Fine-tuned Large Language Models (LLMs) encode rich task-specific features, but the form of these representations—especially within MLP layers—remains unclear. Empirical inspection of LoRA updates shows that new features concentrate in mid-layer MLPs, yet the scale of these layers obscures meaningful structure. Prior probing suggests that statistical priors may strengthen, split, or vanish across depth, motivating the need to study how neurons work together rather than in isolation. We introduce a mechanistic interpretability framework based on coalitional game theory, where neurons mimic agents in a hedonic game whose preferences capture their synergistic contributions to layer-local computations. Using top-responsive utilities and the PAC-Top-Cover algorithm, we extract stable coalitions of neurons—groups whose joint ablation has non-additive effects—and track their transitions across layers as persistence, splitting, merging, or disappearance. Applied to LLaMA, Mistral, and Pythia rerankers fine-tuned on scalar IR tasks, our method finds coalitions with consistently higher synergy than clustering baselines. By revealing how neurons cooperate to encode features, hedonic coalitions uncover higher-order structure beyond disentanglement and yield computational units that are functionally important, interpretable, and predictive across domains.

Recent advances in large language models (LLMs) have enabled agentic systems for sequential decision-making. Such agents must perceive their environment, reason across multiple time steps, and take actions that optimize long-term objectives. However, existing web agents struggle on complex, long-horizon tasks due to limited in-context memory for tracking history, weak planning abilities, and greedy behaviors that lead to premature termination. To address these challenges, we propose STRUCTUREDAGENT, a hierarchical planning framework with two core components: (1) an online hierarchical planner that uses dynamic AND/OR trees for efficient search and (2) a structured memory module that tracks and maintains candidate solutions to improve constraint satisfaction in information-seeking tasks. The framework also produces interpretable hierarchical plans, enabling easier debugging and facilitating human intervention when needed. Our results on WebVoyager, WebArena, and custom shopping benchmarks show that STRUCTUREDAGENT improves performance on long-horizon web-browsing tasks compared to standard LLM-based agents.

We study a two-sided matching model where one side of the market (hospitals) has combinatorial preferences over the other side (doctors). Specifically, we consider the setting where hospitals have matroid rank valuations over the doctors, and doctors have either ordinal or cardinal unit-demand valuations over the hospitals. While this setting has been extensively studied in the context of one-sided markets, it remains unexplored in the context of two-sided markets. When doctors have ordinal preferences over hospitals, we present simple sequential allocation algorithms that guarantee stability, strategyproofness for doctors, and approximate strategyproofness for hospitals. When doctors have cardinal utilities over hospitals, we present an algorithm that finds a stable allocation maximizing doctor welfare; subject to that, we show how one can maximize either the hospital utilitarian or hospital Nash welfare. Moreover, we show that it is NP-hard to compute stable allocations that approximately maximize hospital Nash welfare.

Numerous works propose post-hoc, model-agnostic explanations for learning to rank, focusing on ordering entities by their relevance to a query through feature attribution methods. However, these attributions often weakly correlate or contradict each other, confusing end users. We adopt an axiomatic game-theoretic approach, popular in the feature attribution community, to identify a set of fundamental axioms that every ranking-based feature attribution method should satisfy. We then introduce Rank-SHAP, extending classical Shapley values to ranking. We evaluate the RankSHAP framework through extensive experiments on two datasets, multiple ranking methods and evaluation metrics. Additionally, a user study confirms RankSHAP's alignment with human intuition. We also perform an axiomatic analysis of existing rank attribution algorithms to determine their compliance with our proposed axioms. Ultimately, our aim is to equip practitioners with a set of axiomatically backed feature attribution methods for studying IR ranking models, that ensure generality as well as consistency.

