Stephen Roller

Susan Zhang, Stephen Roller, Naman Goyal et al. · meta-ai, uw

Large language models, which are often trained for hundreds of thousands of compute days, have shown remarkable capabilities for zero- and few-shot learning. Given their computational cost, these models are difficult to replicate without significant capital. For the few that are available through APIs, no access is granted to the full model weights, making them difficult to study. We present Open Pre-trained Transformers (OPT), a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters, which we aim to fully and responsibly share with interested researchers. We show that OPT-175B is comparable to GPT-3, while requiring only 1/7th the carbon footprint to develop. We are also releasing our logbook detailing the infrastructure challenges we faced, along with code for experimenting with all of the released models.

Kurt Shuster, Jing Xu, Mojtaba Komeili et al. · meta-ai, mila

We present BlenderBot 3, a 175B parameter dialogue model capable of open-domain conversation with access to the internet and a long-term memory, and having been trained on a large number of user defined tasks. We release both the model weights and code, and have also deployed the model on a public web page to interact with organic users. This technical report describes how the model was built (architecture, model and training scheme), and details of its deployment, including safety mechanisms. Human evaluations show its superiority to existing open-domain dialogue agents, including its predecessors (Roller et al., 2021; Komeili et al., 2022). Finally, we detail our plan for continual learning using the data collected from deployment, which will also be publicly released. The goal of this research program is thus to enable the community to study ever-improving responsible agents that learn through interaction.

Kurt Shuster, Mojtaba Komeili, Leonard Adolphs et al. · meta-ai

Language models (LMs) have recently been shown to generate more factual responses by employing modularity (Zhou et al., 2021) in combination with retrieval (Adolphs et al., 2021). We extend the recent approach of Adolphs et al. (2021) to include internet search as a module. Our SeeKeR (Search engine->Knowledge->Response) method thus applies a single LM to three modular tasks in succession: search, generating knowledge, and generating a final response. We show that, when using SeeKeR as a dialogue model, it outperforms the state-of-the-art model BlenderBot 2 (Chen et al., 2021) on open-domain knowledge-grounded conversations for the same number of parameters, in terms of consistency, knowledge and per-turn engagingness. SeeKeR applied to topical prompt completions as a standard language model outperforms GPT2 (Radford et al., 2019) and GPT3 (Brown et al., 2020) in terms of factuality and topicality, despite GPT3 being a vastly larger model. Our code and models are made publicly available.

Armen Aghajanyan, Lili Yu, Alexis Conneau et al. · uw

Generative language models define distributions over sequences of tokens that can represent essentially any combination of data modalities (e.g., any permutation of image tokens from VQ-VAEs, speech tokens from HuBERT, BPE tokens for language or code, and so on). To better understand the scaling properties of such mixed-modal models, we conducted over 250 experiments using seven different modalities and model sizes ranging from 8 million to 30 billion, trained on 5-100 billion tokens. We report new mixed-modal scaling laws that unify the contributions of individual modalities and the interactions between them. Specifically, we explicitly model the optimal synergy and competition due to data and model size as an additive term to previous uni-modal scaling laws. We also find four empirical phenomena observed during the training, such as emergent coordinate-ascent style training that naturally alternates between modalities, guidelines for selecting critical hyper-parameters, and connections between mixed-modal competition and training stability. Finally, we test our scaling law by training a 30B speech-text model, which significantly outperforms the corresponding unimodal models. Overall, our research provides valuable insights into the design and training of mixed-modal generative models, an important new class of unified models that have unique distributional properties.

Igor Molybog, Peter Albert, Moya Chen et al. · meta-ai

We present a theory for the previously unexplained divergent behavior noticed in the training of large language models. We argue that the phenomenon is an artifact of the dominant optimization algorithm used for training, called Adam. We observe that Adam can enter a state in which the parameter update vector has a relatively large norm and is essentially uncorrelated with the direction of descent on the training loss landscape, leading to divergence. This artifact is more likely to be observed in the training of a deep model with a large batch size, which is the typical setting of large-scale language model training. To argue the theory, we present observations from the training runs of the language models of different scales: 7 billion, 30 billion, 65 billion, and 546 billion parameters.

