Bryan McCann

Zairah Mustahsan, Abel Lim, Megna Anand et al.

As large language models become components of larger agentic systems, evaluation reliability becomes critical: unreliable sub-agents introduce brittleness into downstream system behavior. Yet current evaluation practice, reporting a single accuracy number from a single run, obscures the variance underlying these results, making it impossible to distinguish genuine capability improvements from lucky sampling. We propose adopting Intraclass Correlation Coefficient (ICC), a metric from measurement science, to characterize this variance. ICC decomposes observed variance into between-query variance (task difficulty) and within-query variance (agent inconsistency), highlighting whether reported results reflect true capability or measurement noise. We evaluated on GAIA (Levels 1-3, measuring agentic capabilities across varying reasoning complexity) and FRAMES (measuring retrieval and factuality across multiple documents). We found that ICC varies dramatically with task structure, with reasoning and retrieval tasks (FRAMES) exhibit ICC=0.4955-0.7118 across models, and agentic tasks (GAIA) exhibiting ICC=0.304-0.774 across models. For sub-agent replacement decisions in agentic systems, accuracy improvements are only trustworthy if ICC also improves. We demonstrate that ICC converges by n=8-16 trials for structured tasks and n>=32 for complex reasoning, enabling practitioners to set evidence-based resampling budgets. We recommend reporting accuracy alongside ICC and within-query variance as standard practice, and propose updated Evaluation Cards capturing these metrics. By making evaluation stability visible, we aim to transform agentic benchmarking from opaque leaderboard competition to trustworthy experimental science. Our code is open-sourced at https://github.com/youdotcom-oss/stochastic-agent-evals.

Junxian He, Wojciech Kryściński, Bryan McCann et al.

Current summarization systems yield generic summaries that are disconnected from users' preferences and expectations. To address this limitation, we present CTRLsum, a novel framework for controllable summarization. Our approach enables users to control multiple aspects of generated summaries by interacting with the summarization system through textual input in the form of a set of keywords or descriptive prompts. Using a single unified model, CTRLsum is able to achieve a broad scope of summary manipulation at inference time without requiring additional human annotations or pre-defining a set of control aspects during training. We quantitatively demonstrate the effectiveness of our approach on three domains of summarization datasets and five control aspects: 1) entity-centric and 2) length-controllable summarization, 3) contribution summarization on scientific papers, 4) invention purpose summarization on patent filings, and 5) question-guided summarization on news articles in a reading comprehension setting. Moreover, when used in a standard, uncontrolled summarization setting, CTRLsum achieves state-of-the-art results on the CNN/DailyMail dataset. Code and model checkpoints are available at https://github.com/salesforce/ctrl-sum

Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney et al.

Large-scale language models show promising text generation capabilities, but users cannot easily control particular aspects of the generated text. We release CTRL, a 1.63 billion-parameter conditional transformer language model, trained to condition on control codes that govern style, content, and task-specific behavior. Control codes were derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning while providing more explicit control over text generation. These codes also allow CTRL to predict which parts of the training data are most likely given a sequence. This provides a potential method for analyzing large amounts of data via model-based source attribution. We have released multiple full-sized, pretrained versions of CTRL at https://github.com/salesforce/ctrl.

Tianxing He, Bryan McCann, Caiming Xiong et al.

In this work, we explore joint energy-based model (EBM) training during the finetuning of pretrained text encoders (e.g., Roberta) for natural language understanding (NLU) tasks. Our experiments show that EBM training can help the model reach a better calibration that is competitive to strong baselines, with little or no loss in accuracy. We discuss three variants of energy functions (namely scalar, hidden, and sharp-hidden) that can be defined on top of a text encoder, and compare them in experiments. Due to the discreteness of text data, we adopt noise contrastive estimation (NCE) to train the energy-based model. To make NCE training more effective, we train an auto-regressive noise model with the masked language model (MLM) objective.

Hiroaki Hayashi, Wojciech Kryściński, Bryan McCann et al.

