Abstract: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.
Abstract: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.
Abstract: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.
Abstract:Training models on synthetic data has emerged as an increasingly important strategy for improving the performance of generative AI. This approach is particularly helpful for large multimodal models (LMMs) due to the relative scarcity of high-quality paired image-text data compared to language-only data. While a variety of methods have been proposed for generating large multimodal datasets, they do not tailor the synthetic data to address specific deficiencies in the reasoning abilities of LMMs which will be trained with the generated dataset. In contrast, humans often learn in a more efficient manner by seeking out examples related to the types of reasoning where they have failed previously. Inspired by this observation, we propose a new approach for synthetic data generation which is grounded in the analysis of an existing LMM's reasoning failures. Our methodology leverages frontier models to automatically analyze errors produced by a weaker LMM and propose new examples which can be used to correct the reasoning failure via additional training, which are then further filtered to ensure high quality. We generate a large multimodal instruction tuning dataset containing over 553k examples using our approach and conduct extensive experiments demonstrating its utility for improving the performance of LMMs on multiple downstream tasks. Our results show that models trained on our synthetic data can even exceed the performance of LMMs trained on an equivalent amount of additional real data, demonstrating the high value of generating synthetic data targeted to specific reasoning failure modes in LMMs. We will make our dataset and code publicly available.
Abstract:Temporal information extraction (IE) aims to extract structured temporal information from unstructured text, thereby uncovering the implicit timelines within. This technique is applied across domains such as healthcare, newswire, and intelligence analysis, aiding models in these areas to perform temporal reasoning and enabling human users to grasp the temporal structure of text. Transformer-based pre-trained language models have produced revolutionary advancements in natural language processing, demonstrating exceptional performance across a multitude of tasks. Despite the achievements garnered by Transformer-based approaches in temporal IE, there is a lack of comprehensive reviews on these endeavors. In this paper, we aim to bridge this gap by systematically summarizing and analyzing the body of work on temporal IE using Transformers while highlighting potential future research directions.
Abstract:CLIP is one of the most popular foundational models and is heavily used for many vision-language tasks. However, little is known about the inner workings of CLIP. While recent work has proposed decomposition-based interpretability methods for identifying textual descriptions of attention heads in CLIP, the implications of conceptual consistency in these text labels on interpretability and model performance has not been explored. To bridge this gap, we study the conceptual consistency of text descriptions for attention heads in CLIP-like models. We conduct extensive experiments on six different models from OpenAI and OpenCLIP which vary by size, type of pre-training data and patch size. We propose Concept Consistency Score (CCS), a novel interpretability metric that measures how consistently individual attention heads in CLIP models align with specific concepts. To assign concept labels to heads, we use in-context learning with ChatGPT, guided by a few manually-curated examples, and validate these labels using an LLM-as-a-judge approach. Our soft-pruning experiments reveal that high CCS heads are critical for preserving model performance, as pruning them leads to a significantly larger performance drop than pruning random or low CCS heads. Notably, we find that high CCS heads capture essential concepts and play a key role in out-of-domain detection, concept-specific reasoning, and video-language understanding. These results position CCS as a powerful interpretability metric for analyzing CLIP-like models.
Abstract:Despite recent efforts in understanding the compression impact on large language models (LLMs) in terms of their downstream task performance and trustworthiness on relatively simpler uni-modal benchmarks (for example, question answering, common sense reasoning), their detailed study on multi-modal Large Vision-Language Models (LVLMs) is yet to be unveiled. Towards mitigating this gap, we present LVLM-Compress-Bench, a framework to first thoroughly study the broad impact of compression on the generative performance of LVLMs with multi-modal input driven tasks. In specific, we consider two major classes of compression for autoregressive models, namely KV cache and weight compression, for the dynamically growing intermediate cache and static weights, respectively. We use four LVLM variants of the popular LLaVA framework to present our analysis via integrating various state-of-the-art KV and weight compression methods including uniform, outlier-reduced, and group quantization for the KV cache and weights. With this framework we demonstrate on ten different multi-modal datasets with different capabilities including recognition, knowledge, language generation, spatial awareness, visual reasoning, hallucination and visual illusion identification, toxicity, stereotypes and bias. In specific, our framework demonstrates the compression impact on both general and ethically critical metrics leveraging a combination of real world and synthetic datasets to encompass diverse societal intersectional attributes. Extensive experimental evaluations yield diverse and intriguing observations on the behavior of LVLMs at different quantization budget of KV and weights, in both maintaining and losing performance as compared to the baseline model with FP16 data format. Code will be open-sourced at https://github.com/opengear-project/LVLM-compress-bench.
Abstract:Peer review is a critical process for ensuring the integrity of published scientific research. Confidence in this process is predicated on the assumption that experts in the relevant domain give careful consideration to the merits of manuscripts which are submitted for publication. With the recent rapid advancements in large language models (LLMs), a new risk to the peer review process is that negligent reviewers will rely on LLMs to perform the often time consuming process of reviewing a paper. However, there is a lack of existing resources for benchmarking the detectability of AI text in the domain of peer review. To address this deficiency, we introduce a comprehensive dataset containing a total of 788,984 AI-written peer reviews paired with corresponding human reviews, covering 8 years of papers submitted to each of two leading AI research conferences (ICLR and NeurIPS). We use this new resource to evaluate the ability of 18 existing AI text detection algorithms to distinguish between peer reviews written by humans and different state-of-the-art LLMs. Motivated by the shortcomings of existing methods, we propose a new detection approach which surpasses existing methods in the identification of AI written peer reviews. Our work reveals the difficulty of identifying AI-generated text at the individual peer review level, highlighting the urgent need for new tools and methods to detect this unethical use of generative AI.
Abstract: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.
Abstract:Large Multi-Modal Models (LMMs) have demonstrated impressive capabilities as general-purpose chatbots that can engage in conversations about a provided input, such as an 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 LMMs that directly removes biased representations during text generation to decrease outputs related to protected attributes, or even representing them internally. Our proposed method is training-free; given a single image and a list of target attributes, we can ablate the corresponding representations with just one step of gradient descent on the image itself. Our experiments show that not only can we can minimize the propensity of LMMs to generate text related to protected attributes, but we can improve sentiment and even simply use synthetic data to inform the ablation while retaining language modeling capabilities on real data such as COCO or FACET. Furthermore, we find the resulting generations from a debiased LMM exhibit similar accuracy as a baseline biased model, showing that debiasing effects can be achieved without sacrificing model performance.