Moving Target Defense (MTD) assumes its pool of candidate configurations is safe to cycle among, i.e. latency and other observables do not trivially fingerprint the active choice, but this assumption has not been quantified at the pool level. We formalize this pool-safety problem as finding the largest ε-close subset of the Cartesian product of per-component implementation choices, reducing pairwise indistinguishability under an additive utility model to a densest-window query over a sum-set. We give four algorithms spanning the scalability spectrum – full enumeration, meet-in-the-middle, FFT convolution, and Monte Carlo sampling – covering configuration spaces from tens to 10^38. We then measure the anonymity gap end-to-end on two production cloud case studies, and find that a component's latency differences do not survive deployment unchanged: a four-runtime serverless rotation, where nothing else masks the interpreter, collapses from four-way to three-way anonymity against a VPC-adjacent adversary, while a 27-configuration three-tier stack, where the same interpreter differences are instead absorbed by a shared 8 ms database round-trip, delivers nine-way effective anonymity. The framework and the two case studies together suggest a diagnostic for MTD design: rotating a component adds anonymity only if the latency differences among its variants are too small for the adversary to identify.

Inference-time decoding methods improve LLM reasoning by exploring multiple candidate trajectories, yet treat each trajectory as atomic: either retaining it whole or discarding it irreversibly. This wastes computation on partially promising candidates whose high-quality prefixes are abandoned alongside degraded suffixes. We introduce Selective Regenerative Decoding (SRD), which routes each candidate to discard, keep, or refine only the degraded portion of the suffix while preserving the useful prefix of borderline candidates, without requiring a larger target model. Under mild assumptions, SRD achieves a provable 1.28-to-1.36-fold gain in sample efficiency over rejection sampling with strictly higher expected trajectory quality, with the gain growing as the candidate pool grows. Across MATH500, GPQA Diamond, HotpotQA, and AlpacaEval with multiple generation-reward model pairs, SRD matches Best-of-N accuracy with substantially fewer generated tokens and outperforms speculative rejection in low-compute regimes. By enabling segment-level intervention rather than whole-trajectory selection, SRD opens a previously underexplored region of the accuracy-compute tradeoff for inference-time reasoning.

In scenarios involving robots generating and executing plans, conflicts can arise between cost-effective robot execution and meeting human expectations for safe behavior. When humans supervise robots, their accountability increases, especially when robot behavior deviates from expectations. To address this, robots may choose a highly constrained plan when monitored and a more optimal one when unobserved. While this behavior is not driven by human-like motives, it stems from robots accommodating diverse supervisors. To optimize monitoring costs while ensuring safety, we model this interaction in a trust-based game-theoretic framework. However, pure-strategy Nash equilibrium often fails to exist in this model. To address this, we introduce the concept of a trust boundary within the mixed strategy space, aiding in the discovery of optimal monitoring strategies. Human studies demonstrate the necessity of optimal strategies and the benefits of our suggested approaches.

The improved competence of generative models can help building multi-modal virtual assistants that leverage modalities beyond language. By observing humans performing multi-step tasks, one can build assistants that have situational awareness of actions and tasks being performed, enabling them to cater assistance based on this understanding. In this paper, we develop a Context-aware Instructional Task Assistant with Multi-modal Large Language Models (InsTALL) that leverages an online visual stream (e.g. a user's screen share or video recording) and responds in real-time to user queries related to the task at hand. To enable useful assistance, InsTALL 1) trains a multi-modal model on task videos and paired textual data, and 2) automatically extracts task graph from video data and leverages it at training and inference time. We show InsTALL achieves state-of-the-art performance across proposed sub-tasks considered for multimodal activity understanding – task recognition (TR), action recognition (AR), next action prediction (AP), and plan prediction (PP) – and outperforms existing baselines on two novel sub-tasks related to automatic error identification.