We study the problem of fairly allocating a set of indivisible goods among agents with matroid rank valuations: Every good provides a marginal value of 0 or 1 when added to a bundle and valuations are submodular. We present a simple algorithmic framework, called General Yankee Swap, that can efficiently compute allocations that maximize any justice criterion (or fairness objective) satisfying some mild assumptions. Along with maximizing a justice criterion, General Yankee Swap is guaranteed to maximize utilitarian social welfare, ensure strategyproofness, and use at most a quadratic number of valuation queries. We show how General Yankee Swap can be used to compute allocations for five different well-studied justice criteria: (a) Prioritized Lorenz dominance, (b) Maximin fairness, (c) Weighted leximin, (d) Max weighted Nash welfare, and (e) Max weighted p-mean welfare. In particular, this framework provides the first polynomial time algorithms to compute weighted leximin, max weighted Nash welfare, and max weighted p-mean welfare allocations for agents with matroid rank valuations. We also extend this framework to the setting of binary chores-items with marginal values-1 or 0-and similarly show that it can be used to maximize any justice criteria satisfying some mild assumptions.

We study the problem of fair allocation of indivisible items when agents have ternary additive valuations -- each agent values each item at some fixed integer values $a$, $b$, or $c$ that are common to all agents. The notions of fairness we consider are max Nash welfare (MNW), when $a$, $b$, and $c$ are non-negative, and max egalitarian welfare (MEW). We show that for any distinct non-negative $a$, $b$, and $c$, maximizing Nash welfare is APX-hard -- i.e., the problem does not admit a PTAS unless P = NP. We also show that for any distinct $a$, $b$, and $c$, maximizing egalitarian welfare is APX-hard except for a few cases when $b = 0$ that admit efficient algorithms. These results make significant progress towards completely characterizing the complexity of computing exact MNW allocations and MEW allocations. En route, we resolve open questions left by prior work regarding the complexity of computing MNW allocations under bivalued valuations, and MEW allocations under ternary mixed manna.

We study fairness in algorithmic recourse, which provides users who receive undesirable outcomes from decision-making systems with explanations on how to achieve their desired results. To produce explanations that emulate real-world settings and preferences, the literature proposes various constraints, e.g., sparsity, on the objective of recourse methods. We argue that these constrained objectives can induce disparate recourse costs between protected groups, especially when groups' recourse objectives differ from the objectives of the methods providing them recourse. We formalize discrimination in algorithmic recourse with respect to legal notions and the cost of recourse. Namely, we introduce a novel notion of induced discrimination that captures the excess cost of recourse induced by having misaligned objectives between groups and methods. Using this definition, we show that recourse objectives can induce discriminatory recourse, independent from the bias of the model being explained. To inhibit discrimination, we extend prior work on welfare-based group-fair supervised learning to measure group welfare under multiple recourse objectives. We then propose a recourse method that maximizes group welfare and empirically show its robust performance over multiple recourse objectives.

We study a hinted heterogeneous multi-agent multi-armed bandits problem (HMA2B), where agents can query low-cost observations (hints) in addition to pulling arms. In this framework, each of the M agents has a unique reward distribution over K arms, and in T rounds, they can observe the reward of the arm they pull only if no other agent pulls that arm. The goal is to maximize the total utility by querying the minimal necessary hints without pulling arms, achieving time-independent regret. We study HMA2B in both centralized and decentralized setups. Our main centralized algorithm, GP-HCLA, which is an extension of HCLA, uses a central decision-maker for arm-pulling and hint queries, achieving O(M^4K) regret with O(MKlog T) adaptive hints. In decentralized setups, we propose two algorithms, HD-ETC and EBHD-ETC, that allow agents to choose actions independently through collision-based communication and query hints uniformly until stopping, yielding O(M^3K^2) regret with O(M^3Klog T) hints, where the former requires knowledge of the minimum gap and the latter does not. Finally, we establish lower bounds to prove the optimality of our results and verify them through numerical simulations.

