Hallucinations -- the generation of untrue claims -- pose a challenge to the application of large language models (LLMs) [1] thereby motivating the development of metrics to evaluate factual precision. We observe that popular metrics using the Decompose-Then-Verify framework, such as FActScore [2], can be manipulated by adding obvious or repetitive claims to artificially inflate scores. We expand the FActScore dataset to design and analyze factual precision metrics, demonstrating that models can be trained to achieve high scores under existing metrics through exploiting the issues we identify. This motivates our new customizable plug-and-play subclaim selection component called Core, which filters down individual subclaims according to their uniqueness and informativeness. Metrics augmented by Core are substantially more robust as shown in head-to-head comparisons. We release an evaluation framework supporting the modular use of Core (https://github.com/zipJiang/Core) and various decomposition strategies, and we suggest its adoption by the LLM community. [1] Hong et al., "The Hallucinations Leaderboard -- An Open Effort to Measure Hallucinations in Large Language Models", arXiv:2404.05904v2 [cs.CL]. [2] Min et al., "FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation", arXiv:2305.14251v2 [cs.CL].

Large language models (LLMs) fine-tuned for text-retrieval have demonstrated state-of-the-art results across several information retrieval (IR) benchmarks. However, supervised training for improving these models requires numerous labeled examples, which are generally unavailable or expensive to acquire. In this work, we explore the effectiveness of extending reverse engineered adaptation to the context of information retrieval (RE-AdaptIR). We use RE-AdaptIR to improve LLM-based IR models using only unlabeled data. We demonstrate improved performance both in training domains as well as zero-shot in domains where the models have seen no queries. We analyze performance changes in various fine-tuning scenarios and offer findings of immediate use to practitioners.

Free-text rationales play a pivotal role in explainable NLP, bridging the knowledge and reasoning gaps behind a model's decision-making. However, due to the diversity of potential reasoning paths and a corresponding lack of definitive ground truth, their evaluation remains a challenge. Existing evaluation metrics rely on the degree to which a rationale supports a target label, but we find these fall short in evaluating rationales that inadvertently leak the labels. To address this problem, we propose RORA, a Robust free-text Rationale evaluation against label leakage. RORA quantifies the new information supplied by a rationale to justify the label. This is achieved by assessing the conditional V-information \citep{hewitt-etal-2021-conditional} with a predictive family robust against leaky features that can be exploited by a small model. RORA consistently outperforms existing approaches in evaluating human-written, synthetic, or model-generated rationales, particularly demonstrating robustness against label leakage. We also show that RORA aligns well with human judgment, providing a more reliable and accurate measurement across diverse free-text rationales.

While existing video benchmarks largely consider specialized downstream tasks like retrieval or question-answering (QA), contemporary multimodal AI systems must be capable of well-rounded common-sense reasoning akin to human visual understanding. A critical component of human temporal-visual perception is our ability to identify and cognitively model "things happening", or events. Historically, video benchmark tasks have implicitly tested for this ability (e.g., video captioning, in which models describe visual events with natural language), but they do not consider video event understanding as a task in itself. Recent work has begun to explore video analogues to textual event extraction but consists of competing task definitions and datasets limited to highly specific event types. Therefore, while there is a rich domain of event-centric video research spanning the past 10+ years, it is unclear how video event understanding should be framed and what resources we have to study it. In this paper, we survey 105 video datasets that require event understanding capability, consider how they contribute to the study of robust event understanding in video, and assess proposed video event extraction tasks in the context of this body of research. We propose suggestions informed by this survey for dataset curation and task framing, with an emphasis on the uniquely temporal nature of video events and ambiguity in visual content.

We introduce RE-Adapt, an approach to fine-tuning large language models on new domains without degrading any pre-existing instruction-tuning. We reverse engineer an adapter which isolates what an instruction-tuned model has learned beyond its corresponding pretrained base model. Importantly, this requires no additional data or training. We can then fine-tune the base model on a new domain and readapt it to instruction following with the reverse engineered adapter. RE-Adapt and our low-rank variant LoRE-Adapt both outperform other methods of fine-tuning, across multiple popular LLMs and datasets, even when the models are used in conjunction with retrieval-augmented generation.