Large Language Models (LLMs) are nowadays expected to generate content aligned with human preferences. Current work focuses on alignment at model training time, through techniques such as Reinforcement Learning with Human Feedback (RLHF). However, it is unclear if such methods are an effective choice to teach alignment objectives to the model. First, the inability to incorporate multiple, custom rewards and reliance on a model developer’s view of universal and static principles are key limitations. Second, the reliability of such approaches is also questionable (e.g. susceptibility to jailbreaking even after safety training). To address these issues, we propose DeAL, a framework that allows the user to customize reward functions and enables Decoding-time Alignment of LLMs (DeAL). At its core, we view decoding as a heuristic-guided search process and facilitate the use of a wide variety of alignment objectives. Our experiments with programmatic constraints such as keyword and length constraints, and abstract objectives such as harmlessness and helpfulness, show that we can DeAL with fine-grained trade-offs and improve adherence to alignment objectives. Lastly, we demonstrate that DeAL is largely complementary to existing alignment strategies, and can be effectively paired with RLHF and prompting techniques to achieve better alignment.

Planning is a crucial task for agents in task oriented dialogs (TODs). Human agents typically resolve user issues by following predefined workflows, decomposing workflow steps into actionable items, and performing actions by executing APIs in order; all of which require reasoning and planning. With the recent advances in LLMs, there have been increasing attempts to use them for task planning and API usage. However, the faithfulness of the plans to predefined workflows and API dependencies, is not guaranteed with LLMs. Moreover, workflows in real life are often custom-defined and prone to changes; hence, adaptation is desirable. To study this, we propose the problem of faithful planning in TODs that needs to resolve user intents by following predefined flows and preserving API dependencies. To solve this problem, we propose FLAP, a Flow-Adhering Planning algorithm based on constrained decoding with lookahead heuristic for LLMs. Our algorithm alleviates the need for finetuning LLMs using domain specific (plan/dependency) data, enables quick adaptation to predefined flows, and outperforms other decoding and prompting-based baselines. Further, our algorithm empowers smaller LLMs (7B) to perform at par larger LLMs (30B-40B).

Task allocation is an important problem in multi-agent systems. It becomes more challenging when the team-members are humans with imperfect knowledge about their teammates' costs and the overall performance metric. In this paper, we propose a centralized Artificial Intelligence Task Allocation (AITA) that simulates a negotiation and produces a negotiation-aware explicable task allocation. If a team-member is unhappy with the proposed allocation, we allow them to question the proposed allocation using a counterfactual. By using parts of the simulated negotiation, we are able to provide contrastive explanations that provide minimum information about other's cost to refute their foil. With human studies, we show that (1) the allocation proposed using our method appears fair to the majority, and (2) when a counterfactual is raised, explanations generated are easy to comprehend and convincing. Finally, we empirically study the effect of different kinds of incompleteness on the explanation-length and find that underestimation of a teammate's costs often increases it.

Direct alignment algorithms (DAAs), such as direct preference optimization (DPO), have become popular alternatives for Reinforcement Learning from Human Feedback (RLHF) due to their simplicity, efficiency, and stability. However, the preferences used in DAAs are usually collected before the alignment training begins and remain unchanged (off-policy). This can lead to two problems where the policy model (1) picks up on spurious correlations in the dataset (as opposed to learning the intended alignment expressed in the human preference labels), and (2) overfits to feedback on off-policy trajectories that have less likelihood of being generated by an updated policy model. To address these issues, we introduce Self-Reviewing and Alignment (SeRA), a cost-efficient and effective method that can be readily combined with existing DAAs. SeRA comprises of two components: (1) sample selection using implicit reward margins, which helps alleviate over-fitting to some undesired features, and (2) preference bootstrapping using implicit rewards to augment preference data with updated policy models in a cost-efficient manner. Extensive experimentation, including some on instruction-following tasks, demonstrate the effectiveness and generality of SeRA in training LLMs on offline preference datasets with DAAs.

