Aligning Large Language Models (LLMs) to cater to different human preferences, learning new skills, and unlearning harmful behavior is an important problem. Search-based methods, such as Best-of-N or Monte-Carlo Tree Search, are performant, but impractical for LLM adaptation due to their high inference cost. On the other hand, using Reinforcement Learning (RL) for adaptation is computationally efficient, but performs worse due to the optimization challenges in co-training the value function and the policy. We present a new framework for reward optimization, Value Augmented Sampling (VAS), that can maximize different reward functions using data sampled from only the initial, frozen LLM. VAS solves for the optimal reward-maximizing policy without co-training the policy and the value function, making the optimization stable, outperforming established baselines, such as PPO and DPO, on standard benchmarks, and achieving comparable results to Best-of-128 with lower inference cost. Unlike existing RL methods that require changing the weights of the LLM, VAS does not require access to the weights of the pre-trained LLM. Thus, it can even adapt LLMs (e.g., ChatGPT), which are available only as APIs. In addition, our algorithm unlocks the new capability of composing several rewards and controlling the extent of each one during deployment time, paving the road ahead for the future of aligned, personalized LLMs.

We study the problem of incentive-compatible online learning with bandit feedback. In this class of problems, the experts are self-interested agents who might misrepresent their preferences with the goal of being selected most often. The goal is to devise algorithms which are simultaneously incentive-compatible, that is the experts are incentivised to report their true preferences, and have no regret with respect to the preferences of the best fixed expert in hindsight. \citet{freeman2020no} propose an algorithm in the full information setting with optimal $O(\sqrt{T \log(K)})$ regret and $O(T^{2/3}(K\log(K))^{1/3})$ regret in the bandit setting. In this work we propose the first incentive-compatible algorithms that enjoy $O(\sqrt{KT})$ regret bounds. We further demonstrate how simple loss-biasing allows the algorithm proposed in Freeman et al. 2020 to enjoy $\tilde O(\sqrt{KT})$ regret. As a byproduct of our approach we obtain the first bandit algorithm with nearly optimal regret bounds in the adversarial setting which works entirely on the observed loss sequence without the need for importance-weighted estimators. Finally, we provide an incentive-compatible algorithm that enjoys asymptotically optimal best-of-both-worlds regret guarantees, i.e., logarithmic regret in the stochastic regime as well as worst-case $O(\sqrt{KT})$ regret.

Deep learning has been able to outperform humans in terms of classification accuracy in many tasks. However, to achieve robustness to adversarial perturbations, the best methodologies require to perform adversarial training on a much larger training set that has been typically augmented using generative models (e.g., diffusion models). Our main objective in this work, is to reduce these data requirements while achieving the same or better accuracy-robustness trade-offs. We focus on data pruning, where some training samples are removed based on the distance to the model classification boundary (i.e., margin). We find that the existing approaches that prune samples with low margin fails to increase robustness when we add a lot of synthetic data, and explain this situation with a perceptron learning task. Moreover, we find that pruning high margin samples for better accuracy increases the harmful impact of mislabeled perturbed data in adversarial training, hurting both robustness and accuracy. We thus propose PUMA, a new data pruning strategy that computes the margin using DeepFool, and prunes the training samples of highest margin without hurting performance by jointly adjusting the training attack norm on the samples of lowest margin. We show that PUMA can be used on top of the current state-of-the-art methodology in robustness, and it is able to significantly improve the model performance unlike the existing data pruning strategies. Not only PUMA achieves similar robustness with less data, but it also significantly increases the model accuracy, improving the performance trade-off.

Although recent advances in machine learning have shown its success to learn from independent and identically distributed (IID) data, it is vulnerable to out-of-distribution (OOD) data in an open world. Domain generalization (DG) deals with such an issue and it aims to learn a model from multiple source domains that can be generalized to unseen target domains. Existing studies on DG have largely focused on stationary settings with homogeneous source domains. However, in many applications, domains may evolve along a specific direction (e.g., time, space). Without accounting for such non-stationary patterns, models trained with existing methods may fail to generalize on OOD data. In this paper, we study domain generalization in non-stationary environment. We first examine the impact of environmental non-stationarity on model performance and establish the theoretical upper bounds for the model error at target domains. Then, we propose a novel algorithm based on adaptive invariant representation learning, which leverages the non-stationary pattern to train a model that attains good performance on target domains. Experiments on both synthetic and real data validate the proposed algorithm.

Online social media platforms, such as Twitter, provide valuable information during disaster events. Existing tweet disaster summarization approaches provide a summary of these events to aid government agencies, humanitarian organizations, etc., to ensure effective disaster response. In the literature, there are two types of approaches for disaster summarization, namely, supervised and unsupervised approaches. Although supervised approaches are typically more effective, they necessitate a sizable number of disaster event summaries for testing and training. However, there is a lack of good number of disaster summary datasets for training and evaluation. This motivates us to add more datasets to make supervised learning approaches more efficient. In this paper, we present ADSumm, which adds annotated ground-truth summaries for eight disaster events which consist of both natural and man-made disaster events belonging to seven different countries. Our experimental analysis shows that the newly added datasets improve the performance of the supervised summarization approaches by 8-28% in terms of ROUGE-N F1-score. Moreover, in newly annotated dataset, we have added a category label for each input tweet which helps to ensure good coverage from different categories in summary. Additionally, we have added two other features relevance label and key-phrase, which provide information about the quality of a tweet and explanation about the inclusion of the tweet into summary, respectively. For ground-truth summary creation, we provide the annotation procedure adapted in detail, which has not been described in existing literature. Experimental analysis shows the quality of ground-truth summary is very good with Coverage, Relevance and Diversity.

