We introduce a zero-shot merging framework for large language models (LLMs) that consolidates specialized domain experts into a single model without any further training.Our core contribution lies in leveraging relative task vectors-difference representations encoding each expert's unique traits with respect to a shared base model-to guide a principled and efficient merging process.By dissecting parameters into common dimensions (averaged across experts) and complementary dimensions (unique to each expert), we strike an optimal balance between generalization and specialization.We further devise a compression mechanism for the complementary parameters, retaining only principal components and scalar multipliers per expert, thereby minimizing overhead.A dynamic router then selects the most relevant domain at inference, ensuring that domain-specific precision is preserved.Experiments on code generation, mathematical reasoning, medical question answering, and instruction-following benchmarks confirm the versatility and effectiveness of our approach.Altogether, this framework enables truly adaptive and scalable LLMs that seamlessly integrate specialized knowledge for improved zero-shot performance.

Ranking systems based on pairwise comparisons are fundamental in decision-making applications, yet fairness concerns remain largely unaddressed, particularly in active ranking frameworks. We propose a novel fairness-aware active ranking approach that adaptively queries pairwise preferences to construct rankings that are both probably approximately correct (PAC) and fair. We propose a flexible cascaded norm-based objective that balances error distribution within and across socially salient groups, providing a unified framework to mitigate systemic disparities. Adopting our objective function allows us to explore fundamental fairness concepts like equal or proportionate errors within a unified framework. We develop both group-blind and group-aware algorithms and derive their sample complexity bounds. Empirical evaluations on real-world datasets, including COMPAS and German Credit, demonstrate the efficiency of our approach, reducing sample complexity while achieving fairer rankings. Our findings offer theoretical insights and practical methods to enhance fairness in active ranking systems, improving their reliability in hiring, recommendations, and other applications.

Learning-to-rank (LTR) models rank items based on specific features, aiming to maximize ranking utility by prioritizing highly relevant items. However, optimizing only for ranking utility can lead to representational harm and may fail to address implicit bias in relevance scores. Prior studies introduced algorithms to train stochastic ranking models, such as the Plackett-Luce ranking model, that maximize expected ranking utility while achieving fairness in expectation (ex-ante fairness). Still, every sampled ranking may not satisfy group fairness (ex-post fairness). Post-processing methods ensure ex-post fairness; however, the LTR model lacks awareness of this step, creating a mismatch between the objective function the LTR model optimizes and the one it is supposed to optimize. In this paper, we first propose a novel objective where the relevance (or the expected ranking utility) is computed over only those rankings that satisfy given representation constraints for groups of items. We call this the ex-post fair relevance. We then give a framework for training Group-Fair LTR models to maximize our proposed ranking objective.

Underserved communities face critical health challenges due to lack of access to timely and reliable information. Nongovernmental organizations are leveraging the widespread use of cellphones to combat these healthcare challenges and spread preventative awareness. The health workers at these organizations reach out individually to beneficiaries; however such programs still suffer from declining engagement. We have deployed SAHELI, a system to efficiently utilize the limited availability of health workers for improving maternal and child health in India. SAHELI uses the Restless Multiarmed Bandit (RMAB) framework to identify beneficiaries for outreach. It is the first deployed application for RMABs in public health, and is already in continuous use by our partner NGO, ARMMAN. We have already reached ~100K beneficiaries with SAHELI, and are on track to serve 1 million beneficiaries by the end of 2023. This scale and impact has been achieved through multiple innovations in the RMAB model and its development, in preparation of real world data, and in deployment practices; and through careful consideration of responsible AI practices. Specifically, in this paper, we describe our approach to learn from past data to improve the performance of SAHELI’s RMAB model, the real-world challenges faced during deployment and adoption of SAHELI, and the end-to-end pipeline.

Rankings on online platforms help their end-users find the relevant information—people, news, media, and products—quickly. Fair ranking tasks, which ask to rank a set of items to maximize utility subject to satisfying group-fairness constraints, have gained significant interest in the Algorithmic Fairness, Information Retrieval, and Machine Learning literature. Recent works, however, identify uncertainty in the utilities of items as a primary cause of unfairness and propose introducing randomness in the output. This randomness is carefully chosen to guarantee an adequate representation of each item (while accounting for the uncertainty). However, due to this randomness, the output rankings may violate group fairness constraints. We give an efficient algorithm that samples rankings from an individually-fair distribution while ensuring that every output ranking is group fair. The expected utility of the output ranking is at least α times the utility of the optimal fair solution. Here, α depends on the utilities, position-discounts, and constraints—it approaches 1 as the range of utilities or the position-discounts shrinks, or when utilities satisfy distributional assumptions. Empirically, we observe that our algorithm achieves individual and group fairness and that Pareto dominates the state-of-the-art baselines.

Randomized rankings have been of recent interest to achieve ex-ante fairer exposure and better robustness than deterministic rankings. We propose a set of natural axioms for randomized group-fair rankings and prove that there exists a unique distribution D that satisfies our axioms and is supported only over ex-post group-fair rankings, i.e., rankings that satisfy given lower and upper bounds on group-wise representation in the top-k ranks. Our problem formulation works even when there is implicit bias, incomplete relevance information, or only ordinal ranking is available instead of relevance scores or utility values. We propose two algorithms to sample a random group-fair ranking from the distribution D mentioned above. Our first dynamic programming-based algorithm samples ex-post group-fair rankings uniformly at random in time O(k^2 ell), where "ell" is the number of groups. Our second random walk-based algorithm samples ex-post group-fair rankings from a distribution epsilon-close to D in total variation distance and has expected running time O*(k^2 ell^2), when there is a sufficient gap between the given upper and lower bounds on the group-wise representation. The former does exact sampling, but the latter runs significantly faster on real-world data sets for larger values of k. We give empirical evidence that our algorithms compare favorably against recent baselines for fairness and ranking utility on real-world data sets.

Bias in ranking systems, especially among the top ranks, can worsen social and economic inequalities, polarize opinions, and reinforce stereotypes. On the other hand, a bias correction for minority groups can cause more harm if perceived as favoring group-fair outcomes over meritocracy. Most group-fair ranking algorithms post-process a given ranking and output a group-fair ranking. In this paper, we formulate the problem of underranking in group-fair rankings based on how close the group-fair rank of each item is to its original rank, and prove a lower bound on the trade-off achievable for simultaneous underranking and group fairness in ranking. We give a fair ranking algorithm that takes any given ranking and outputs another ranking with simultaneous underranking and group fairness guarantees comparable to the lower bound we prove. Our experimental results confirm the theoretical trade-off between underranking and group fairness, and also show that our algorithm achieves the best of both when compared to the state-of-the-art baselines.

The literature in automated sarcasm detection has mainly focused on lexical-, syntactic- and semantic-level analysis of text. However, a sarcastic sentence can be expressed with contextual presumptions, background and commonsense knowledge. In this paper, we propose a ContextuAl SarCasm DEtector (CASCADE), which adopts a hybrid approach of both content- and context-driven modeling for sarcasm detection in online social media discussions. For the latter, CASCADE aims at extracting contextual information from the discourse of a discussion thread. Also, since the sarcastic nature and form of expression can vary from person to person, CASCADE utilizes user embeddings that encode stylometric and personality features of users. When used along with content-based feature extractors such as convolutional neural networks, we see a significant boost in the classification performance on a large Reddit corpus.