Search

While many capabilities of language models (LMs) improve with increased training budget, the influence of scale on hallucinations is not yet fully understood. Hallucinations come in many forms, and there is no universally accepted definition. We thus focus on studying only those hallucinations where a correct answer appears verbatim in the training set. To fully control the training data content, we construct a knowledge graph (KG)-based dataset, and use it to train a set of increasingly large LMs. We find that for a fixed dataset, larger and longer-trained LMs hallucinate less. However, hallucinating on $\leq5$% of the training data requires an order of magnitude larger model, and thus an order of magnitude more compute, than Hoffmann et al. (2022) reported was optimal. Given this costliness, we study how hallucination detectors depend on scale. While we see detector size improves performance on fixed LM's outputs, we find an inverse relationship between the scale of the LM and the detectability of its hallucinations., Comment: Published at COLM 2024. 16 pages, 11 figures

In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community.

Aligning language models (LMs) with curated human feedback is critical to control their behaviors in real-world applications. Several recent policy optimization methods, such as DPO and SLiC, serve as promising alternatives to the traditional Reinforcement Learning from Human Feedback (RLHF) approach. In practice, human feedback often comes in a format of a ranked list over multiple responses to amortize the cost of reading prompt. Multiple responses can also be ranked by reward models or AI feedback. There lacks such a thorough study on directly fitting upon a list of responses. In this work, we formulate the LM alignment as a \textit{listwise} ranking problem and describe the LiPO framework, where the policy can potentially learn more effectively from a ranked list of plausible responses given the prompt. This view draws an explicit connection to Learning-to-Rank (LTR), where most existing preference optimization work can be mapped to existing ranking objectives. Following this connection, we provide an examination of ranking objectives that are not well studied for LM alignment with DPO and SLiC as special cases when list size is two. In particular, we highlight a specific method, LiPO-$\lambda$, which leverages a state-of-the-art \textit{listwise} ranking objective and weights each preference pair in a more advanced manner. We show that LiPO-$\lambda$ can outperform DPO variants and SLiC by a clear margin on several preference alignment tasks with both curated and real rankwise preference data.

Safe deployment of large language models (LLMs) may benefit from a reliable method for assessing their generated content to determine when to abstain or to selectively generate. While likelihood-based metrics such as perplexity are widely employed, recent research has demonstrated the limitations of using sequence-level probability estimates given by LLMs as reliable indicators of generation quality. Conversely, LLMs have demonstrated strong calibration at the token level, particularly when it comes to choosing correct answers in multiple-choice questions or evaluating true/false statements. In this work, we reformulate open-ended generation tasks into token-level prediction tasks, and leverage LLMs' superior calibration at the token level. We instruct an LLM to self-evaluate its answers, employing either a multi-way comparison or a point-wise evaluation approach, with the option to include a ``None of the above'' option to express the model's uncertainty explicitly. We benchmark a range of scoring methods based on self-evaluation and evaluate their performance in selective generation using TruthfulQA and TL;DR. Through experiments with PaLM-2 and GPT-3, we demonstrate that self-evaluation based scores not only improve accuracy, but also correlate better with the overall quality of generated content.

We introduce and study the problem of adversarial arithmetic, which provides a simple yet challenging testbed for language model alignment. This problem is comprised of arithmetic questions posed in natural language, with an arbitrary adversarial string inserted before the question is complete. Even in the simple setting of 1-digit addition problems, it is easy to find adversarial prompts that make all tested models (including PaLM2, GPT4, Claude2) misbehave, and even to steer models to a particular wrong answer. We additionally provide a simple algorithm for finding successful attacks by querying those same models, which we name "prompt inversion rejection sampling" (PIRS). We finally show that models can be partially hardened against these attacks via reinforcement learning and via agentic constitutional loops. However, we were not able to make a language model fully robust against adversarial arithmetic attacks.