Conversational systems often rely on embedding models for intent classification and intent clustering tasks. The advent of Large Language Models (LLMs), which enable instructional embeddings allowing one to adjust semantics over the embedding space using prompts, are being viewed as a panacea for these downstream conversational tasks. However, traditional evaluation benchmarks rely solely on task metrics that don't particularly measure gaps related to semantic understanding. Thus, we propose an intent semantic toolkit that gives a more holistic view of intent embedding models by considering three tasks -- (1) intent classification, (2) intent clustering, and (3) a novel triplet task. The triplet task gauges the model's understanding of two semantic concepts paramount in real-world conversational systems -- negation and implicature. We observe that current embedding models fare poorly in semantic understanding of these concepts. To address this, we propose a pre-training approach to improve the embedding model by leveraging augmentation with data generated by an auto-regressive model and a contrastive loss term. Our approach improves the semantic understanding of the intent embedding model on the aforementioned linguistic dimensions while slightly effecting their performance on downstream task metrics.

Multi-objective alignment from human feedback (MOAHF) in large language models (LLMs) is a challenging problem as human preferences are complex, multifaceted, and often conflicting. Recent works on MOAHF considered a-priori multi-objective optimization (MOO), where human preferences are known at training or inference time. In contrast, when human preferences are unknown or difficult to quantify, a natural approach is to cover the Pareto front by multiple diverse solutions. We propose an algorithm HaM for learning diverse LLM policies that maximizes their hypervolume. This is the first application of a-posteriori MOO to MOAHF. HaM is computationally and space efficient, and empirically superior across objectives such as harmlessness, helpfulness, humor, faithfulness, and hallucination, on various datasets.

Advances in neural modeling have achieved state-of-the-art (SOTA) results on public natural language processing (NLP) benchmarks, at times surpassing human performance. However, there is a gap between public benchmarks and real-world applications where noise, such as typographical or grammatical mistakes, is abundant and can result in degraded performance. Unfortunately, works which evaluate the robustness of neural models on noisy data and propose improvements, are limited to the English language. Upon analyzing noise in different languages, we observe that noise types vary greatly across languages. Thus, existing investigations do not generalize trivially to multilingual settings. To benchmark the performance of pretrained multilingual language models, we construct noisy datasets covering five languages and four NLP tasks and observe a clear gap in the performance between clean and noisy data in the zero-shot cross-lingual setting. After investigating several ways to boost the robustness of multilingual models in this setting, we propose Robust Contrastive Pretraining (RCP). RCP combines data augmentation with a contrastive loss term at the pretraining stage and achieves large improvements on noisy (and original test data) across two sentence-level (+3.2%) and two sequence-labeling (+10 F1-score) multilingual classification tasks.

In this work, we design an Artificially Intelligent Task Allocator (AITA) that proposes a task allocation for multi-agent systems especially with humans. A key property of this allocation is that when an agent with imperfect knowledge (about their teammate's costs and/or the team's performance metric) questions the allocation by contesting with a counterfactual, a contrastive explanation is provided to answer their challenge. For this, we consider a negotiation process that produces a negotiation-aware task allocation and, in turn, leverages a negotiation tree to provide a contrastive explanation. With human subject studies, we show that the proposed allocation indeed appears fair to a majority of participants, and the explanations generated are easy to comprehend and convincing.

In executable task-oriented semantic parsing, the system aims to translate users' utterances in natural language to machine-interpretable programs (API calls) that can be executed according to pre-defined API specifications. With the popularity of Large Language Models (LLMs), in-context learning offers a strong baseline for such scenarios, especially in data-limited regimes. However, LLMs are known to hallucinate and therefore pose a formidable challenge in constraining generated content. Thus, it remains uncertain if LLMs can effectively perform task-oriented utterance-to-API generation where respecting API's structural and task-specific constraints is crucial. In this work, we seek to measure, analyze and mitigate such constraints violations. First, we identify the categories of various constraints in obtaining API-semantics from task-oriented utterances, and define fine-grained metrics that complement traditional ones. Second, we leverage these metrics to conduct a detailed error analysis of constraints violations seen in state-of-the-art LLMs, which motivates us to investigate two mitigation strategies: Semantic-Retrieval of Demonstrations (SRD) and API-aware Constrained Decoding (API-CD). Our experiments show that these strategies are effective at reducing constraints violations and improving the quality of the generated API calls, but require careful consideration given their implementation complexity and latency.