Jianguo Zhang, Stephen Roller, Kun Qian et al. · salesforce

End-to-end task-oriented dialogue (TOD) systems have achieved promising performance by leveraging sophisticated natural language understanding and natural language generation capabilities of pre-trained models. This work enables the TOD systems with more flexibility through a simple cache. The cache provides the flexibility to dynamically update the TOD systems and handle both existing and unseen dialogue scenarios. Towards this end, we first fine-tune a retrieval module to effectively retrieve the most relevant information entries from the cache. We then train end-to-end TOD models that can refer to and ground on both dialogue history and retrieved information during TOD generation. The cache is straightforward to construct, and the backbone models of TOD systems are compatible with existing pre-trained generative models. Extensive experiments demonstrate the superior performance of our framework, with a notable improvement in non-empty joint goal accuracy by 6.7% compared to strong baselines.

Richard Yuanzhe Pang, Stephen Roller, Kyunghyun Cho et al.

We study improving social conversational agents by learning from natural dialogue between users and a deployed model, without extra annotations. To implicitly measure the quality of a machine-generated utterance, we leverage signals like user response length, sentiment and reaction of the future human utterances in the collected dialogue episodes. Our experiments use the publicly released deployment data from BlenderBot (Xu et al., 2023). Human evaluation indicates improvements in our new models over baseline responses; however, we find that some proxy signals can lead to more generations with undesirable properties as well. For example, optimizing for conversation length can lead to more controversial or unfriendly generations compared to the baseline, whereas optimizing for positive sentiment or reaction can decrease these behaviors.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Margaret Li, Jason Weston, Stephen Roller

While dialogue remains an important end-goal of natural language research, the difficulty of evaluation is an oft-quoted reason why it remains troublesome to make real progress towards its solution. Evaluation difficulties are actually two-fold: not only do automatic metrics not correlate well with human judgments, but also human judgments themselves are in fact difficult to measure. The two most used human judgment tests, single-turn pairwise evaluation and multi-turn Likert scores, both have serious flaws as we discuss in this work. We instead provide a novel procedure involving comparing two full dialogues, where a human judge is asked to pay attention to only one speaker within each, and make a pairwise judgment. The questions themselves are optimized to maximize the robustness of judgments across different annotators, resulting in better tests. We also show how these tests work in self-play model chat setups, resulting in faster, cheaper tests. We hope these tests become the de facto standard, and will release open-source code to that end.

Eric Michael Smith, Orion Hsu, Rebecca Qian et al.

At the heart of improving conversational AI is the open problem of how to evaluate conversations. Issues with automatic metrics are well known (Liu et al., 2016, arXiv:1603.08023), with human evaluations still considered the gold standard. Unfortunately, how to perform human evaluations is also an open problem: differing data collection methods have varying levels of human agreement and statistical sensitivity, resulting in differing amounts of human annotation hours and labor costs. In this work we compare five different crowdworker-based human evaluation methods and find that different methods are best depending on the types of models compared, with no clear winner across the board. While this highlights the open problems in the area, our analysis leads to advice of when to use which one, and possible future directions.

Moya Chen, Paul A. Crook, Stephen Roller

We demonstrate that large language models are able to simulate Task Oriented Dialogues in novel domains, provided only with an API implementation and a list of goals. We show these simulations can formulate online, automatic metrics that correlate well with human evaluations. Furthermore, by checking for whether the User's goals are met, we can use simulation to repeatedly generate training data and improve the quality of simulations themselves. With no human intervention or domain-specific training data, our simulations bootstrap end-to-end models which achieve a 37\% error reduction in previously unseen domains. By including as few as 32 domain-specific conversations, bootstrapped models can match the performance of a fully-supervised model with $10\times$ more data. To our knowledge, this is the first time simulations have been shown to be effective at bootstrapping models without explicitly requiring any domain-specific training data, rule-engineering, or humans-in-the-loop.

Stephen Roller, Sainbayar Sukhbaatar, Arthur Szlam et al.

We investigate the training of sparse layers that use different parameters for different inputs based on hashing in large Transformer models. Specifically, we modify the feedforward layer to hash to different sets of weights depending on the current token, over all tokens in the sequence. We show that this procedure either outperforms or is competitive with learning-to-route mixture-of-expert methods such as Switch Transformers and BASE Layers, while requiring no routing parameters or extra terms in the objective function such as a load balancing loss, and no sophisticated assignment algorithm. We study the performance of different hashing techniques, hash sizes and input features, and show that balanced and random hashes focused on the most local features work best, compared to either learning clusters or using longer-range context. We show our approach works well both on large language modeling and dialogue tasks, and on downstream fine-tuning tasks.

