Catherine Chen

Catherine Chen, Zejiang Shen, Dan Klein et al. · berkeley, mit

Recent work has shown that infusing layout features into language models (LMs) improves processing of visually-rich documents such as scientific papers. Layout-infused LMs are often evaluated on documents with familiar layout features (e.g., papers from the same publisher), but in practice models encounter documents with unfamiliar distributions of layout features, such as new combinations of text sizes and styles, or new spatial configurations of textual elements. In this work we test whether layout-infused LMs are robust to layout distribution shifts. As a case study we use the task of scientific document structure recovery, segmenting a scientific paper into its structural categories (e.g., "title", "caption", "reference"). To emulate distribution shifts that occur in practice we re-partition the GROTOAP2 dataset. We find that under layout distribution shifts model performance degrades by up to 20 F1. Simple training strategies, such as increasing training diversity, can reduce this degradation by over 35% relative F1; however, models fail to reach in-distribution performance in any tested out-of-distribution conditions. This work highlights the need to consider layout distribution shifts during model evaluation, and presents a methodology for conducting such evaluations.

Boyi Li, Rodolfo Corona, Karttikeya Mangalam et al. · berkeley

Are multimodal inputs necessary for grammar induction? Recent work has shown that multimodal training inputs can improve grammar induction. However, these improvements are based on comparisons to weak text-only baselines that were trained on relatively little textual data. To determine whether multimodal inputs are needed in regimes with large amounts of textual training data, we design a stronger text-only baseline, which we refer to as LC-PCFG. LC-PCFG is a C-PFCG that incorporates em-beddings from text-only large language models (LLMs). We use a fixed grammar family to directly compare LC-PCFG to various multi-modal grammar induction methods. We compare performance on four benchmark datasets. LC-PCFG provides an up to 17% relative improvement in Corpus-F1 compared to state-of-the-art multimodal grammar induction methods. LC-PCFG is also more computationally efficient, providing an up to 85% reduction in parameter count and 8.8x reduction in training time compared to multimodal approaches. These results suggest that multimodal inputs may not be necessary for grammar induction, and emphasize the importance of strong vision-free baselines for evaluating the benefit of multimodal approaches.

Catherine Chen, Jingyan Shen, Zhun Deng et al.

We present an online, distribution-free framework for controlling the Conditional Value-at-Risk (CVaR), extending conformal tail risk control to non-stationary and adversarial environments. Unlike classical risk control methods, which rely on stationarity or linearity of expectation, our approach provides provable safety guarantees for a nonlinear tail risk functional under arbitrary data-generating processes that may drift or shift strategically over time. By leveraging deep connections between conformal tail risk control, online learning, and the variational representation of CVaR introduced by Rockafellar and Uryasev, we develop a novel procedure for online CVaR control with adversarial regret guarantees. The proposed method operates without assumptions on the underlying data-generating process, making it broadly applicable in modern high-stakes deployment settings. We prove that the realized empirical CVaR is asymptotically controlled at the target level, and that the resulting control is asymptotically tight up to a finite-sample conservatism gap. We demonstrate the effectiveness of our approach on portfolio risk management and toxicity mitigation for Large Language Models (LLMs), where rare but catastrophic failures dominate system risk.

William Rudman, Catherine Chen, Carsten Eickhoff

Representations from large language models (LLMs) are known to be dominated by a small subset of dimensions with exceedingly high variance. Previous works have argued that although ablating these outlier dimensions in LLM representations hurts downstream performance, outlier dimensions are detrimental to the representational quality of embeddings. In this study, we investigate how fine-tuning impacts outlier dimensions and show that 1) outlier dimensions that occur in pre-training persist in fine-tuned models and 2) a single outlier dimension can complete downstream tasks with a minimal error rate. Our results suggest that outlier dimensions can encode crucial task-specific knowledge and that the value of a representation in a single outlier dimension drives downstream model decisions.

Mainak Kundu, Catherine Chen, Rifatul Islam et al.

Human activity recognition (HAR) has become a key component of intelligent systems for healthcare monitoring, assistive living, smart environments, and human-computer interaction. Although deep learning has substantially improved HAR performance on multivariate sensor data, the resulting models often remain opaque, limiting trust, reliability, and real-world deployment. Explainable artificial intelligence (XAI) has therefore emerged as a critical direction for making HAR systems more transparent and human-centered. This paper presents a comprehensive review of explainable HAR methods across wearable, ambient, physiological, and multimodal sensing settings. We introduce a unified perspective that separates conceptual dimensions of explainability from algorithmic explanation mechanisms, reducing ambiguities in prior surveys. Building on this distinction, we present a mechanism-centric taxonomy of XAI-HAR methods covering major explanation paradigms. The review examines how these methods address the temporal, multimodal, and semantic complexities of HAR, and summarize their interpretability objectives, explanation targets, and limitations. In addition, we discuss current evaluation practices, highlight key challenges in achieving reliable and deployable XAI-HAR, and outline directions toward trustworthy activity recognition systems that better support human understanding and decision-making.

