determine whether to call a tool significantly improves over strong baselines on a new metric, expected tool-calling utility. Further experiments show that MICE is sample-efficient, can generalize zero-shot to unseen APIs, and results in higher tool-calling utility in scenarios with varying risk levels. Our code is open source, available at https://github.com/microsoft/mice_for_cats.

This paper introduces the Bayesian Confidence Estimator (BACON) for deep neural networks. Current practice of interpreting Softmax values in the output layer as probabilities of outcomes is prone to extreme predictions of class probability. In this work we extend Waagen's method of representing the terminal layers with a geometric model, where the probability associated with an output vector is estimated with Bayes' Rule using validation data to provide likelihood and normalization values. This estimator provides superior ECE and ACE calibration error compared to Softmax for ResNet-18 at 85% network accuracy, and EfficientNet-B0 at 95% network accuracy, on the CIFAR-10 dataset with an imbalanced test set, except for very high accuracy edge cases. In addition, when using the ACE metric, BACON demonstrated improved calibration error when estimating probabilities for the imbalanced test set when using actual class distribution fractions.

In high-stakes decision-making tasks within legal NLP, such as Case Outcome Classification (COC), quantifying a model's predictive confidence is crucial. Confidence estimation enables humans to make more informed decisions, particularly when the model's certainty is low, or where the consequences of a mistake are significant. However, most existing COC works prioritize high task performance over model reliability. This paper conducts an empirical investigation into how various design choices including pre-training corpus, confidence estimator and fine-tuning loss affect the reliability of COC models within the framework of selective prediction. Our experiments on the multi-label COC task, focusing on European Court of Human Rights (ECtHR) cases, highlight the importance of a diverse yet domain-specific pre-training corpus for better calibration. Additionally, we demonstrate that larger models tend to exhibit overconfidence, Monte Carlo dropout methods produce reliable confidence estimates, and confident error regularization effectively mitigates overconfidence. To our knowledge, this is the first systematic exploration of selective prediction in legal NLP. Our findings underscore the need for further research on enhancing confidence measurement and improving the trustworthiness of models in the legal domain.

Temporal Knowledge Graph (TKG), which characterizes temporally evolving facts in the form of (subject, relation, object, timestamp), has attracted much attention recently. TKG reasoning aims to predict future facts based on given historical ones. However, existing TKG reasoning models are unable to abstain from predictions they are uncertain, which will inevitably bring risks in real-world applications. Thus, in this paper, we propose an abstention mechanism for TKG reasoning, which helps the existing models make selective, instead of indiscriminate, predictions. Specifically, we develop a confidence estimator, called Confidence Estimator with History (CEHis), to enable the existing TKG reasoning models to first estimate their confidence in making predictions, and then abstain from those with low confidence. To do so, CEHis takes two kinds of information into consideration, namely, the certainty of the current prediction and the accuracy of historical predictions. Experiments with representative TKG reasoning models on two benchmark datasets demonstrate the effectiveness of the proposed CEHis.

Semantic segmentation plays a crucial role in various computer vision applications, yet its efficacy is often hindered by the lack of high-quality labeled data. To address this challenge, a common strategy is to leverage models trained on data from different populations, such as publicly available datasets. This approach, however, leads to the distribution shift problem, presenting a reduced performance on the population of interest. In scenarios where model errors can have significant consequences, selective prediction methods offer a means to mitigate risks and reduce reliance on expert supervision. This paper investigates selective prediction for semantic segmentation in low-resource settings, thus focusing on post-hoc confidence estimators applied to pre-trained models operating under distribution shift. We propose a novel image-level confidence measure tailored for semantic segmentation and demonstrate its effectiveness through experiments on three medical imaging tasks. Our findings show that post-hoc confidence estimators offer a cost-effective approach to reducing the impacts of distribution shift.