Da Ju, Stephen Roller, Sainbayar Sukhbaatar et al.

Attention mechanisms have become a standard tool for sequence modeling tasks, in particular by stacking self-attention layers over the entire input sequence as in the Transformer architecture. In this work we introduce a novel attention procedure called staircase attention that, unlike self-attention, operates across the sequence (in time) recurrently processing the input by adding another step of processing. A step in the staircase comprises of backward tokens (encoding the sequence so far seen) and forward tokens (ingesting a new part of the sequence), or an extreme Ladder version with a forward step of zero that simply repeats the Transformer on each step of the ladder, sharing the weights. We thus describe a family of such models that can trade off performance and compute, by either increasing the amount of recurrence through time, the amount of sequential processing via recurrence in depth, or both. Staircase attention is shown to be able to solve tasks that involve tracking that conventional Transformers cannot, due to this recurrence. Further, it is shown to provide improved modeling power for the same size model (number of parameters) compared to self-attentive Transformers on large language modeling and dialogue tasks, yielding significant perplexity gains.

Sainbayar Sukhbaatar, Da Ju, Spencer Poff et al.

Attention mechanisms have shown promising results in sequence modeling tasks that require long-term memory. Recent work investigated mechanisms to reduce the computational cost of preserving and storing memories. However, not all content in the past is equally important to remember. We propose Expire-Span, a method that learns to retain the most important information and expire the irrelevant information. This forgetting of memories enables Transformers to scale to attend over tens of thousands of previous timesteps efficiently, as not all states from previous timesteps are preserved. We demonstrate that Expire-Span can help models identify and retain critical information and show it can achieve strong performance on reinforcement learning tasks specifically designed to challenge this functionality. Next, we show that Expire-Span can scale to memories that are tens of thousands in size, setting a new state of the art on incredibly long context tasks such as character-level language modeling and a frame-by-frame moving objects task. Finally, we analyze the efficiency of Expire-Span compared to existing approaches and demonstrate that it trains faster and uses less memory.

Kai Sun, Seungwhan Moon, Paul Crook et al.

Existing dialogue corpora and models are typically designed under two disjoint motives: while task-oriented systems focus on achieving functional goals (e.g., booking hotels), open-domain chatbots aim at making socially engaging conversations. In this work, we propose to integrate both types of systems by Adding Chit-Chat to ENhance Task-ORiented dialogues (ACCENTOR), with the goal of making virtual assistant conversations more engaging and interactive. Specifically, we propose a Human <-> AI collaborative data collection approach for generating diverse chit-chat responses to augment task-oriented dialogues with minimal annotation effort. We then present our new chit-chat-based annotations to 23.8K dialogues from two popular task-oriented datasets (Schema-Guided Dialogue and MultiWOZ 2.1) and demonstrate their advantage over the originals via human evaluation. Lastly, we propose three new models for adding chit-chat to task-oriented dialogues, explicitly trained to predict user goals and to generate contextually relevant chit-chat responses. Automatic and human evaluations show that, compared with the state-of-the-art task-oriented baseline, our models can code-switch between task and chit-chat to be more engaging, interesting, knowledgeable, and humanlike, while maintaining competitive task performance.

Stephen Roller, Y-Lan Boureau, Jason Weston et al.

We present our view of what is necessary to build an engaging open-domain conversational agent: covering the qualities of such an agent, the pieces of the puzzle that have been built so far, and the gaping holes we have not filled yet. We present a biased view, focusing on work done by our own group, while citing related work in each area. In particular, we discuss in detail the properties of continual learning, providing engaging content, and being well-behaved -- and how to measure success in providing them. We end with a discussion of our experience and learnings, and our recommendations to the community.

Stephen Roller, Emily Dinan, Naman Goyal et al.