Despite their success in many natural language tasks, solving math problems remains a significant challenge for large language models (LLMs). A large gap exists between LLMs' pass-at-one and pass-at-N performance in solving math problems, suggesting LLMs might be close to finding correct solutions, motivating our exploration of fine-tuning methods to unlock LLMs' performance. Using the challenging MATH dataset, we investigate three fine-tuning strategies: (1) solution fine-tuning, where we fine-tune to generate a detailed solution for a given math problem; (2) solution-cluster re-ranking, where the LLM is fine-tuned as a solution verifier/evaluator to choose among generated candidate solution clusters; (3) multi-task sequential fine-tuning, which integrates both solution generation and evaluation tasks together efficiently to enhance the LLM performance. With these methods, we present a thorough empirical study on a series of PaLM 2 models and find: (1) The quality and style of the step-by-step solutions used for fine-tuning can make a significant impact on the model performance; (2) While solution re-ranking and majority voting are both effective for improving the model performance when used separately, they can also be used together for an even greater performance boost; (3) Multi-task fine-tuning that sequentially separates the solution generation and evaluation tasks can offer improved performance compared with the solution fine-tuning baseline. Guided by these insights, we design a fine-tuning recipe that yields approximately 58.8% accuracy on the MATH dataset with fine-tuned PaLM 2-L models, an 11.2% accuracy improvement over the few-shot performance of pre-trained PaLM 2-L model with majority voting.

Improving the alignment of language models with human preferences remains an active research challenge. Previous approaches have primarily utilized Reinforcement Learning from Human Feedback (RLHF) via online RL methods such as Proximal Policy Optimization (PPO). Recently, offline methods such as Sequence Likelihood Calibration (SLiC) and Direct Preference Optimization (DPO) have emerged as attractive alternatives, offering improvements in stability and scalability while maintaining competitive performance. SLiC refines its loss function using sequence pairs sampled from a supervised fine-tuned (SFT) policy, while DPO directly optimizes language models based on preference data, foregoing the need for a separate reward model. However, the maximum likelihood estimator (MLE) of the target optimal policy requires labeled preference pairs sampled from that policy. DPO's lack of a reward model constrains its ability to sample preference pairs from the optimal policy, and SLiC is restricted to sampling preference pairs only from the SFT policy. To address these limitations, we introduce a novel approach called Statistical Rejection Sampling Optimization (RSO) that aims to source preference data from the target optimal policy using rejection sampling, enabling a more accurate estimation of the optimal policy. We also propose a unified framework that enhances the loss functions used in both SLiC and DPO from a preference modeling standpoint. Through extensive experiments across three diverse tasks, we demonstrate that RSO consistently outperforms both SLiC and DPO on evaluations from both Large Language Model (LLM) and human raters., Comment: Accepted in ICLR 2024

Learning from human feedback has been shown to be effective at aligning language models with human preferences. Past work has often relied on Reinforcement Learning from Human Feedback (RLHF), which optimizes the language model using reward scores assigned from a reward model trained on human preference data. In this work we show how the recently introduced Sequence Likelihood Calibration (SLiC), can also be used to effectively learn from human preferences (SLiC-HF). Furthermore, we demonstrate this can be done with human feedback data collected for a different model, similar to off-policy, offline RL data. Automatic and human evaluation experiments on the TL;DR summarization task show that SLiC-HF significantly improves supervised fine-tuning baselines. Furthermore, SLiC-HF presents a competitive alternative to the PPO RLHF implementation used in past work while being much simpler to implement, easier to tune and more computationally efficient in practice.

The evaluation of abstractive summarization models typically uses test data that is identically distributed as training data. In real-world practice, documents to be summarized may contain input noise caused by text extraction artifacts or data pipeline bugs. The robustness of model performance under distribution shift caused by such noise is relatively under-studied. We present a large empirical study quantifying the sometimes severe loss in performance (up to 12 ROUGE-1 points) from different types of input noise for a range of datasets and model sizes. We then propose a light-weight method for detecting and removing such noise in the input during model inference without requiring any extra training, auxiliary models, or even prior knowledge of the type of noise. Our proposed approach effectively mitigates the loss in performance, recovering a large fraction of the performance drop, sometimes as large as 11 ROUGE-1 points., Comment: EMNLP Findings 2023 Camera Ready

