Recent work has aimed to improve LLM generations by filtering out hallucinations, thereby improving the precision of the information in responses. Correctness of a long-form response, however, also depends on the recall of multiple pieces of information relevant to the question. In this paper, we introduce Atomic Self-Consistency (ASC), a technique for improving the recall of relevant information in an LLM response. ASC follows recent work, Universal Self-Consistency (USC) in using multiple stochastic samples from an LLM to improve the long-form response. Unlike USC which only focuses on selecting the best single generation, ASC picks authentic subparts from the samples and merges them into a superior composite answer. Through extensive experiments and ablations, we show that merging relevant subparts of multiple samples performs significantly better than picking a single sample. ASC demonstrates significant gains over USC on multiple factoids and open-ended QA datasets - ASQA, QAMPARI, QUEST, ELI5 with ChatGPT and Llama3. Our analysis also reveals untapped potential for enhancing long-form generations using the approach of merging multiple samples.

Sentence embedding models are typically trained using contrastive learning (CL), either using human annotations directly or by repurposing other annotated datasets. In this work, we explore the recently introduced paradigm of generating CL data using generative language models (LM). In CL for computer vision (CV), compositional transformations (series of operations applied over an image. e.g. cropping + color distortion) which modify the input/image to retain minimal information were shown to be very effective. We show that composition of a ‘Summary’ transformation with diverse paraphrasing/contradicting transformations accomplishes the same and works very well in CL for sentence embeddings. Our final generated dataset (using Vicuna-13B) significantly outperforms the previous best unsupervised method (using ChatGPT) by 1.8 points, and SimCSE, a strong supervised baseline by 0.3 points on the semantic text similarity (STS) benchmark.

To enhance Large Language Models’ (LLMs) reliability, calibration is essential—the model’s confidence scores should align with the likelihood of its responses being correct. However, traditional calibration methods typically rely on a binary true/false assessment of response correctness, unsuitable for long-form generations where an answer can be partially correct. Addressing this gap, we introduce a unified calibration framework, in which both the correctness of the LLMs’ responses and their associated confidence levels are treated as distributions across a range of scores. We develop three metrics for assessing LLM calibration and propose confidence elicitation methods based on self-consistency and self-evaluation. Our experiments demonstrate that larger models don’t necessarily guarantee better calibration, that various calibration metrics complement each other, and that self-consistency methods excel in factoid datasets. We also find that calibration can be enhanced through techniques such as fine-tuning, scaling the temperature. Finally, we illustrate one application of long-form calibration through selective answering in long-form responses, optimizing correctness within a constrained API budget.

Data selection techniques, which adaptively select datapoints inside the training loop, have demonstrated empirical benefits in reducing the number of gradient steps to train neural models. However, these techniques have so far largely been applied to classification. In this work, we study their applicability to language model pretraining, a highly time-intensive task. We propose a simple modification to an existing data selection technique (reducible hold-out loss training) in order to adapt it to the sequence losses typical in language modeling. We experiment on both autoregressive and masked language modelling, and show that applying data selection to pretraining offers notable benefits including a 4.3% reduction in total number of steps, a 21.5% steps reduction in average, to an intermediate target perplexity, over the course of pretraining an autoregressive language model. Further, data selection trained language models demonstrate significantly better generalization ability on out of domain datasets - 7.9% reduction in total number of steps and 23.2% average steps reduction to an intermediate target perplexity.

Ontonotes has served as the most important benchmark for coreference resolution. However, for ease of annotation, several long documents in Ontonotes were split into smaller parts. In this work, we build a corpus of coreference-annotated documents of significantly longer length than what is currently available. We do so by providing an accurate, manually-curated, merging of annotations from documents that were split into multiple parts in the original Ontonotes annotation process. The resulting corpus, which we call LongtoNotes contains documents in multiple genres of the English language with varying lengths, the longest of which are up to 8x the length of documents in Ontonotes, and 2x those in Litbank.We evaluate state-of-the-art neural coreference systems on this new corpus, analyze the relationships between model architectures/hyperparameters and document length on performance and efficiency of the models, and demonstrate areas of improvement in long-document coreference modelling revealed by our new corpus.

Smart assistants are tasked to answer various questions regarding world knowledge. These questions range from retrieval of simple facts to retrieval of complex, multi-hops question followed by various operators (i.e., filter, argmax). Semantic parsing has emerged as the state-of-the-art for answering these kinds of questions by forming queries to extract information from knowledge bases (KBs). Specially, neural semantic parsers (NSPs) effectively translate natural questions to logical forms, which execute on KB and give desirable answers. Yet, NSPs suffer from non-executable logical forms for some instances in the generated logical forms might be missing due to the incompleteness of KBs. Intuitively, knowing the KB structure informs NSP with changes of the global logical forms structures with respect to changes in KB instances. In this work, we propose a novel knowledge-informed decoder variant of NSP. We consider the conversational question answering settings, where a natural language query, its context and its final answers are available at training. Experimental results show that our method outperformed strong baselines by 1.8 F1 points overall across 10 types of questions of the CSQA dataset. Especially for the “Logical Reasoning” category, our model improves by 7 F1 points. Furthermore, our results are achieved with 90.3% fewer parameters, allowing faster training for large-scale datasets.

Pre-trained language models have emerged as highly successful methods for learning good text representations. However, the amount of structured knowledge retained in such models, and how (if at all) it can be extracted, remains an open question. In this work, we aim at directly learning text representations which leverage structured knowledge about entities mentioned in the text. This can be particularly beneficial for downstream tasks which are knowledge-intensive. Our approach utilizes self-attention between words in the text and knowledge graph (KG) entities mentioned in the text. While existing methods require entity-linked data for pre-training, we train using a mention-span masking objective and a candidate ranking objective – which doesn’t require any entity-links and only assumes access to an alias table for retrieving candidates, enabling large-scale pre-training. We show that the proposed model learns knowledge-informed text representations that yield improvements on the downstream tasks over existing methods.