Emily Dinan

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Weiyan Shi, Emily Dinan, Kurt Shuster et al. · meta-ai

Deployed dialogue agents have the potential to integrate human feedback to continuously improve themselves. However, humans may not always provide explicit signals when the chatbot makes mistakes during interactions. In this work, we propose Juicer, a framework to make use of both binary and free-form textual human feedback. It works by: (i) extending sparse binary feedback by training a satisfaction classifier to label the unlabeled data; and (ii) training a reply corrector to map the bad replies to good ones. We find that augmenting training with model-corrected replies improves the final dialogue model, and we can further improve performance by using both positive and negative replies through the recently proposed Director model.

Weiyan Shi, Emily Dinan, Adi Renduchintala et al. · cmu

Existing approaches built separate classifiers to detect nonsense in dialogues. In this paper, we show that without external classifiers, dialogue models can detect errors in their own messages introspectively, by calculating the likelihood of replies that are indicative of poor messages. For example, if an agent believes its partner is likely to respond "I don't understand" to a candidate message, that message may not make sense, so an alternative message should be chosen. We evaluate our approach on a dataset from the game Diplomacy, which contains long dialogues richly grounded in the game state, on which existing models make many errors. We first show that hand-crafted replies can be effective for the task of detecting nonsense in applications as complex as Diplomacy. We then design AutoReply, an algorithm to search for such discriminative replies automatically, given a small number of annotated dialogue examples. We find that AutoReply-generated replies outperform handcrafted replies and perform on par with carefully fine-tuned large supervised models. Results also show that one single reply without much computation overheads can also detect dialogue nonsense reasonably well.

Yushi Hu, Reyhane Askari-Hemmat, Melissa Hall et al.

Reward models (RMs) are essential for training large language models (LLMs), but remain underexplored for omni models that handle interleaved image and text sequences. We introduce Multimodal RewardBench 2 (MMRB2), the first comprehensive benchmark for reward models on multimodal understanding and (interleaved) generation. MMRB2 spans four tasks: text-to-image, image editing, interleaved generation, and multimodal reasoning ("thinking-with-images"), providing 1,000 expert-annotated preference pairs per task from 23 models and agents across 21 source tasks. MMRB2 is designed with: (1) practical but challenging prompts; (2) responses from state-of-the-art models and agents; and (3) preference pairs with strong human-expert consensus, curated via an ensemble filtering strategy. Using MMRB2, we study existing judges for each subtask, including multimodal LLM-as-a-judge and models trained with human preferences. The latest Gemini 3 Pro attains 75-80% accuracy. GPT-5 and Gemini 2.5 Pro reach 66-75% accuracy, compared to >90% for humans, yet surpass the widely used GPT-4o (59%). The best performing open-source model Qwen3-VL-32B achieves similar accuracies as Gemini 2.5 Flash (64%). We also show that MMRB2 performance strongly correlates with downstream task success using Best-of-N sampling and conduct an in-depth analysis that shows key areas to improve the reward models going forward.

Xiaochuang Han, Youssef Emad, Melissa Hall et al. · meta-ai

Video generation models are rapidly advancing, but can still struggle with complex video outputs that require significant semantic branching or repeated high-level reasoning about what should happen next. In this paper, we introduce a new class of omni video-text models that integrate ideas from recent LM reasoning advances to address this challenge. More specifically, we present TV2TV, a unified generative modeling framework which decomposes video generation into an interleaved text and video generation process. TV2TV jointly learns language modeling (next-token prediction) and video flow matching (next-frame prediction) using a Mixture-of-Transformers (MoT) architecture. At inference time, TV2TV decides when to alternate between generating text and video frames, allowing the model to "think in words" about subsequent content before ``acting in pixels'' to produce frames. This design offloads much of the responsibility for deciding what should happen next to the language modeling tower, enabling improved visual quality and prompt alignment of generated videos. It also enables fine-grained controllability, allowing users to modify the video generation trajectory through text interventions at any point in the process. In controlled experiments on video game data, TV2TV demonstrates substantial improvements in both visual quality and controllability. TV2TV also scales to natural videos, as we show by augmenting sports videos with interleaved natural language action descriptions using vision-language models (VLMs). Training TV2TV on this corpus yields strong visual quality and prompt alignment, showcasing the model's ability to reason about and generate complex real-world action sequences. Together, these results highlight TV2TV as a promising step toward video generation with open-ended textual reasoning and control.

Andrew Lee, David Wu, Emily Dinan et al.

Despite many recent advancements in language modeling, state-of-the-art language models lack grounding in the real world and struggle with tasks involving complex reasoning. Meanwhile, advances in the symbolic reasoning capabilities of AI have led to systems that outperform humans in games like chess and Go (Silver et al., 2018). Chess commentary provides an interesting domain for bridging these two fields of research, as it requires reasoning over a complex board state and providing analyses in natural language. In this work we demonstrate how to combine symbolic reasoning engines with controllable language models to generate chess commentaries. We conduct experiments to demonstrate that our approach generates commentaries that are preferred by human judges over previous baselines.