Universities regularly face the challenging task of assigning classes to thousands of students while considering their preferences, along with course schedules and capacities. Ensuring the effectiveness and fairness of course allocation mechanisms is crucial to guaranteeing student satisfaction and optimizing resource utilization. We approach this problem from an economic perspective, using formal justice criteria to evaluate different algorithmic frameworks. To evaluate our frameworks, we conduct a large scale survey of university students at University of Massachusetts Amherst, collecting over 1,000 student preferences. This is, to our knowledge, the largest publicly available dataset of student preferences. We develop software for generating synthetic student preferences over courses, and implement four allocation algorithms: the serial dictatorship algorithm used by University of Massachusetts Amherst; Round Robin; an Integer Linear Program; and the Yankee Swap algorithm. We propose improvements to the Yankee Swap framework to handle scenarios with item multiplicities. Through experimentation with the Fall 2024 Computer Science course schedule at University of Massachusetts Amherst, we evaluate each algorithm's performance relative to standard justice criteria, providing insights into fair course allocation in large university settings.

The recent criticisms of the robustness of post hoc model approximation explanation methods (like LIME and SHAP) have led to the rise of model-precise abductive explanations. For each data point, abductive explanations provide a minimal subset of features that are sufficient to generate the outcome. While theoretically sound and rigorous, abductive explanations suffer from a major issue -- there can be several valid abductive explanations for the same data point. In such cases, providing a single abductive explanation can be insufficient; on the other hand, providing all valid abductive explanations can be incomprehensible due to their size. In this work, we solve this issue by aggregating the many possible abductive explanations into feature importance scores. We propose three aggregation methods: two based on power indices from cooperative game theory and a third based on a well-known measure of causal strength. We characterize these three methods axiomatically, showing that each of them uniquely satisfies a set of desirable properties. We also evaluate them on multiple datasets and show that these explanations are robust to the attacks that fool SHAP and LIME.

We study fair allocation of constrained resources, where a market designer optimizes overall welfare while maintaining group fairness. In many large-scale settings, utilities are not known in advance, but are instead observed after realizing the allocation. We therefore estimate agent utilities using machine learning. Optimizing over estimates requires trading-off between mean utilities and their predictive variances. We discuss these trade-offs under two paradigms for preference modeling – in the stochastic optimization regime, the market designer has access to a probability distribution over utilities, and in the robust optimization regime they have access to an uncertainty set containing the true utilities with high probability. We discuss utilitarian and egalitarian welfare objectives, and we explore how to optimize for them under stochastic and robust paradigms. We demonstrate the efficacy of our approaches on three publicly available conference reviewer assignment datasets. The approaches presented enable scalable constrained resource allocation under uncertainty for many combinations of objectives and preference models.

We propose a novel data-driven framework for algorithmic recourse that offers users interventions to change their predicted outcome. Existing approaches to compute recourse find a set of points that satisfy some desiderata -- e.g. an intervention in the underlying causal graph, or minimizing a cost function. Satisfying these criteria, however, requires extensive knowledge of the underlying model structure, often an unrealistic amount of information in several domains. We propose a data-driven, computationally efficient approach to computing algorithmic recourse. We do so by suggesting directions in the data manifold that users can take to change their predicted outcome. We present Stepwise Explainable Paths (StEP), an axiomatically justified framework to compute direction-based algorithmic recourse. We offer a thorough empirical and theoretical investigation of StEP. StEP offers provable privacy and robustness guarantees, and outperforms the state-of-the-art on several established recourse desiderata.

In reinforcement learning, robust policies for high-stakes decision-making problems with limited data are usually computed by optimizing the percentile criterion. The percentile criterion is optimized by constructing an uncertainty set that contains the true model with high probability and optimizing the policy for the worst model in the set. Since the percentile criterion is non-convex, constructing these sets itself is challenging. Existing works use Bayesian credible regions as uncertainty sets, but they are often unnecessarily large and result in learning overly conservative policies. To overcome these shortcomings, we propose a novel Value-at-Risk based dynamic programming algorithm to optimize the percentile criterion without explicitly constructing any uncertainty sets. Our theoretical and empirical results show that our algorithm implicitly constructs much smaller uncertainty sets and learns less-conservative robust policies.