Modern Language Models (LMs) are capable of following long and complex instructions that enable a large and diverse set of user requests. While Information Retrieval (IR) models use these LMs as the backbone of their architectures, virtually none of them allow users to provide detailed instructions alongside queries, thus limiting their ability to satisfy complex information needs. In this work, we study the use of instructions in IR systems. First, we introduce our dataset FollowIR, which contains a rigorous instruction evaluation benchmark as well as a training set for helping IR models learn to better follow real-world instructions. FollowIR repurposes detailed instructions -- also known as narratives -- developed for professional assessors to evaluate retrieval systems. In particular, we build our benchmark from three collections curated for shared tasks at the Text REtrieval Conference (TREC). These collections contains hundreds to thousands of labeled documents per query, making them suitable for our exploration. Through this process, we can measure how well IR models follow instructions, through a new pairwise evaluation framework. Our results indicate that existing retrieval models fail to correctly use instructions, using them for basic keywords and struggling to understand long-form information. However, we show that it is possible for IR models to learn to follow complex instructions: our new FollowIR-7B model has significant improvements after fine-tuning on our training set.

Large language models (LLMs) are increasingly capable of completing knowledge intensive tasks by recalling information from a static pretraining corpus. Here we are concerned with LLMs in the context of evolving data requirements. For instance: batches of new data that are introduced periodically; subsets of data with user-based access controls; or requirements on dynamic removal of documents with guarantees that associated knowledge cannot be recalled. We wish to satisfy these requirements while at the same time ensuring a model does not forget old information when new data becomes available. To address these issues, we introduce AdapterSwap, a training and inference scheme that organizes knowledge from a data collection into a set of low-rank adapters, which are dynamically composed during inference. Our experiments demonstrate AdapterSwap's ability to support efficient continual learning, while also enabling organizations to have fine-grained control over data access and deletion.

Can LLMs continually improve their previous outputs for better results? An affirmative answer would require LLMs to be better at discriminating among previously-generated alternatives, than generating initial responses. We explore the validity of this hypothesis in practice. We first introduce a unified framework that allows us to compare the generative and discriminative capability of any model on any task. Then, in our resulting experimental analysis of several LLMs, we do not observe the performance of those models on discrimination to be reliably better than generation. We hope these findings inform the growing literature on self-improvement AI systems.

For humans to trust the fluent generations of large language models (LLMs), they must be able to verify their correctness against trusted, external sources. Recent efforts aim to increase verifiability through citations of retrieved documents or post-hoc provenance. However, such citations are prone to mistakes that further complicate their verifiability. To address these limitations, we tackle the verifiability goal with a different philosophy: we trivialize the verification process by developing models that quote verbatim statements from trusted sources in pre-training data. We propose Quote-Tuning, which demonstrates the feasibility of aligning LLMs to leverage memorized information and quote from pre-training data. Quote-Tuning quantifies quoting against large corpora with efficient membership inference tools, and uses the amount of quotes as an implicit reward signal to construct a synthetic preference dataset for quoting, without any human annotation. Next, the target model is aligned to quote using preference optimization algorithms. Experimental results show that Quote-Tuning significantly increases the percentage of LLM generation quoted verbatim from high-quality pre-training documents by 55% to 130% relative to untuned models while maintaining response quality. Further experiments demonstrate that Quote-Tuning generalizes quoting to out-of-domain data, is applicable in different tasks, and provides additional benefits to truthfulness. Quote-Tuning not only serves as a hassle-free method to increase quoting but also opens up avenues for improving LLM trustworthiness through better verifiability.

Recent chatbots have demonstrated impressive ability to understand and communicate in raw-text form. However, there is more to the world than raw text. For example, humans spend long hours of their time on web pages, where text is intertwined with other modalities and tasks are accomplished in the form of various complex interactions. Can state-of-the-art multi-modal models generalize to such complex domains? To address this question, we introduce TurkingBench, a benchmark of tasks formulated as web pages containing textual instructions with multi-modal context. Unlike existing work which employs artificially synthesized web pages, here we use natural HTML pages that were originally designed for crowdsourcing workers for various annotation purposes. The HTML instructions of each task are also instantiated with various values (obtained from the crowdsourcing tasks) to form new instances of the task. This benchmark contains 32.2K instances distributed across 158 tasks. Additionally, to facilitate the evaluation on TurkingBench, we develop an evaluation framework that connects the responses of chatbots to modifications on web pages (modifying a text box, checking a radio, etc.). We evaluate the performance of state-of-the-art models, including language-only, vision-only, and layout-only models, and their combinations, on this benchmark. Our findings reveal that these models perform significantly better than random chance, yet considerable room exists for improvement. We hope this benchmark will help facilitate the evaluation and development of web-based agents.