Emily Dinan, Sho Yaida, Susan Zhang

We perform an effective-theory analysis of forward-backward signal propagation in wide and deep Transformers, i.e., residual neural networks with multi-head self-attention blocks and multilayer perceptron blocks. This analysis suggests particular width scalings of initialization and training hyperparameters for these models. We then take up such suggestions, training Vision and Language Transformers in practical setups.

Emily Dinan, Samuel Humeau, Bharath Chintagunta et al.

The detection of offensive language in the context of a dialogue has become an increasingly important application of natural language processing. The detection of trolls in public forums (Galán-García et al., 2016), and the deployment of chatbots in the public domain (Wolf et al., 2017) are two examples that show the necessity of guarding against adversarially offensive behavior on the part of humans. In this work, we develop a training scheme for a model to become robust to such human attacks by an iterative build it, break it, fix it strategy with humans and models in the loop. In detailed experiments we show this approach is considerably more robust than previous systems. Further, we show that offensive language used within a conversation critically depends on the dialogue context, and cannot be viewed as a single sentence offensive detection task as in most previous work. Our newly collected tasks and methods will be made open source and publicly available.

Qizhen Zhang, Prajjwal Bhargava, Chloe Bi et al. · allen-ai

We present Branch-Train-Stitch (BTS), an efficient and flexible training algorithm for combining independently trained large language model (LLM) experts into a single, capable generalist model. Following Li et al., we start with a single seed language model which is branched into domain-specific (e.g., coding or math) experts with continual pretraining. BTS combines experts into a generalist model using lightweight stitch layers, which are inserted between frozen experts and the seed LLM, and trained on a small datamix of the expert domains. Stitch layers enable the seed LLM to integrate representations from any number of experts during the forward pass, allowing it to generalize to new domains, despite remaining frozen. Because BTS does not alter the constituent LLMs, BTS provides a modular and flexible approach: experts can be easily removed and new experts can be added with only a small amount of training. Compared to alternative model merging approaches, BTS yields the best generalist performance on a variety of downstream tasks, retaining the specialized capabilities of each of the experts.

Jiahui Chen, Philippe Hansen-Estruch, Xiaochuang Han et al.

Unified multimodal generation architectures that jointly produce text and images have recently emerged as a promising direction for text-to-image (T2I) synthesis. However, many existing systems rely on explicit modality switching, generating reasoning text before switching manually to image generation. This separate, sequential inference process limits cross-modal coupling and prohibits automatic multimodal generation. This work explores post-training to achieve fully unified text-image generation, where models autonomously transition from textual reasoning to visual synthesis within a single inference process. We examine the impact of joint text-image generation on T2I performance and the relative importance of each modality during post-training. We additionally explore different post-training data strategies, showing that a targeted dataset addressing specific limitations achieves superior results compared to broad image-caption corpora or benchmark-aligned data. Using offline, reward-weighted post-training with fully self-generated synthetic data, our approach enables improvements in multimodal image generation across four diverse T2I benchmarks, demonstrating the effectiveness of reward-weighting both modalities and strategically designed post-training data.

Raghuveer Thirukovalluru, Xiaochuang Han, Bhuwan Dhingra et al.

Image encoders, a fundamental component of vision-language models (VLMs), are typically pretrained independently before being aligned with a language model. This standard paradigm results in encoders that process images agnostically, without regard to the specific downstream task or text query. To address this limitation, we propose the Text-Guided Semantic Image Encoder (TIE), which generates image representations conditioned on the input text query. VLMs equipped with TIE outperform their conventional counterparts by +1.5 and +1.3 points on average across nine image-to-text benchmarks at the 1B and 3B scales, respectively, with gains reaching up to 6 points on tasks such as DocVQA and InfoVQA. Moreover, TIE-based VLMs attain superior performance while utilizing only half as many image tiles (tokens), resulting in notably improved inference efficiency. TIE also generalizes well with generic queries, indicating that text-conditioned training effectively optimizes the encoder to capture key visual features. Qualitative analysis confirms that TIE consistently attends to query-relevant regions, enhancing both interpretability and query-specific grounding.

Emily Dinan, Gavin Abercrombie, A. Stevie Bergman et al.

Over the last several years, end-to-end neural conversational agents have vastly improved in their ability to carry a chit-chat conversation with humans. However, these models are often trained on large datasets from the internet, and as a result, may learn undesirable behaviors from this data, such as toxic or otherwise harmful language. Researchers must thus wrestle with the issue of how and when to release these models. In this paper, we survey the problem landscape for safety for end-to-end conversational AI and discuss recent and related work. We highlight tensions between values, potential positive impact and potential harms, and provide a framework for making decisions about whether and how to release these models, following the tenets of value-sensitive design. We additionally provide a suite of tools to enable researchers to make better-informed decisions about training and releasing end-to-end conversational AI models.