This white paper is a response to the "AI Accountability Policy Request for Comments" by the National Telecommunications and Information Administration of the United States. The question numbers for which comments were requested are provided in superscripts at the end of key sentences answering the respective questions. The white paper offers a set of interconnected recommendations for an AI accountability policy.

Fair allocation of indivisible goods is a well-explored problem. Traditionally, research focused on individual fairness - are individual agents satisfied with their allotted share? - and group fairness - are groups of agents treated fairly? In this paper, we explore the coexistence of individual envy-freeness (i-EF) and its group counterpart, group weighted envy-freeness (g-WEF), in the allocation of indivisible goods. We propose several polynomial-time algorithms that provably achieve i-EF and g-WEF simultaneously in various degrees of approximation under three different conditions on the agents' (i) when agents have identical additive valuation functions, i-EFX and i-WEF1 can be achieved simultaneously; (ii) when agents within a group share a common valuation function, an allocation satisfying both i-EF1 and g-WEF1 exists; and (iii) when agents' valuations for goods within a group differ, we show that while maintaining i-EF1, we can achieve a 1/3-approximation to ex-ante g-WEF1. Our results thus provide a first step towards connecting individual and group fairness in the allocation of indivisible goods, in hopes of its useful application to domains requiring the reconciliation of diversity with individual demands.

We study the problem of fairly allocating indivisible goods (positively valued items) and chores (negatively valued items) among agents with decreasing marginal utilities over items. Our focus is on instances where all the agents have simple preferences; specifically, we assume the marginal value of an item can be either -1, 0 or some positive integer c. Under this assumption, we present an efficient algorithm to compute leximin allocations for a broad class of valuation functions we call order-neutral submodular valuations. Order-neutral submodular valuations strictly contain the well-studied class of additive valuations but are a strict subset of the class of submodular valuations. We show that these leximin allocations are Lorenz dominating and approximately proportional. We also show that, under further restriction to additive valuations, these leximin allocations are approximately envy-free and guarantee each agent their maxmin share. We complement this algorithmic result with a lower bound showing that the problem of computing leximin allocations is NP-hard when c is a rational number.

A successful peer review process requires that qualified and interested reviewers are assigned to each paper. Most automated reviewer assignment approaches estimate a real-valued affinity score for each paper-reviewer pair that acts as a proxy for the quality of the match, and then assign reviewers to maximize the sum of affinity scores. Most affinity score estimation methods are inherently noisy: reviewers can only bid on a small number of papers, and textual similarity models and subject-area matching are inherently noisy estimators. Current paper assignment systems are not designed to rigorously handle noise in the peer-review matching market. In this work, we assume paper-reviewer affinity scores are located in or near a high-probability region called an uncertainty set. We maximize the worst-case sum of scores for a reviewer assignment over the uncertainty set. We demonstrate how to robustly maximize the sum of scores across various classes of uncertainty sets, avoiding potentially serious mistakes in assignment. Our general approach can be used to integrate a large variety of paper-reviewer affinity models into reviewer assignment, opening the door to a much more robust peer review process.

We study the problem of allocating indivisible chores among agents with binary supermodular cost functions. In other words, each chore has a marginal cost of $0$ or $1$ and chores exhibit increasing marginal costs (or decreasing marginal utilities). In this note, we combine the techniques of Viswanathan and Zick (2022) and Barman et al. (2023) to present a general framework for fair allocation with this class of valuation functions. Our framework allows us to generalize the results of Barman et al. (2023) and efficiently compute allocations which satisfy weighted notions of fairness like weighted leximin or min weighted $p$-mean malfare for any $p \ge 1$.