With thousands of academic articles shared on a daily basis, it has become increasingly difficult to keep up with the latest scientific findings. To overcome this problem, we introduce a new task of disentangled paper summarization, which seeks to generate separate summaries for the paper contributions and the context of the work, making it easier to identify the key findings shared in articles. For this purpose, we extend the S2ORC corpus of academic articles, which spans a diverse set of domains ranging from economics to psychology, by adding disentangled "contribution" and "context" reference labels. Together with the dataset, we introduce and analyze three baseline approaches: 1) a unified model controlled by input code prefixes, 2) a model with separate generation heads specialized in generating the disentangled outputs, and 3) a training strategy that guides the model using additional supervision coming from inbound and outbound citations. We also propose a comprehensive automatic evaluation protocol which reports the relevance, novelty, and disentanglement of generated outputs. Through a human study involving expert annotators, we show that in 79%, of cases our new task is considered more helpful than traditional scientific paper summarization.

Gustavo Aguilar, Bryan McCann, Tong Niu et al.

Byte-pair encoding (BPE) is a ubiquitous algorithm in the subword tokenization process of language models as it provides multiple benefits. However, this process is solely based on pre-training data statistics, making it hard for the tokenizer to handle infrequent spellings. On the other hand, though robust to misspellings, pure character-level models often lead to unreasonably long sequences and make it harder for the model to learn meaningful words. To alleviate these challenges, we propose a character-based subword module (char2subword) that learns the subword embedding table in pre-trained models like BERT. Our char2subword module builds representations from characters out of the subword vocabulary, and it can be used as a drop-in replacement of the subword embedding table. The module is robust to character-level alterations such as misspellings, word inflection, casing, and punctuation. We integrate it further with BERT through pre-training while keeping BERT transformer parameters fixed--and thus, providing a practical method. Finally, we show that incorporating our module to mBERT significantly improves the performance on the social media linguistic code-switching evaluation (LinCE) benchmark.

Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann et al.

While large-scale language models (LMs) are able to imitate the distribution of natural language well enough to generate realistic text, it is difficult to control which regions of the distribution they generate. This is especially problematic because datasets used for training large LMs usually contain significant toxicity, hate, bias, and negativity. We propose GeDi as an efficient method for using smaller LMs as generative discriminators to guide generation from large LMs to make them safer and more controllable. GeDi guides generation at each step by computing classification probabilities for all possible next tokens via Bayes rule by normalizing over two class-conditional distributions; one conditioned on the desired attribute, or control code, and another conditioned on the undesired attribute, or anti control code. We find that GeDi gives stronger controllability than the state of the art method while also achieving generation speeds more than 30 times faster. Additionally, training GeDi on only four topics allows us to controllably generate new topics zero-shot from just a keyword, unlocking a new capability that previous controllable generation methods do not have. Lastly, we show that GeDi can make GPT-2 (1.5B parameters) significantly less toxic without sacrificing linguistic quality, making it by far the most practical existing method for detoxifying large language models while maintaining a fast generation speed.

Alexander R. Fabbri, Wojciech Kryściński, Bryan McCann et al.

The scarcity of comprehensive up-to-date studies on evaluation metrics for text summarization and the lack of consensus regarding evaluation protocols continue to inhibit progress. We address the existing shortcomings of summarization evaluation methods along five dimensions: 1) we re-evaluate 14 automatic evaluation metrics in a comprehensive and consistent fashion using neural summarization model outputs along with expert and crowd-sourced human annotations, 2) we consistently benchmark 23 recent summarization models using the aforementioned automatic evaluation metrics, 3) we assemble the largest collection of summaries generated by models trained on the CNN/DailyMail news dataset and share it in a unified format, 4) we implement and share a toolkit that provides an extensible and unified API for evaluating summarization models across a broad range of automatic metrics, 5) we assemble and share the largest and most diverse, in terms of model types, collection of human judgments of model-generated summaries on the CNN/Daily Mail dataset annotated by both expert judges and crowd-source workers. We hope that this work will help promote a more complete evaluation protocol for text summarization as well as advance research in developing evaluation metrics that better correlate with human judgments.

Tianlu Wang, Xi Victoria Lin, Nazneen Fatema Rajani et al.

