Phillip Howard

Xin Su, Tiep Le, Steven Bethard et al.

An important open question in the use of large language models for knowledge-intensive tasks is how to effectively integrate knowledge from three sources: the model's parametric memory, external structured knowledge, and external unstructured knowledge. Most existing prompting methods either rely on one or two of these sources, or require repeatedly invoking large language models to generate similar or identical content. In this work, we overcome these limitations by introducing a novel semi-structured prompting approach that seamlessly integrates the model's parametric memory with unstructured knowledge from text documents and structured knowledge from knowledge graphs. Experimental results on open-domain multi-hop question answering datasets demonstrate that our prompting method significantly surpasses existing techniques, even exceeding those that require fine-tuning.

Xin Su, Phillip Howard, Nagib Hakim et al.

Answering time-sensitive questions from long documents requires temporal reasoning over the times in questions and documents. An important open question is whether large language models can perform such reasoning solely using a provided text document, or whether they can benefit from additional temporal information extracted using other systems. We address this research question by applying existing temporal information extraction systems to construct temporal graphs of events, times, and temporal relations in questions and documents. We then investigate different approaches for fusing these graphs into Transformer models. Experimental results show that our proposed approach for fusing temporal graphs into input text substantially enhances the temporal reasoning capabilities of Transformer models with or without fine-tuning. Additionally, our proposed method outperforms various graph convolution-based approaches and establishes a new state-of-the-art performance on SituatedQA and three splits of TimeQA.

Phillip Howard, Arden Ma, Vasudev Lal et al.

The extraction of aspect terms is a critical step in fine-grained sentiment analysis of text. Existing approaches for this task have yielded impressive results when the training and testing data are from the same domain. However, these methods show a drastic decrease in performance when applied to cross-domain settings where the domain of the testing data differs from that of the training data. To address this lack of extensibility and robustness, we propose a novel approach for automatically constructing domain-specific knowledge graphs that contain information relevant to the identification of aspect terms. We introduce a methodology for injecting information from these knowledge graphs into Transformer models, including two alternative mechanisms for knowledge insertion: via query enrichment and via manipulation of attention patterns. We demonstrate state-of-the-art performance on benchmark datasets for cross-domain aspect term extraction using our approach and investigate how the amount of external knowledge available to the Transformer impacts model performance.

Gadi Singer, Joscha Bach, Tetiana Grinberg et al.

While end-to-end learning systems are rapidly gaining capabilities and popularity, the increasing computational demands for deploying such systems, along with a lack of flexibility, adaptability, explainability, reasoning and verification capabilities, require new types of architectures. Here we introduce a classification of hybrid systems which, based on an analysis of human knowledge and intelligence, combines neural learning with various types of knowledge and knowledge sources. We present the Thrill-K architecture as a prototypical solution for integrating instantaneous knowledge, standby knowledge and external knowledge sources in a framework capable of inference, learning and intelligent control.

Man Luo, Bradley Peterson, Rafael Gan et al.

Peer review is crucial for advancing and improving science through constructive criticism. However, toxic feedback can discourage authors and hinder scientific progress. This work explores an important but underexplored area: detecting toxicity in peer reviews. We first define toxicity in peer reviews across four distinct categories and curate a dataset of peer reviews from the OpenReview platform, annotated by human experts according to these definitions. Leveraging this dataset, we benchmark a variety of models, including a dedicated toxicity detection model, a sentiment analysis model, several open-source large language models (LLMs), and two closed-source LLMs. Our experiments explore the impact of different prompt granularities, from coarse to fine-grained instructions, on model performance. Notably, state-of-the-art LLMs like GPT-4 exhibit low alignment with human judgments under simple prompts but achieve improved alignment with detailed instructions. Moreover, the model's confidence score is a good indicator of better alignment with human judgments. For example, GPT-4 achieves a Cohen's Kappa score of 0.56 with human judgments, which increases to 0.63 when using only predictions with a confidence score higher than 95%. Overall, our dataset and benchmarks underscore the need for continued research to enhance toxicity detection capabilities of LLMs. By addressing this issue, our work aims to contribute to a healthy and responsible environment for constructive academic discourse and scientific collaboration.

Neale Ratzlaff, Man Luo, Xin Su et al.

Multimodal models typically combine a powerful large language model (LLM) with a vision encoder and are then trained on multimodal data via instruction tuning. While this process adapts LLMs to multimodal settings, it remains unclear whether this adaptation compromises their original language reasoning capabilities. In this work, we explore the effects of multimodal instruction tuning on language reasoning performance. We focus on LLaVA, a leading multimodal framework that integrates LLMs such as Vicuna or Mistral with the CLIP vision encoder. We compare the performance of the original LLMs with their multimodal-adapted counterparts across eight language reasoning tasks. Our experiments yield several key insights. First, the impact of multimodal learning varies between Vicuna and Mistral: we observe a degradation in language reasoning for Mistral but improvements for Vicuna across most tasks. Second, while multimodal instruction learning consistently degrades performance on mathematical reasoning tasks (e.g., GSM8K), it enhances performance on commonsense reasoning tasks (e.g., CommonsenseQA). Finally, we demonstrate that a training-free model merging technique can effectively mitigate the language reasoning degradation observed in multimodal-adapted Mistral and even improve performance on visual tasks.

Sezen Perçin, Xin Su, Qutub Sha Syed et al.

