Large Language Models (LLMs) are known to exhibit social, demographic, and gender biases, often as a consequence of the data on which they are trained. In this work, we adopt a mechanistic interpretability approach to analyze how such biases are structurally represented within models such as GPT-2 and Llama2. Focusing on demographic and gender biases, we explore different metrics to identify the internal edges responsible for biased behavior. We then assess the stability, localization, and generalizability of these components across dataset and linguistic variations. Through systematic ablations, we demonstrate that bias-related computations are highly localized, often concentrated in a small subset of layers. Moreover, the identified components change across fine-tuning settings, including those unrelated to bias. Finally, we show that removing these components not only reduces biased outputs but also affects other NLP tasks, such as named entity recognition and linguistic acceptability judgment because of the sharing of important components with these tasks.

Pre-trained language models have become an integral component of question-answering systems, achieving remarkable performance. For practical deployment, it is critical to carry out knowledge distillation to preserve high performance under computational constraints. In this paper, we address a key question: given the importance of unsupervised distillation for student performance, how does one effectively ensemble knowledge from multiple teachers at this stage without the guidance of ground-truth labels? We propose a novel algorithm, GOVERN, to tackle this issue. GOVERN has demonstrated significant improvements in both offline and online experiments. The proposed algorithm has been successfully deployed in a real-world commercial question-answering system.

We present an automatic large language model (LLM) conversion approach that produces uncertainty-aware LLMs capable of estimating uncertainty with every prediction. Our approach is model- and data-agnostic, is computationally-efficient, and does not rely on external models or systems. We evaluate converted models on the selective question answering setting -- to answer as many questions as possible while maintaining a given accuracy, forgoing providing predictions when necessary. As part of our results, we test BERT and Llama 2 model variants on the SQuAD extractive QA task and the TruthfulQA generative QA task. We show that using the uncertainty estimates provided by our approach to selectively answer questions leads to significantly higher accuracy over directly using model probabilities.

Hyperparameter tuning of deep learning models can lead to order-of-magnitude performance gains for the same amount of compute. Despite this, systematic tuning is uncommon, particularly for large models, which are expensive to evaluate and tend to have many hyperparameters, necessitating difficult judgment calls about tradeoffs, budgets, and search bounds. To address these issues and propose a practical method for robustly tuning large models, we present Cost-Aware Pareto Region Bayesian Search (CARBS), a Bayesian optimization algorithm that performs local search around the performance-cost Pareto frontier. CARBS does well even in unbounded search spaces with many hyperparameters, learns scaling relationships so that it can tune models even as they are scaled up, and automates much of the "black magic" of tuning. Among our results, we effectively solve the entire ProcGen benchmark just by tuning a simple baseline (PPO, as provided in the original ProcGen paper). We also reproduce the model size vs. training tokens scaling result from the Chinchilla project (Hoffmann et al. 2022), while simultaneously discovering scaling laws for every other hyperparameter, via an easy automated process that uses significantly less compute and is applicable to any deep learning problem (not just language models).

Out-of-distribution (OOD) detection for text applications is a rapidly growing field due to new robustness and security requirements driven by an increased number of AI-based systems. Existing OOD textual detectors often rely on an anomaly score (e.g., Mahalanobis distance) computed on the embedding output of the last layer of the encoder. In this work, we begin by uncovering that the fact that performance of existent methods varies greatly depending on the task and choice of the layer output. More importantly, we show that the usual choice (the last layer) is rarely the best one and thus, far better results could be achieved if the best layer were chosen. To leverage our key observation, we propose a data-driven, unsupervised method to combine layer-wise anomaly scores. In addition, we extend classical textual OOD benchmarks by including classification tasks with a greater number of classes (up to 77), which reflects more realistic settings. On this augmented benchmark, we show that the proposed post-aggregation methods achieve robust and consistent results while removing manual feature selection altogether. Their performance achieves near oracle's best layer performance.

We consider high dimensional $M$-estimation in settings where the response $Y$ is possibly missing at random and the covariates $\mathbf{X} \in \mathbb{R}^p$ can be high dimensional compared to the sample size $n$. The parameter of interest $\boldsymbol{\theta}_0 \in \mathbb{R}^d$ is defined as the minimizer of the risk of a convex loss, under a fully non-parametric model, and $\boldsymbol{\theta}_0$ itself is high dimensional which is a key distinction from existing works. Standard high dimensional regression and series estimation with possibly misspecified models and missing $Y$ are included as special cases, as well as their counterparts in causal inference using 'potential outcomes'. Assuming $\boldsymbol{\theta}_0$ is $s$-sparse ($s \ll n$), we propose an $L_1$-regularized debiased and doubly robust (DDR) estimator of $\boldsymbol{\theta}_0$ based on a high dimensional adaptation of the traditional double robust (DR) estimator's construction. Under mild tail assumptions and arbitrarily chosen (working) models for the propensity score (PS) and the outcome regression (OR) estimators, satisfying only some high-level conditions, we establish finite sample performance bounds for the DDR estimator showing its (optimal) $L_2$ error rate to be $\sqrt{s (\log d)/ n}$ when both models are correct, and its consistency and DR properties when only one of them is correct. Further, when both the models are correct, we propose a desparsified version of our DDR estimator that satisfies an asymptotic linear expansion and facilitates inference on low dimensional components of $\boldsymbol{\theta}_0$. Finally, we discuss various of choices of high dimensional parametric/semi-parametric working models for the PS and OR estimators. All results are validated via detailed simulations.

In this paper we address a classification problem where two sources of labels with different levels of fidelity are available. Our approach is to combine data from both sources by applying a co-kriging schema on latent functions, which allows the model to account item-dependent labeling discrepancy. We provide an extension for Laplace inference for Gaussian process classification, that takes into account multi-fidelity data. We evaluate the proposed method on real and synthetic datasets and show that it is more resistant to different levels of discrepancy between sources, than other approaches for data fusion. Our method can provide accuracy/cost trade-off for a number of practical tasks such as crowd-sourced data annotation and feasibility regions construction in engineering design.