Conditional language models are predominantly trained with maximum likelihood estimation (MLE), giving probability mass to sparsely observed target sequences. While MLE trained models assign high probability to plausible sequences given the context, the model probabilities often do not accurately rank-order generated sequences by quality. This has been empirically observed in beam search decoding as output quality degrading with large beam sizes, and decoding strategies benefiting from heuristics such as length normalization and repetition-blocking. In this work, we introduce sequence likelihood calibration (SLiC) where the likelihood of model generated sequences are calibrated to better align with reference sequences in the model's latent space. With SLiC, decoding heuristics become unnecessary and decoding candidates' quality significantly improves regardless of the decoding method. Furthermore, SLiC shows no sign of diminishing returns with model scale, and presents alternative ways to improve quality with limited training and inference budgets. With SLiC, we exceed or match SOTA results on a wide range of generation tasks spanning abstractive summarization, question generation, abstractive question answering and data-to-text generation, even with modest-sized models.

Machine learning algorithms typically assume independent and identically distributed samples in training and at test time. Much work has shown that high-performing ML classifiers can degrade significantly and provide overly-confident, wrong classification predictions, particularly for out-of-distribution (OOD) inputs. Conditional language models (CLMs) are predominantly trained to classify the next token in an output sequence, and may suffer even worse degradation on OOD inputs as the prediction is done auto-regressively over many steps. Furthermore, the space of potential low-quality outputs is larger as arbitrary text can be generated and it is important to know when to trust the generated output. We present a highly accurate and lightweight OOD detection method for CLMs, and demonstrate its effectiveness on abstractive summarization and translation. We also show how our method can be used under the common and realistic setting of distribution shift for selective generation (analogous to selective prediction for classification) of high-quality outputs, while automatically abstaining from low-quality ones, enabling safer deployment of generative language models., Comment: Published in ICLR 2023

While large pretrained Transformer models have proven highly capable at tackling natural language tasks, handling long sequence inputs continues to be a significant challenge. One such task is long input summarization, where inputs are longer than the maximum input context of most pretrained models. Through an extensive set of experiments, we investigate what model architectural changes and pretraining paradigms can most efficiently adapt a pretrained Transformer for long input summarization. We find that a staggered, block-local Transformer with global encoder tokens strikes a good balance of performance and efficiency, and that an additional pretraining phase on long sequences meaningfully improves downstream summarization performance. Based on our findings, we introduce PEGASUS-X, an extension of the PEGASUS model with additional long input pretraining to handle inputs of up to 16K tokens. PEGASUS-X achieves strong performance on long input summarization tasks comparable with much larger models while adding few additional parameters and not requiring model parallelism to train.

Widely used evaluation metrics for text generation either do not work well with longer texts or fail to evaluate all aspects of text quality. In this paper, we introduce a new metric called SMART to mitigate such limitations. Specifically, We treat sentences as basic units of matching instead of tokens, and use a sentence matching function to soft-match candidate and reference sentences. Candidate sentences are also compared to sentences in the source documents to allow grounding (e.g., factuality) evaluation. Our results show that system-level correlations of our proposed metric with a model-based matching function outperforms all competing metrics on the SummEval summarization meta-evaluation dataset, while the same metric with a string-based matching function is competitive with current model-based metrics. The latter does not use any neural model, which is useful during model development phases where resources can be limited and fast evaluation is required. Finally, we also conducted extensive analyses showing that our proposed metrics work well with longer summaries and are less biased towards specific models., Comment: code coming soon

Most prior work in the sequence-to-sequence paradigm focused on datasets with input sequence lengths in the hundreds of tokens due to the computational constraints of common RNN and Transformer architectures. In this paper, we study long-form abstractive text summarization, a sequence-to-sequence setting with input sequence lengths up to 100,000 tokens and output sequence lengths up to 768 tokens. We propose SEAL, a Transformer-based model, featuring a new encoder-decoder attention that dynamically extracts/selects input snippets to sparsely attend to for each output segment. Using only the original documents and summaries, we derive proxy labels that provide weak supervision for extractive layers simultaneously with regular supervision from abstractive summaries. The SEAL model achieves state-of-the-art results on existing long-form summarization tasks, and outperforms strong baseline models on a new dataset/task we introduce, Search2Wiki, with much longer input text. Since content selection is explicit in the SEAL model, a desirable side effect is that the selection can be inspected for enhanced interpretability.