Building open-domain chatbots is a challenging area for machine learning research. While prior work has shown that scaling neural models in the number of parameters and the size of the data they are trained on gives improved results, we show that other ingredients are important for a high-performing chatbot. Good conversation requires a number of skills that an expert conversationalist blends in a seamless way: providing engaging talking points and listening to their partners, and displaying knowledge, empathy and personality appropriately, while maintaining a consistent persona. We show that large scale models can learn these skills when given appropriate training data and choice of generation strategy. We build variants of these recipes with 90M, 2.7B and 9.4B parameter models, and make our models and code publicly available. Human evaluations show our best models are superior to existing approaches in multi-turn dialogue in terms of engagingness and humanness measurements. We then discuss the limitations of this work by analyzing failure cases of our models.

Margaret Li, Stephen Roller, Ilia Kulikov et al.

Generative dialogue models currently suffer from a number of problems which standard maximum likelihood training does not address. They tend to produce generations that (i) rely too much on copying from the context, (ii) contain repetitions within utterances, (iii) overuse frequent words, and (iv) at a deeper level, contain logical flaws. In this work we show how all of these problems can be addressed by extending the recently introduced unlikelihood loss (Welleck et al., 2019) to these cases. We show that appropriate loss functions which regularize generated outputs to match human distributions are effective for the first three issues. For the last important general issue, we show applying unlikelihood to collected data of what a model should not do is effective for improving logical consistency, potentially paving the way to generative models with greater reasoning ability. We demonstrate the efficacy of our approach across several dialogue tasks.

Kurt Shuster, Da Ju, Stephen Roller et al.

We introduce dodecaDialogue: a set of 12 tasks that measures if a conversational agent can communicate engagingly with personality and empathy, ask questions, answer questions by utilizing knowledge resources, discuss topics and situations, and perceive and converse about images. By multi-tasking on such a broad large-scale set of data, we hope to both move towards and measure progress in producing a single unified agent that can perceive, reason and converse with humans in an open-domain setting. We show that such multi-tasking improves over a BERT pre-trained baseline, largely due to multi-tasking with very large dialogue datasets in a similar domain, and that the multi-tasking in general provides gains to both text and image-based tasks using several metrics in both the fine-tune and task transfer settings. We obtain state-of-the-art results on many of the tasks, providing a strong baseline for this challenge.

Sean Welleck, Ilia Kulikov, Stephen Roller et al.

Neural text generation is a key tool in natural language applications, but it is well known there are major problems at its core. In particular, standard likelihood training and decoding leads to dull and repetitive outputs. While some post-hoc fixes have been proposed, in particular top-$k$ and nucleus sampling, they do not address the fact that the token-level probabilities predicted by the model are poor. In this paper we show that the likelihood objective itself is at fault, resulting in a model that assigns too much probability to sequences containing repeats and frequent words, unlike those from the human training distribution. We propose a new objective, unlikelihood training, which forces unlikely generations to be assigned lower probability by the model. We show that both token and sequence level unlikelihood training give less repetitive, less dull text while maintaining perplexity, giving superior generations using standard greedy or beam search. According to human evaluations, our approach with standard beam search also outperforms the currently popular decoding methods of nucleus sampling or beam blocking, thus providing a strong alternative to existing techniques.

Abigail See, Stephen Roller, Douwe Kiela et al.

A good conversation requires balance -- between simplicity and detail; staying on topic and changing it; asking questions and answering them. Although dialogue agents are commonly evaluated via human judgments of overall quality, the relationship between quality and these individual factors is less well-studied. In this work, we examine two controllable neural text generation methods, conditional training and weighted decoding, in order to control four important attributes for chitchat dialogue: repetition, specificity, response-relatedness and question-asking. We conduct a large-scale human evaluation to measure the effect of these control parameters on multi-turn interactive conversations on the PersonaChat task. We provide a detailed analysis of their relationship to high-level aspects of conversation, and show that by controlling combinations of these variables our models obtain clear improvements in human quality judgments.

Matt Le, Stephen Roller, Laetitia Papaxanthos et al.

We consider the task of inferring is-a relationships from large text corpora. For this purpose, we propose a new method combining hyperbolic embeddings and Hearst patterns. This approach allows us to set appropriate constraints for inferring concept hierarchies from distributional contexts while also being able to predict missing is-a relationships and to correct wrong extractions. Moreover -- and in contrast with other methods -- the hierarchical nature of hyperbolic space allows us to learn highly efficient representations and to improve the taxonomic consistency of the inferred hierarchies. Experimentally, we show that our approach achieves state-of-the-art performance on several commonly-used benchmarks.