The use of multilingual language models for tasks in low and high-resource languages has been a success story in deep learning. In recent times, Arabic has been receiving widespread attention on account of its dialectal variance. While prior research studies have tried to adapt these multilingual models for dialectal variants of Arabic, it still remains a challenging problem owing to the lack of sufficient monolingual dialectal data and parallel translation data of such dialectal variants. It remains an open problem on whether the limited dialectical data can be used to improve the models trained in Arabic on its dialectal variants. First, we show that multilingual-BERT (mBERT) incrementally pretrained on Arabic monolingual data takes less training time and yields comparable accuracy when compared to our custom monolingual Arabic model and beat existing models (by an avg metric of +$6.41$). We then explore two continual pre-training methods -- (1) using small amounts of dialectical data for continual finetuning and (2) parallel Arabic to English data and a Translation Language Modeling loss function. We show that both approaches help improve performance on dialectal classification tasks ($+4.64$ avg. gain) when used on monolingual models.

In this paper, we show that popular Generative Adversarial Network (GAN) variants exacerbate biases along the axes of gender and skin tone in the generated data. The use of synthetic data generated by GANs is widely used for a variety of tasks ranging from data augmentation to stylizing images. While practitioners celebrate this method as an economical way to obtain synthetic data to train data-hungry machine learning models or provide new features to users of mobile applications, it is unclear whether they recognize the perils of such techniques when applied to real world datasets biased along latent dimensions. Although one expects GANs to replicate the distribution of the original data, in real-world settings with limited data and finite network capacity, GANs suffer from mode collapse. First, we show readily-accessible GAN variants such as DCGANs ‘imagine’ faces of synthetic engineering professors that have masculine facial features and fair skin tones. When using popular GAN architectures that attempt to address mode-collapse, we observe that these variants either provide a false sense of security or suffer from other inherent limitations due to their design choice. Second, we show that a conditional GAN variant transforms input images of female and nonwhite faces to have more masculine features and lighter skin when asked to generate faces of engineering professors. Worse yet, prevalent filters on Snapchat end up consistently lightening the skin tones in people of color when trying to make face images appear more feminine. Thus, our study is meant to serve as a cautionary tale for practitioners and educate them about the side-effect of bias amplification when applying GAN-based techniques.

The field of cybersecurity has mostly been a cat-and-mouse game with the discovery of new attacks leading the way. To take away an attacker's advantage of reconnaissance, researchers have proposed proactive defense methods such as Moving Target Defense (MTD). To find good movement strategies, researchers have modeled MTD as leader-follower games between the defender and a cyber-adversary. We argue that existing models are inadequate in sequential settings when there is incomplete information about rational adversary and yield sub-optimal movement strategies. Further, while there exists an array of work on learning defense policies in sequential settings for cyber-security, they are either unpopular due to scalability issues arising out of incomplete information or tend to ignore the strategic nature of the adversary simplifying the scenario to use single-agent reinforcement learning techniques. To address these concerns, we propose (1) a unifying game-theoretic model, called the Bayesian Stackelberg Markov Games (BSMGs), that can model uncertainty over attacker types and the nuances of an MTD system and (2) a Bayesian Strong Stackelberg Q-learning (BSS-Q) approach that can, via interaction, learn the optimal movement policy for BSMGs within a reasonable time. We situate BSMGs in the landscape of incomplete-information Markov games and characterize the notion of Strong Stackelberg Equilibrium (SSE) in them. We show that our learning approach converges to an SSE of a BSMG and then highlight that the learned movement policy (1) improves the state-of-the-art in MTD for web-application security and (2) converges to an optimal policy in MTD domains with incomplete information about adversaries even when prior information about rewards and transitions is absent.