Sabrina J. Mielke, Arthur Szlam, Emily Dinan et al.

While improving neural dialogue agents' factual accuracy is the object of much research, another important aspect of communication, less studied in the setting of neural dialogue, is transparency about ignorance. In this work, we analyze to what extent state-of-the-art chit-chat models are linguistically calibrated in the sense that their verbalized expression of doubt (or confidence) matches the likelihood that the model's responses are factually incorrect (or correct). We find that these models are poorly calibrated, yet we show that likelihood of correctness can accurately be predicted. By incorporating such metacognitive features into the training of a controllable generation model, we obtain a dialogue agent with greatly improved linguistic calibration. While improving neural dialogue agents' factual accuracy is the object of much research, another important aspect of communication, less studied in the setting of neural dialogue, is transparency about ignorance. In this work, we analyze to what extent state-of-the-art chit-chat models are linguistically calibrated in the sense that their verbalized expression of doubt (or confidence) matches the likelihood that the model's responses are factually incorrect (or correct). We find that these models are poorly calibrated, yet we show that likelihood of correctness can accurately be predicted. By incorporating such metacognitive features into the training of a controllable generation model, we obtain a dialogue agent with greatly improved linguistic calibration.

Jing Xu, Da Ju, Margaret Li et al.

Models trained on large unlabeled corpora of human interactions will learn patterns and mimic behaviors therein, which include offensive or otherwise toxic behavior and unwanted biases. We investigate a variety of methods to mitigate these issues in the context of open-domain generative dialogue models. We introduce a new human-and-model-in-the-loop framework for both training safer models and for evaluating them, as well as a novel method to distill safety considerations inside generative models without the use of an external classifier at deployment time. We conduct experiments comparing these methods and find our new techniques are (i) safer than existing models as measured by automatic and human evaluations while (ii) maintaining usability metrics such as engagingness relative to the state of the art. We then discuss the limitations of this work by analyzing failure cases of our models.

Eric Michael Smith, Diana Gonzalez-Rico, Emily Dinan et al.

Open-domain conversation models have become good at generating natural-sounding dialogue, using very large architectures with billions of trainable parameters. The vast training data required to train these architectures aggregates many different styles, tones, and qualities. Using that data to train a single model makes it difficult to use the model as a consistent conversational agent, e.g. with a stable set of persona traits and a typical style of expression. Several architectures affording control mechanisms over generation architectures have been proposed, each with different trade-offs. However, it remains unclear whether their use in dialogue is viable, and what the trade-offs look like with the most recent state-of-the-art conversational architectures. In this work, we adapt three previously proposed controllable generation architectures to open-domain dialogue generation, controlling the style of the generation to match one among about 200 possible styles. We compare their respective performance and tradeoffs, and show how they can be used to provide insights into existing conversational datasets, and generate a varied set of styled conversation replies.

Kurt Shuster, Jack Urbanek, Emily Dinan et al.

Much of NLP research has focused on crowdsourced static datasets and the supervised learning paradigm of training once and then evaluating test performance. As argued in de Vries et al. (2020), crowdsourced data has the issues of lack of naturalness and relevance to real-world use cases, while the static dataset paradigm does not allow for a model to learn from its experiences of using language (Silver et al., 2013). In contrast, one might hope for machine learning systems that become more useful as they interact with people. In this work, we build and deploy a role-playing game, whereby human players converse with learning agents situated in an open-domain fantasy world. We show that by training models on the conversations they have with humans in the game the models progressively improve, as measured by automatic metrics and online engagement scores. This learning is shown to be more efficient than crowdsourced data when applied to conversations with real users, as well as being far cheaper to collect.

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.

Emily Dinan, Angela Fan, Ledell Wu et al.

Machine learning models are trained to find patterns in data. NLP models can inadvertently learn socially undesirable patterns when training on gender biased text. In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions: bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker. Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information. In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites. Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers. We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models, detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.

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.

Shrimai Prabhumoye, Margaret Li, Jack Urbanek et al.

Dialogue research tends to distinguish between chit-chat and goal-oriented tasks. While the former is arguably more naturalistic and has a wider use of language, the latter has clearer metrics and a straightforward learning signal. Humans effortlessly combine the two, for example engaging in chit-chat with the goal of exchanging information or eliciting a specific response. Here, we bridge the divide between these two domains in the setting of a rich multi-player text-based fantasy environment where agents and humans engage in both actions and dialogue. Specifically, we train a goal-oriented model with reinforcement learning against an imitation-learned ``chit-chat'' model with two approaches: the policy either learns to pick a topic or learns to pick an utterance given the top-K utterances from the chit-chat model. We show that both models outperform an inverse model baseline and can converse naturally with their dialogue partner in order to achieve goals.