This paper addresses the problem of selective classification for deep neural networks, where a model is allowed to abstain from low-confidence predictions to avoid potential errors. We focus on so-called post-hoc methods, which replace the confidence estimator of a given classifier without retraining or modifying it, thus being practically appealing. Considering neural networks with softmax outputs, our goal is to identify the best confidence estimator that can be computed directly from the unnormalized logits. This problem is motivated by the intriguing observation in recent work that many classifiers appear to have a "broken" confidence estimator, in the sense that their selective classification performance is much worse than what could be expected by their corresponding accuracies. We perform an extensive experimental study of many existing and proposed confidence estimators applied to 84 pretrained ImageNet classifiers available from popular repositories. Our results show that a simple $p$-norm normalization of the logits, followed by taking the maximum logit as the confidence estimator, can lead to considerable gains in selective classification performance, completely fixing the pathological behavior observed in many classifiers. As a consequence, the selective classification performance of any classifier becomes almost entirely determined by its corresponding accuracy. Moreover, these results are shown to be consistent under distribution shift. We also investigate why certain classifiers innately have a good confidence estimator that apparently cannot be improved by post-hoc methods.

We studied the capability of automated machine translation in the online video education space by automatically translating Khan Academy videos with state-of-the-art translation models and applying text-to-speech synthesis and audio/video synchronization to build engaging videos in target languages. We also analyzed and established two reliable translation confidence estimators based on round-trip translations in order to efficiently manage translation quality and reduce human translation effort. Finally, we developed a deployable system to deliver translated videos to end users and collect user corrections for iterative improvement.

Existing models for named entity recognition (NER) are mainly based on large-scale labeled datasets, which always obtain using crowdsourcing. However, it is hard to obtain a unified and correct label via majority voting from multiple annotators for NER due to the large labeling space and complexity of this task. To address this problem, we aim to utilize the original multi-annotator labels directly. Particularly, we propose a Confidence-based Partial Label Learning (CPLL) method to integrate the prior confidence (given by annotators) and posterior confidences (learned by models) for crowd-annotated NER. This model learns a token- and content-dependent confidence via an Expectation-Maximization (EM) algorithm by minimizing empirical risk. The true posterior estimator and confidence estimator perform iteratively to update the true posterior and confidence respectively. We conduct extensive experimental results on both real-world and synthetic datasets, which show that our model can improve performance effectively compared with strong baselines.

Motivated by the above limitations, we propose a framework called POsthoc pseudo Ood REgularization Detecting Out-of-Distribution (OOD) (Goodfellow (POORE) that generates pseudo-OOD data et al., 2014; Hendrycks and Gimpel, 2016; using the trained classifier and the In-Distribution Yang et al., 2021) samples is vital for developing (IND) samples. As opposed to methods that use reliable machine learning systems for various outlier exposure, our framework doesn't rely on any industry-scale applications of natural language understanding external OOD set. Moreover, POORE can be easily (NLP) (Shen et al., 2019; Sundararaman applied to already deployed large-scale models et al., 2020) including intent understanding trained on a classification task, without requiring in conversational dialogues (Zheng et al., 2020; to re-train the classifier from scratch. In summary, Li et al., 2017), language translation (Denkowski we make the following contributions: and Lavie, 2011; Sundararaman et al., 2019), and text classification (Aggarwal and Zhai, 2012; Sundararaman 1. We propose a Mahalanobis-based context et al., 2022). For instance, a language masking scheme for generating pseudo-OOD understanding model deployed to support a chat samples that can be used during the finetuning.

Although modern automatic speech recognition (ASR) systems can achieve high performance, they may produce errors that weaken readers' experience and do harm to downstream tasks. To improve the accuracy and reliability of ASR hypotheses, we propose a cross-modal post-processing system for speech recognizers, which 1) fuses acoustic features and textual features from different modalities, 2) joints a confidence estimator and an error corrector in multi-task learning fashion and 3) unifies error correction and utterance rejection modules. Compared with single-modal or single-task models, our proposed system is proved to be more effective and efficient. Experiment result shows that our post-processing system leads to more than 10% relative reduction of character error rate (CER) for both single-speaker and multi-speaker speech on our industrial ASR system, with about 1.7ms latency for each token, which ensures that extra latency introduced by post-processing is acceptable in streaming speech recognition.