Emily Dinan, Stephen Roller, Kurt Shuster et al.

In open-domain dialogue intelligent agents should exhibit the use of knowledge, however there are few convincing demonstrations of this to date. The most popular sequence to sequence models typically "generate and hope" generic utterances that can be memorized in the weights of the model when mapping from input utterance(s) to output, rather than employing recalled knowledge as context. Use of knowledge has so far proved difficult, in part because of the lack of a supervised learning benchmark task which exhibits knowledgeable open dialogue with clear grounding. To that end we collect and release a large dataset with conversations directly grounded with knowledge retrieved from Wikipedia. We then design architectures capable of retrieving knowledge, reading and conditioning on it, and finally generating natural responses. Our best performing dialogue models are able to conduct knowledgeable discussions on open-domain topics as evaluated by automatic metrics and human evaluations, while our new benchmark allows for measuring further improvements in this important research direction.

Stephen Roller, Douwe Kiela, Maximilian Nickel

Methods for unsupervised hypernym detection may broadly be categorized according to two paradigms: pattern-based and distributional methods. In this paper, we study the performance of both approaches on several hypernymy tasks and find that simple pattern-based methods consistently outperform distributional methods on common benchmark datasets. Our results show that pattern-based models provide important contextual constraints which are not yet captured in distributional methods.

Su Wang, Stephen Roller, Katrin Erk

We test whether distributional models can do one-shot learning of definitional properties from text only. Using Bayesian models, we find that first learning overarching structure in the known data, regularities in textual contexts and in properties, helps one-shot learning, and that individual context items can be highly informative. Our experiments show that our model can learn properties from a single exposure when given an informative utterance.

Stephen Roller, Katrin Erk

We consider the task of predicting lexical entailment using distributional vectors. We perform a novel qualitative analysis of one existing model which was previously shown to only measure the prototypicality of word pairs. We find that the model strongly learns to identify hypernyms using Hearst patterns, which are well known to be predictive of lexical relations. We present a novel model which exploits this behavior as a method of feature extraction in an iterative procedure similar to Principal Component Analysis. Our model combines the extracted features with the strengths of other proposed models in the literature, and matches or outperforms prior work on multiple data sets.

Ye Zhang, Stephen Roller, Byron Wallace

We introduce a novel, simple convolution neural network (CNN) architecture - multi-group norm constraint CNN (MGNC-CNN) that capitalizes on multiple sets of word embeddings for sentence classification. MGNC-CNN extracts features from input embedding sets independently and then joins these at the penultimate layer in the network to form a final feature vector. We then adopt a group regularization strategy that differentially penalizes weights associated with the subcomponents generated from the respective embedding sets. This model is much simpler than comparable alternative architectures and requires substantially less training time. Furthermore, it is flexible in that it does not require input word embeddings to be of the same dimensionality. We show that MGNC-CNN consistently outperforms baseline models.

I. Beltagy, Stephen Roller, Pengxiang Cheng et al.

NLP tasks differ in the semantic information they require, and at this time no single se- mantic representation fulfills all requirements. Logic-based representations characterize sentence structure, but do not capture the graded aspect of meaning. Distributional models give graded similarity ratings for words and phrases, but do not capture sentence structure in the same detail as logic-based approaches. So it has been argued that the two are complementary. We adopt a hybrid approach that combines logic-based and distributional semantics through probabilistic logic inference in Markov Logic Networks (MLNs). In this paper, we focus on the three components of a practical system integrating logical and distributional models: 1) Parsing and task representation is the logic-based part where input problems are represented in probabilistic logic. This is quite different from representing them in standard first-order logic. 2) For knowledge base construction we form weighted inference rules. We integrate and compare distributional information with other sources, notably WordNet and an existing paraphrase collection. In particular, we use our system to evaluate distributional lexical entailment approaches. We use a variant of Robinson resolution to determine the necessary inference rules. More sources can easily be added by mapping them to logical rules; our system learns a resource-specific weight that corrects for scaling differences between resources. 3) In discussing probabilistic inference, we show how to solve the inference problems efficiently. To evaluate our approach, we use the task of textual entailment (RTE), which can utilize the strengths of both logic-based and distributional representations. In particular we focus on the SICK dataset, where we achieve state-of-the-art results.