Word embeddings derived from human-generated corpora inherit strong gender bias which can be further amplified by downstream models. Some commonly adopted debiasing approaches, including the seminal Hard Debias algorithm, apply post-processing procedures that project pre-trained word embeddings into a subspace orthogonal to an inferred gender subspace. We discover that semantic-agnostic corpus regularities such as word frequency captured by the word embeddings negatively impact the performance of these algorithms. We propose a simple but effective technique, Double Hard Debias, which purifies the word embeddings against such corpus regularities prior to inferring and removing the gender subspace. Experiments on three bias mitigation benchmarks show that our approach preserves the distributional semantics of the pre-trained word embeddings while reducing gender bias to a significantly larger degree than prior approaches.

Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu et al.

Task-oriented dialogue is often decomposed into three tasks: understanding user input, deciding actions, and generating a response. While such decomposition might suggest a dedicated model for each sub-task, we find a simple, unified approach leads to state-of-the-art performance on the MultiWOZ dataset. SimpleTOD is a simple approach to task-oriented dialogue that uses a single, causal language model trained on all sub-tasks recast as a single sequence prediction problem. This allows SimpleTOD to fully leverage transfer learning from pre-trained, open domain, causal language models such as GPT-2. SimpleTOD improves over the prior state-of-the-art in joint goal accuracy for dialogue state tracking, and our analysis reveals robustness to noisy annotations in this setting. SimpleTOD also improves the main metrics used to evaluate action decisions and response generation in an end-to-end setting: inform rate by 8.1 points, success rate by 9.7 points, and combined score by 7.2 points.

Ali Madani, Bryan McCann, Nikhil Naik et al.

Generative modeling for protein engineering is key to solving fundamental problems in synthetic biology, medicine, and material science. We pose protein engineering as an unsupervised sequence generation problem in order to leverage the exponentially growing set of proteins that lack costly, structural annotations. We train a 1.2B-parameter language model, ProGen, on ~280M protein sequences conditioned on taxonomic and keyword tags such as molecular function and cellular component. This provides ProGen with an unprecedented range of evolutionary sequence diversity and allows it to generate with fine-grained control as demonstrated by metrics based on primary sequence similarity, secondary structure accuracy, and conformational energy.

Wojciech Kryściński, Bryan McCann, Caiming Xiong et al.

Currently used metrics for assessing summarization algorithms do not account for whether summaries are factually consistent with source documents. We propose a weakly-supervised, model-based approach for verifying factual consistency and identifying conflicts between source documents and a generated summary. Training data is generated by applying a series of rule-based transformations to the sentences of source documents. The factual consistency model is then trained jointly for three tasks: 1) identify whether sentences remain factually consistent after transformation, 2) extract a span in the source documents to support the consistency prediction, 3) extract a span in the summary sentence that is inconsistent if one exists. Transferring this model to summaries generated by several state-of-the art models reveals that this highly scalable approach substantially outperforms previous models, including those trained with strong supervision using standard datasets for natural language inference and fact checking. Additionally, human evaluation shows that the auxiliary span extraction tasks provide useful assistance in the process of verifying factual consistency.

Wojciech Kryściński, Nitish Shirish Keskar, Bryan McCann et al.

Text summarization aims at compressing long documents into a shorter form that conveys the most important parts of the original document. Despite increased interest in the community and notable research effort, progress on benchmark datasets has stagnated. We critically evaluate key ingredients of the current research setup: datasets, evaluation metrics, and models, and highlight three primary shortcomings: 1) automatically collected datasets leave the task underconstrained and may contain noise detrimental to training and evaluation, 2) current evaluation protocol is weakly correlated with human judgment and does not account for important characteristics such as factual correctness, 3) models overfit to layout biases of current datasets and offer limited diversity in their outputs.

Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong et al.

Deep learning models perform poorly on tasks that require commonsense reasoning, which often necessitates some form of world-knowledge or reasoning over information not immediately present in the input. We collect human explanations for commonsense reasoning in the form of natural language sequences and highlighted annotations in a new dataset called Common Sense Explanations (CoS-E). We use CoS-E to train language models to automatically generate explanations that can be used during training and inference in a novel Commonsense Auto-Generated Explanation (CAGE) framework. CAGE improves the state-of-the-art by 10% on the challenging CommonsenseQA task. We further study commonsense reasoning in DNNs using both human and auto-generated explanations including transfer to out-of-domain tasks. Empirical results indicate that we can effectively leverage language models for commonsense reasoning.