Recent work pre-training Transformers with self-supervised objectives on large text corpora has shown great success when fine-tuned on downstream NLP tasks including text summarization. However, pre-training objectives tailored for abstractive text summarization have not been explored. Furthermore there is a lack of systematic evaluation across diverse domains. In this work, we propose pre-training large Transformer-based encoder-decoder models on massive text corpora with a new self-supervised objective. In PEGASUS, important sentences are removed/masked from an input document and are generated together as one output sequence from the remaining sentences, similar to an extractive summary. We evaluated our best PEGASUS model on 12 downstream summarization tasks spanning news, science, stories, instructions, emails, patents, and legislative bills. Experiments demonstrate it achieves state-of-the-art performance on all 12 downstream datasets measured by ROUGE scores. Our model also shows surprising performance on low-resource summarization, surpassing previous state-of-the-art results on 6 datasets with only 1000 examples. Finally we validated our results using human evaluation and show that our model summaries achieve human performance on multiple datasets., Comment: Added results from mixed+stochastic model, test-set overlapping analysis; Code link added; Accepted for ICML 2020. arXiv admin note: text overlap with arXiv:1605.06560, arXiv:1205.2395, arXiv:0902.4351, arXiv:1610.09932, arXiv:nucl-ex/0512029 by other authors

Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts all text-based language problems into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled data sets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new ``Colossal Clean Crawled Corpus'', we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our data set, pre-trained models, and code.

We propose an end-to-end neural model for zero-shot abstractive text summarization of paragraphs, and introduce a benchmark task, ROCSumm, based on ROCStories, a subset for which we collected human summaries. In this task, five-sentence stories (paragraphs) are summarized with one sentence, using human summaries only for evaluation. We show results for extractive and human baselines to demonstrate a large abstractive gap in performance. Our model, SummAE, consists of a denoising auto-encoder that embeds sentences and paragraphs in a common space, from which either can be decoded. Summaries for paragraphs are generated by decoding a sentence from the paragraph representations. We find that traditional sequence-to-sequence auto-encoders fail to produce good summaries and describe how specific architectural choices and pre-training techniques can significantly improve performance, outperforming extractive baselines. The data, training, evaluation code, and best model weights are open-sourced., Comment: first two authors contributed equally

We propose a model-based metric to estimate the factual accuracy of generated text that is complementary to typical scoring schemes like ROUGE (Recall-Oriented Understudy for Gisting Evaluation) and BLEU (Bilingual Evaluation Understudy). We introduce and release a new large-scale dataset based on Wikipedia and Wikidata to train relation classifiers and end-to-end fact extraction models. The end-to-end models are shown to be able to extract complete sets of facts from datasets with full pages of text. We then analyse multiple models that estimate factual accuracy on a Wikipedia text summarization task, and show their efficacy compared to ROUGE and other model-free variants by conducting a human evaluation study.

Abstractive summarization has been studied using neural sequence transduction methods with datasets of large, paired document-summary examples. However, such datasets are rare and the models trained from them do not generalize to other domains. Recently, some progress has been made in learning sequence-to-sequence mappings with only unpaired examples. In our work, we consider the setting where there are only documents (product or business reviews) with no summaries provided, and propose an end-to-end, neural model architecture to perform unsupervised abstractive summarization. Our proposed model consists of an auto-encoder where the mean of the representations of the input reviews decodes to a reasonable summary-review while not relying on any review-specific features. We consider variants of the proposed architecture and perform an ablation study to show the importance of specific components. We show through automated metrics and human evaluation that the generated summaries are highly abstractive, fluent, relevant, and representative of the average sentiment of the input reviews. Finally, we collect a reference evaluation dataset and show that our model outperforms a strong extractive baseline., Comment: Accepted to ICML 2019