Large language models (LLMs) are very costly and inefficient to update with new information. To address this limitation, retrieval-augmented generation (RAG) has been proposed as a solution that dynamically incorporates external knowledge during inference, improving factual consistency and reducing hallucinations. Despite its promise, RAG systems face practical challenges-most notably, a strong dependence on the quality of the input query for accurate retrieval. In this paper, we investigate the sensitivity of different components in the RAG pipeline to various types of query perturbations. Our analysis reveals that the performance of commonly used retrievers can degrade significantly even under minor query variations. We study each module in isolation as well as their combined effect in an end-to-end question answering setting, using both general-domain and domain-specific datasets. Additionally, we propose an evaluation framework to systematically assess the query-level robustness of RAG pipelines and offer actionable recommendations for practitioners based on the results of more than 1092 experiments we performed.

Matthew Lyle Olson, Musashi Hinck, Neale Ratzlaff et al.

The ImageNet hierarchy provides a structured taxonomy of object categories, offering a valuable lens through which to analyze the representations learned by deep vision models. In this work, we conduct a comprehensive analysis of how vision models encode the ImageNet hierarchy, leveraging Sparse Autoencoders (SAEs) to probe their internal representations. SAEs have been widely used as an explanation tool for large language models (LLMs), where they enable the discovery of semantically meaningful features. Here, we extend their use to vision models to investigate whether learned representations align with the ontological structure defined by the ImageNet taxonomy. Our results show that SAEs uncover hierarchical relationships in model activations, revealing an implicit encoding of taxonomic structure. We analyze the consistency of these representations across different layers of the popular vision foundation model DINOv2 and provide insights into how deep vision models internalize hierarchical category information by increasing information in the class token through each layer. Our study establishes a framework for systematic hierarchical analysis of vision model representations and highlights the potential of SAEs as a tool for probing semantic structure in deep networks.

Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck et al.

Large Multi-Modal Models (LMMs) have demonstrated impressive capabilities as general-purpose chatbots able to engage in conversations about visual inputs. However, their responses are influenced by societal biases present in their training datasets, leading to undesirable differences in how the model responds when presented with images depicting people of different demographics. In this work, we propose a training-free debiasing framework for LMMs that intervenes on the model's representations during text generation by constructing a steering vector that reduces reference on protected attributes. Our framework introduces two complementary methods: (1) a dataset-based approach that constructs a steering vector by contrasting model activations on biased and neutral inputs, and (2) a novel optimization-based approach designed for low-resource settings, which constructs the steering vector using a single step of gradient-based perturbation without requiring additional data. Our experiments show that these interventions effectively reduce the propensity of LMMs to generate text related to protected attributes while maintaining sentiment and fluency. Furthermore, we demonstrate that debiased LMMs achieve comparable accuracy to their unmodified counterparts on downstream tasks, indicating that bias mitigation can be achieved without sacrificing model performance.

Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck et al.

Large Vision Language Models (LVLMs) such as LLaVA have demonstrated impressive capabilities as general-purpose chatbots that can engage in conversations about a provided input image. However, their responses are influenced by societal biases present in their training datasets, leading to undesirable differences in how the model responds when presented with images depicting people of different demographics. In this work, we propose a novel debiasing framework for LVLMs by directly ablating biased attributes during text generation to avoid generating text related to protected attributes, or even representing them internally. Our method requires no training and a relatively small amount of representative biased outputs (~1000 samples). Our experiments show that not only can we can minimize the propensity of LVLMs to generate text related to protected attributes, but we can even use synthetic data to inform the ablation while retaining captioning performance on real data such as COCO. Furthermore, we find the resulting generations from a debiased LVLM exhibit similar accuracy as a baseline biased model, showing that debiasing effects can be achieved without sacrificing model performance.

Xin Su, Sungduk Yu, Phillip Howard et al.

Time normalization is the task of converting natural language temporal expressions into machine-readable representations. It underpins many downstream applications in information retrieval, question answering, and clinical decision-making. Traditional systems based on the ISO-TimeML schema limit expressivity and struggle with complex constructs such as compositional, event-relative, and multi-span time expressions. In this work, we introduce a novel formulation of time normalization as a code generation task grounded in the SCATE framework, which defines temporal semantics through symbolic and compositional operators. We implement a fully executable SCATE Python library and demonstrate that large language models (LLMs) can generate executable SCATE code. Leveraging this capability, we develop an automatic data augmentation pipeline using LLMs to synthesize large-scale annotated data with code-level validation. Our experiments show that small, locally deployable models trained on this augmented data can achieve strong performance, outperforming even their LLM parents and enabling practical, accurate, and interpretable time normalization.

Costas Mavromatis, Prasanna Lakkur Subramanyam, Vassilis N. Ioannidis et al.

Knowledge Graph Question Answering (KGQA) involves retrieving facts from a Knowledge Graph (KG) using natural language queries. A KG is a curated set of facts consisting of entities linked by relations. Certain facts include also temporal information forming a Temporal KG (TKG). Although many natural questions involve explicit or implicit time constraints, question answering (QA) over TKGs has been a relatively unexplored area. Existing solutions are mainly designed for simple temporal questions that can be answered directly by a single TKG fact. This paper puts forth a comprehensive embedding-based framework for answering complex questions over TKGs. Our method termed temporal question reasoning (TempoQR) exploits TKG embeddings to ground the question to the specific entities and time scope it refers to. It does so by augmenting the question embeddings with context, entity and time-aware information by employing three specialized modules. The first computes a textual representation of a given question, the second combines it with the entity embeddings for entities involved in the question, and the third generates question-specific time embeddings. Finally, a transformer-based encoder learns to fuse the generated temporal information with the question representation, which is used for answer predictions. Extensive experiments show that TempoQR improves accuracy by 25--45 percentage points on complex temporal questions over state-of-the-art approaches and it generalizes better to unseen question types.