Uncertainty estimation is a significant issue for current large language models (LLMs) that are generally poorly calibrated and over-confident, especially with reinforcement learning from human feedback (RLHF). Unlike humans, whose decisions and confidences not only stem from intrinsic beliefs but can also be adjusted through daily observations, existing calibration methods for LLMs focus on estimating or eliciting individual confidence without taking full advantage of the "Collective Wisdom": the interaction among multiple LLMs that can collectively improve both accuracy and calibration. In this work, we propose Collaborative Calibration, a post-hoc training-free calibration strategy that leverages the collaborative and expressive capabilities of multiple tool-augmented LLM agents in a simulated group deliberation process. While contemporary large language models (LLMs) have achieved remarkable performance in a variety of tasks ranging from question answering to complex reasoning (Brown et al., 2020; Bubeck et al., 2023), it remains a significant bottleneck for them to produce well-calibrated confidence estimates for their predictions, meaning that their individual confidence is not a reliable indicator of accuracy. Models still often generate hallucinations (Bubeck et al., 2023) or wildly wrong predictions, unknowingly and over-confidently, which is found to be more evident for models fine-tuned with RLHF (Kadavath et al., 2022; Tian et al., 2023). On the other hand, models can exhibit inconsistencies and lack of confidence, by blindly altering decisions and prioritizing incorrect user opinions (Wei et al., 2023).

Nowadays, individuals tend to engage in dialogues with Large Language Models, seeking answers to their questions. In times when such answers are readily accessible to anyone, the stimulation and preservation of human's cognitive abilities, as well as the assurance of maintaining good reasoning skills by humans becomes crucial. This study addresses such needs by proposing hints (instead of final answers or before giving answers) as a viable solution. We introduce a framework for the automatic hint generation for factoid questions, employing it to construct TriviaHG, a novel large-scale dataset featuring 160,230 hints corresponding to 16,645 questions from the TriviaQA dataset. Additionally, we present an automatic evaluation method that measures the Convergence and Familiarity quality attributes of hints. To evaluate the TriviaHG dataset and the proposed evaluation method, we enlisted 10 individuals to annotate 2,791 hints and tasked 6 humans with answering questions using the provided hints. The effectiveness of hints varied, with success rates of 96%, 78%, and 36% for questions with easy, medium, and hard answers, respectively. Moreover, the proposed automatic evaluation methods showed a robust correlation with annotators' results. Conclusively, the findings highlight three key insights: the facilitative role of hints in resolving unknown questions, the dependence of hint quality on answer difficulty, and the feasibility of employing automatic evaluation methods for hint assessment.

While considerable theoretical progress has been devoted to the study of metric and similarity learning, the generalization mystery is still missing. In this paper, we study the generalization performance of metric and similarity learning by leveraging the specific structure of the true metric (the target function). Specifically, by deriving the explicit form of the true metric for metric and similarity learning with the hinge loss, we construct a structured deep ReLU neural network as an approximation of the true metric, whose approximation ability relies on the network complexity. Here, the network complexity corresponds to the depth, the number of nonzero weights and the computation units of the network. Consider the hypothesis space which consists of the structured deep ReLU networks, we develop the excess generalization error bounds for a metric and similarity learning problem by estimating the approximation error and the estimation error carefully. An optimal excess risk rate is derived by choosing the proper capacity of the constructed hypothesis space. To the best of our knowledge, this is the first-ever-known generalization analysis providing the excess generalization error for metric and similarity learning. In addition, we investigate the properties of the true metric of metric and similarity learning with general losses.

As Large Language Models (LLMs) are nondeterministic, the same input can generate different outputs, some of which may be incorrect or hallucinated. If run again, the LLM may correct itself and produce the correct answer. Unfortunately, most LLM-powered systems resort to single results which, correct or not, users accept. Having the LLM produce multiple outputs may help identify disagreements or alternatives. However, it is not obvious how the user will interpret conflicts or inconsistencies. To this end, we investigate how users perceive the AI model and comprehend the generated information when they receive multiple, potentially inconsistent, outputs. Through a preliminary study, we identified five types of output inconsistencies. Based on these categories, we conducted a study (N=252) in which participants were given one or more LLM-generated passages to an information-seeking question. We found that inconsistency within multiple LLM-generated outputs lowered the participants' perceived AI capacity, while also increasing their comprehension of the given information. Specifically, we observed that this positive effect of inconsistencies was most significant for participants who read two passages, compared to those who read three. Based on these findings, we present design implications that, instead of regarding LLM output inconsistencies as a drawback, we can reveal the potential inconsistencies to transparently indicate the limitations of these models and promote critical LLM usage.

Large Language Models (LLMs) have become capable of generating highly fluent text in certain languages, without modules specially designed to capture grammar or semantic coherence. What does this mean for the future of linguistic expertise in NLP? We highlight several aspects in which NLP (still) relies on linguistics, or where linguistic thinking can illuminate new directions. We argue our case around the acronym $RELIES$ that encapsulates six major facets where linguistics contributes to NLP: $R$esources, $E$valuation, $L$ow-resource settings, $I$nterpretability, $E$xplanation, and the $S$tudy of language. This list is not exhaustive, nor is linguistics the main point of reference for every effort under these themes; but at a macro level, these facets highlight the enduring importance of studying machine systems vis-a-vis systems of human language.