Clinicians spend a significant amount of time inputting free-form textual notes into Electronic Health Records (EHR) systems. Much of this documentation work is seen as a burden, reducing time spent with patients and contributing to clinician burnout. With the aspiration of AI-assisted note-writing, we propose a new language modeling task predicting the content of notes conditioned on past data from a patient's medical record, including patient demographics, labs, medications, and past notes. We train generative models using the public, de-identified MIMIC-III dataset and compare generated notes with those in the dataset on multiple measures. We find that much of the content can be predicted, and that many common templates found in notes can be learned. We discuss how such models can be useful in supporting assistive note-writing features such as error-detection and auto-complete., Comment: preprint

We show that generating English Wikipedia articles can be approached as a multi- document summarization of source documents. We use extractive summarization to coarsely identify salient information and a neural abstractive model to generate the article. For the abstractive model, we introduce a decoder-only architecture that can scalably attend to very long sequences, much longer than typical encoder- decoder architectures used in sequence transduction. We show that this model can generate fluent, coherent multi-sentence paragraphs and even whole Wikipedia articles. When given reference documents, we show it can extract relevant factual information as reflected in perplexity, ROUGE scores and human evaluations., Comment: Published as a conference paper at ICLR 2018

The driving force behind the recent success of LSTMs has been their ability to learn complex and non-linear relationships. Consequently, our inability to describe these relationships has led to LSTMs being characterized as black boxes. To this end, we introduce contextual decomposition (CD), an interpretation algorithm for analysing individual predictions made by standard LSTMs, without any changes to the underlying model. By decomposing the output of a LSTM, CD captures the contributions of combinations of words or variables to the final prediction of an LSTM. On the task of sentiment analysis with the Yelp and SST data sets, we show that CD is able to reliably identify words and phrases of contrasting sentiment, and how they are combined to yield the LSTM's final prediction. Using the phrase-level labels in SST, we also demonstrate that CD is able to successfully extract positive and negative negations from an LSTM, something which has not previously been done., Comment: Oral presentation at ICLR 2018

Neural sequence-to-sequence models have provided a viable new approach for abstractive text summarization (meaning they are not restricted to simply selecting and rearranging passages from the original text). However, these models have two shortcomings: they are liable to reproduce factual details inaccurately, and they tend to repeat themselves. In this work we propose a novel architecture that augments the standard sequence-to-sequence attentional model in two orthogonal ways. First, we use a hybrid pointer-generator network that can copy words from the source text via pointing, which aids accurate reproduction of information, while retaining the ability to produce novel words through the generator. Second, we use coverage to keep track of what has been summarized, which discourages repetition. We apply our model to the CNN / Daily Mail summarization task, outperforming the current abstractive state-of-the-art by at least 2 ROUGE points., Comment: Add METEOR evaluation results, add some citations, fix some equations (what are now equations 1, 8 and 11 were missing a bias term), fix url to pyrouge package, add acknowledgments

Recurrent neural network models with an attention mechanism have proven to be extremely effective on a wide variety of sequence-to-sequence problems. However, the fact that soft attention mechanisms perform a pass over the entire input sequence when producing each element in the output sequence precludes their use in online settings and results in a quadratic time complexity. Based on the insight that the alignment between input and output sequence elements is monotonic in many problems of interest, we propose an end-to-end differentiable method for learning monotonic alignments which, at test time, enables computing attention online and in linear time. We validate our approach on sentence summarization, machine translation, and online speech recognition problems and achieve results competitive with existing sequence-to-sequence models., Comment: ICML camera-ready version; 10 pages + 9 page appendix

This work presents a general unsupervised learning method to improve the accuracy of sequence to sequence (seq2seq) models. In our method, the weights of the encoder and decoder of a seq2seq model are initialized with the pretrained weights of two language models and then fine-tuned with labeled data. We apply this method to challenging benchmarks in machine translation and abstractive summarization and find that it significantly improves the subsequent supervised models. Our main result is that pretraining improves the generalization of seq2seq models. We achieve state-of-the art results on the WMT English$\rightarrow$German task, surpassing a range of methods using both phrase-based machine translation and neural machine translation. Our method achieves a significant improvement of 1.3 BLEU from the previous best models on both WMT'14 and WMT'15 English$\rightarrow$German. We also conduct human evaluations on abstractive summarization and find that our method outperforms a purely supervised learning baseline in a statistically significant manner., Comment: Updated to accepted EMNLP 2017 version