Jasdeep Singh, Bryan McCann, Nitish Shirish Keskar et al.

While natural language processing systems often focus on a single language, multilingual transfer learning has the potential to improve performance, especially for low-resource languages. We introduce XLDA, cross-lingual data augmentation, a method that replaces a segment of the input text with its translation in another language. XLDA enhances performance of all 14 tested languages of the cross-lingual natural language inference (XNLI) benchmark. With improvements of up to $4.8\%$, training with XLDA achieves state-of-the-art performance for Greek, Turkish, and Urdu. XLDA is in contrast to, and performs markedly better than, a more naive approach that aggregates examples in various languages in a way that each example is solely in one language. On the SQuAD question answering task, we see that XLDA provides a $1.0\%$ performance increase on the English evaluation set. Comprehensive experiments suggest that most languages are effective as cross-lingual augmentors, that XLDA is robust to a wide range of translation quality, and that XLDA is even more effective for randomly initialized models than for pretrained models.

Nitish Shirish Keskar, Bryan McCann, Caiming Xiong et al.

Even as pre-trained language encoders such as BERT are shared across many tasks, the output layers of question answering, text classification, and regression models are significantly different. Span decoders are frequently used for question answering, fixed-class, classification layers for text classification, and similarity-scoring layers for regression tasks, We show that this distinction is not necessary and that all three can be unified as span extraction. A unified, span-extraction approach leads to superior or comparable performance in supplementary supervised pre-trained, low-data, and multi-task learning experiments on several question answering, text classification, and regression benchmarks.

Bryan McCann, Nitish Shirish Keskar, Caiming Xiong et al.

Deep learning has improved performance on many natural language processing (NLP) tasks individually. However, general NLP models cannot emerge within a paradigm that focuses on the particularities of a single metric, dataset, and task. We introduce the Natural Language Decathlon (decaNLP), a challenge that spans ten tasks: question answering, machine translation, summarization, natural language inference, sentiment analysis, semantic role labeling, zero-shot relation extraction, goal-oriented dialogue, semantic parsing, and commonsense pronoun resolution. We cast all tasks as question answering over a context. Furthermore, we present a new Multitask Question Answering Network (MQAN) jointly learns all tasks in decaNLP without any task-specific modules or parameters in the multitask setting. MQAN shows improvements in transfer learning for machine translation and named entity recognition, domain adaptation for sentiment analysis and natural language inference, and zero-shot capabilities for text classification. We demonstrate that the MQAN's multi-pointer-generator decoder is key to this success and performance further improves with an anti-curriculum training strategy. Though designed for decaNLP, MQAN also achieves state of the art results on the WikiSQL semantic parsing task in the single-task setting. We also release code for procuring and processing data, training and evaluating models, and reproducing all experiments for decaNLP.

Stephen Merity, Bryan McCann, Richard Socher

Recurrent neural networks (RNNs) serve as a fundamental building block for many sequence tasks across natural language processing. Recent research has focused on recurrent dropout techniques or custom RNN cells in order to improve performance. Both of these can require substantial modifications to the machine learning model or to the underlying RNN configurations. We revisit traditional regularization techniques, specifically L2 regularization on RNN activations and slowness regularization over successive hidden states, to improve the performance of RNNs on the task of language modeling. Both of these techniques require minimal modification to existing RNN architectures and result in performance improvements comparable or superior to more complicated regularization techniques or custom cell architectures. These regularization techniques can be used without any modification on optimized LSTM implementations such as the NVIDIA cuDNN LSTM.

Bryan McCann, James Bradbury, Caiming Xiong et al.

Computer vision has benefited from initializing multiple deep layers with weights pretrained on large supervised training sets like ImageNet. Natural language processing (NLP) typically sees initialization of only the lowest layer of deep models with pretrained word vectors. In this paper, we use a deep LSTM encoder from an attentional sequence-to-sequence model trained for machine translation (MT) to contextualize word vectors. We show that adding these context vectors (CoVe) improves performance over using only unsupervised word and character vectors on a wide variety of common NLP tasks: sentiment analysis (SST, IMDb), question classification (TREC), entailment (SNLI), and question answering (SQuAD). For fine-grained sentiment analysis and entailment, CoVe improves performance of our baseline models to the state of the art.