Determined cyber adversaries often strategize their attacks by carefully selecting high-value target machines that host insecure (e.g., unpatched) legacy software. In this paper, we propose a moving-target approach to thwart and countersurveil such adversaries, wherein live (non-decoy) enterprise software services are automatically modified to deceptively emulate vulnerable legacy versions that entice attackers. A game-theoretic framework chooses which emulated software stacks, versions, configurations, and vulnerabilities yield the best defensive payoffs and most useful threat data given a specific attack model. The results show that effective movement strategies can be computed to account for pragmatic aspects of deception, such as the utility of various intelligence-gathering actions, impact of vulnerabilities, performance costs of patch deployment, complexity of exploits, and attacker profile.

Intent Classification (IC) and Slot Labeling (SL) models, which form the basis of dialogue systems, often encounter noisy data in real-word environments. In this work, we investigate how robust IC/SL models are to noisy data. We collect and publicly release a test-suite for seven common noise types found in production human-to-bot conversations (abbreviations, casing, misspellings, morphological variants, paraphrases, punctuation and synonyms). On this test-suite, we show that common noise types substantially degrade the IC accuracy and SL F1 performance of state-of-the-art BERT-based IC/SL models. By leveraging cross-noise robustness transfer -- training on one noise type to improve robustness on another noise type -- we design aggregate data-augmentation approaches that increase the model performance across all seven noise types by +10.8% for IC accuracy and +15 points for SL F1 on average. To the best of our knowledge, this is the first work to present a single IC/SL model that is robust to a wide range of noise phenomena.

Decision support systems seek to enable informed decision-making. In the recent years, automated planning techniques have been leveraged to empower such systems to better aid the human-in-the-loop. The central idea for such decision support systems is to augment the capabilities of the human-in-the-loop with automated planning techniques and enhance the quality of decision-making. In addition to providing planning support, effective decision support systems must be able to provide intuitive explanations based on specific user queries for proposed decisions to its end users. Using this as motivation, we present our decision support system RADAR-X that showcases the ability to engage the user in an interactive explanatory dialogue by first enabling them to specify an alternative to a proposed decision (which we refer to as foils), and then providing contrastive explanations to these user-specified foils which helps the user understand why a specific plan was chosen over the alternative (or foil). Furthermore, the system uses this dialogue to elicit the user's latent preferences and provides revised plan suggestions through three different interaction strategies.

Deep Neural Networks are often brittle on image classification tasks and known to misclassify inputs. While these misclassifications may be inevitable, all failure modes cannot be considered equal. Certain misclassifications (eg. classifying the image of a dog to an airplane) can create surprise and result in the loss of human trust in the system. Even worse, certain errors (eg. a person misclassified as a primate) can have societal impacts. Thus, in this work, we aim to reduce inexplicable errors. To address this challenge, we first discuss how to obtain the class-level semantics that captures the human's expectation ($M^h$) regarding which classes are semantically close vs. ones that are far away. We show that for data-sets like CIFAR-10 and CIFAR-100, class-level semantics can be obtained by leveraging human subject studies (significantly inexpensive compared to existing works) and, whenever possible, by utilizing publicly available human-curated knowledge. Second, we propose the use of Weighted Loss Functions to penalize misclassifications by the weight of their inexplicability. Finally, we show that training (or even fine-tuning) existing classifiers with the two proposed methods lead to Deep Neural Networks that have (1) comparable top-1 accuracy, an important metric in operational contexts, (2) more explicable failure modes and (3) require significantly less cost in teams of additional human labels compared to existing work.