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.

Large language models have emerged as powerful tools for general intelligence, showcasing advanced natural language processing capabilities that find applications across diverse domains. Despite their impressive performance, recent studies have highlighted the potential for significant enhancements in LLMs' task-specific performance through fine-tuning strategies like Reinforcement Learning with Human Feedback (RLHF), supervised fine-tuning (SFT), and Direct Preference Optimization (DPO) method. However, previous works have shown that while fine-tuning offers significant performance gains, it also leads to challenges such as catastrophic forgetting and privacy and safety risks. To this end, there has been little to no work in understanding the impact of fine-tuning on the reasoning capabilities of LLMs. Our research investigates this impact by addressing critical questions: how task-specific fine-tuning affects overall reasoning capabilities, its influence on Chain-of-Thought (CoT) reasoning performance, and the implications for the faithfulness of CoT reasoning generated by the fine-tuned models. Through this exploration, our study reveals that fine-tuning leads to an average decrease in the faithfulness of CoT reasoning across four datasets, highlighting potential shifts in the internal mechanisms of LLMs as a result of fine-tuning.

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.

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.

Enterprises often own large collections of structured data in the form of large databases or an enterprise data lake. Such data collections come with limited metadata and strict access policies that could limit access to the data contents and, therefore, limit the application of classic retrieval and analysis solutions. As a result, there is a need for solutions that can effectively utilize the available metadata. In this paper, we study the problem of matching table metadata to a business glossary containing data labels and descriptions. The resulting matching enables the use of an available or curated business glossary for retrieval and analysis without or before requesting access to the data contents. One solution to this problem is to use manually-defined rules or similarity measures on column names and glossary descriptions (or their vector embeddings) to find the closest match. However, such approaches need to be tuned through manual labeling and cannot handle many business glossaries that contain a combination of simple as well as complex and long descriptions. In this work, we leverage the power of large language models (LLMs) to design generic matching methods that do not require manual tuning and can identify complex relations between column names and glossaries. We propose methods that utilize LLMs in two ways: a) by generating additional context for column names that can aid with matching b) by using LLMs to directly infer if there is a relation between column names and glossary descriptions. Our preliminary experimental results show the effectiveness of our proposed methods.

Off-policy Evaluation (OPE) methods are a crucial tool for evaluating policies in high-stakes domains such as healthcare, where exploration is often infeasible, unethical, or expensive. However, the extent to which such methods can be trusted under adversarial threats to data quality is largely unexplored. In this work, we make the first attempt at investigating the sensitivity of OPE methods to marginal adversarial perturbations to the data. We design a generic data poisoning attack framework leveraging influence functions from robust statistics to carefully construct perturbations that maximize error in the policy value estimates. We carry out extensive experimentation with multiple healthcare and control datasets. Our results demonstrate that many existing OPE methods are highly prone to generating value estimates with large errors when subject to data poisoning attacks, even for small adversarial perturbations. These findings question the reliability of policy values derived using OPE methods and motivate the need for developing OPE methods that are statistically robust to train-time data poisoning attacks.

In real-world sequential decision problems, exploration is expensive, and the risk of expert decision policies must be evaluated from limited data. In this setting, Monte Carlo (MC) risk estimators are typically used to estimate the risk of decision policies. Unfortunately, while these estimators have the desired low bias property, they often suffer from large variance. In this paper, we consider the problem of minimizing the asymptotic mean squared error and hence variance of MC risk estimators. We show that by carefully choosing the data sampling policy (behavior policy), we can obtain low variance estimates of the risk of any given decision policy.