Lincan Li, Jiaqi Li, Catherine Chen et al.

In recent years, large language models (LLMs) have been widely adopted in political science tasks such as election prediction, sentiment analysis, policy impact assessment, and misinformation detection. Meanwhile, the need to systematically understand how LLMs can further revolutionize the field also becomes urgent. In this work, we--a multidisciplinary team of researchers spanning computer science and political science--present the first principled framework termed Political-LLM to advance the comprehensive understanding of integrating LLMs into computational political science. Specifically, we first introduce a fundamental taxonomy classifying the existing explorations into two perspectives: political science and computational methodologies. In particular, from the political science perspective, we highlight the role of LLMs in automating predictive and generative tasks, simulating behavior dynamics, and improving causal inference through tools like counterfactual generation; from a computational perspective, we introduce advancements in data preparation, fine-tuning, and evaluation methods for LLMs that are tailored to political contexts. We identify key challenges and future directions, emphasizing the development of domain-specific datasets, addressing issues of bias and fairness, incorporating human expertise, and redefining evaluation criteria to align with the unique requirements of computational political science. Political-LLM seeks to serve as a guidebook for researchers to foster an informed, ethical, and impactful use of Artificial Intelligence in political science. Our online resource is available at: http://political-llm.org/.

Meng Lu, Catherine Chen, Carsten Eickhoff

Neural Ranking Models (NRMs) have rapidly advanced state-of-the-art performance on information retrieval tasks. In this work, we investigate a Cross-Encoder variant of MiniLM to determine which relevance features it computes and where they are stored. We find that it employs a semantic variant of the traditional BM25 in an interpretable manner, featuring localized components: (1) Transformer attention heads that compute soft term frequency while controlling for term saturation and document length effects, and (2) a low-rank component of its embedding matrix that encodes inverse document frequency information for the vocabulary. This suggests that the Cross-Encoder uses the same fundamental mechanisms as BM25, but further leverages their capacity to capture semantics for improved retrieval performance. The granular understanding lays the groundwork for model editing to enhance model transparency, addressing safety concerns, and improving scalability in training and real-world applications.

Catherine Chen, Kevin Lin, Dan Klein

We present a method for constructing taxonomic trees (e.g., WordNet) using pretrained language models. Our approach is composed of two modules, one that predicts parenthood relations and another that reconciles those predictions into trees. The parenthood prediction module produces likelihood scores for each potential parent-child pair, creating a graph of parent-child relation scores. The tree reconciliation module treats the task as a graph optimization problem and outputs the maximum spanning tree of this graph. We train our model on subtrees sampled from WordNet, and test on non-overlapping WordNet subtrees. We show that incorporating web-retrieved glosses can further improve performance. On the task of constructing subtrees of English WordNet, the model achieves 66.7 ancestor F1, a 20.0% relative increase over the previous best published result on this task. In addition, we convert the original English dataset into nine other languages using Open Multilingual WordNet and extend our results across these languages.

Catherine Chen, Qihong Lu, Andre Beukers et al.

Through specific experiences, humans learn relationships underlying the structure of events in the world. Schema theory suggests that we organize this information in mental frameworks called "schemata," which represent our knowledge of the structure of the world. Generalizing knowledge of structural relationships to new situations requires role-filler binding, the ability to associate specific "fillers" with abstract "roles." For instance, when we hear the sentence "Alice ordered a tea from Bob," the role-filler bindings "Alice:customer," "tea:drink," and "Bob:barista" allow us to understand and make inferences about the sentence. We can perform these bindings for arbitrary fillers -- we understand this sentence even if we have never heard the names "Alice," "tea," or "Bob" before. In this work, we define a model as capable of performing role-filler binding if it can recall arbitrary fillers corresponding to a specified role, even when these pairings violate correlations seen during training. Previous work found that models can learn this ability when explicitly told what the roles and fillers are, or when given fillers seen during training. We show that networks with external memory can learn these relationships with fillers not seen during training and without explicitly labeled role-filler bindings, and show that analyses inspired by neural decoding can provide a means of understanding what the networks have learned.