Angela Fan, Jack Urbanek, Pratik Ringshia et al.

Procedurally generating cohesive and interesting game environments is challenging and time-consuming. In order for the relationships between the game elements to be natural, common-sense has to be encoded into arrangement of the elements. In this work, we investigate a machine learning approach for world creation using content from the multi-player text adventure game environment LIGHT. We introduce neural network based models to compositionally arrange locations, characters, and objects into a coherent whole. In addition to creating worlds based on existing elements, our models can generate new game content. Humans can also leverage our models to interactively aid in worldbuilding. We show that the game environments created with our approach are cohesive, diverse, and preferred by human evaluators compared to other machine learning based world construction algorithms.

Eric Michael Smith, Diana Gonzalez-Rico, Emily Dinan et al.

Text style transfer is usually performed using attributes that can take a handful of discrete values (e.g., positive to negative reviews). In this work, we introduce an architecture that can leverage pre-trained consistent continuous distributed style representations and use them to transfer to an attribute unseen during training, without requiring any re-tuning of the style transfer model. We demonstrate the method by training an architecture to transfer text conveying one sentiment to another sentiment, using a fine-grained set of over 20 sentiment labels rather than the binary positive/negative often used in style transfer. Our experiments show that this model can then rewrite text to match a target sentiment that was unseen during training.

Emily Dinan, Angela Fan, Adina Williams et al.

Models often easily learn biases present in the training data, and their predictions directly reflect this bias. We analyze gender bias in dialogue data, and examine how this bias is actually amplified in subsequent generative chit-chat dialogue models. We measure gender bias in six existing dialogue datasets, and focus on the most biased one, the multi-player text-based fantasy adventure dataset LIGHT, as a testbed for our bias mitigation techniques. The LIGHT dataset is highly imbalanced with respect to gender, containing predominantly male characters, likely because it is entirely collected by crowdworkers and reflects common biases that exist in fantasy or medieval settings. We consider three techniques to mitigate gender bias: counterfactual data augmentation, targeted data collection, and bias controlled training. We show that our proposed techniques mitigate gender bias in LIGHT by balancing the genderedness of generated dialogue utterances and are particularly effective in combination. We quantify performance using various evaluation methods---such as quantity of gendered words, a dialogue safety classifier, and human studies---all of which show that our models generate less gendered, but equally engaging chit-chat responses.

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.

Yixin Nie, Adina Williams, Emily Dinan et al.

We introduce a new large-scale NLI benchmark dataset, collected via an iterative, adversarial human-and-model-in-the-loop procedure. We show that training models on this new dataset leads to state-of-the-art performance on a variety of popular NLI benchmarks, while posing a more difficult challenge with its new test set. Our analysis sheds light on the shortcomings of current state-of-the-art models, and shows that non-expert annotators are successful at finding their weaknesses. The data collection method can be applied in a never-ending learning scenario, becoming a moving target for NLU, rather than a static benchmark that will quickly saturate.

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.

Jack Urbanek, Angela Fan, Siddharth Karamcheti et al.

We introduce a large scale crowdsourced text adventure game as a research platform for studying grounded dialogue. In it, agents can perceive, emote, and act whilst conducting dialogue with other agents. Models and humans can both act as characters within the game. We describe the results of training state-of-the-art generative and retrieval models in this setting. We show that in addition to using past dialogue, these models are able to effectively use the state of the underlying world to condition their predictions. In particular, we show that grounding on the details of the local environment, including location descriptions, and the objects (and their affordances) and characters (and their previous actions) present within it allows better predictions of agent behavior and dialogue. We analyze the ingredients necessary for successful grounding in this setting, and how each of these factors relate to agents that can talk and act successfully.

Emily Dinan, Varvara Logacheva, Valentin Malykh et al.

We describe the setting and results of the ConvAI2 NeurIPS competition that aims to further the state-of-the-art in open-domain chatbots. Some key takeaways from the competition are: (i) pretrained Transformer variants are currently the best performing models on this task, (ii) but to improve performance on multi-turn conversations with humans, future systems must go beyond single word metrics like perplexity to measure the performance across sequences of utterances (conversations) -- in terms of repetition, consistency and balance of dialogue acts (e.g. how many questions asked vs. answered).

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.

Jason Weston, Emily Dinan, Alexander H. Miller

Sequence generation models for dialogue are known to have several problems: they tend to produce short, generic sentences that are uninformative and unengaging. Retrieval models on the other hand can surface interesting responses, but are restricted to the given retrieval set leading to erroneous replies that cannot be tuned to the specific context. In this work we develop a model that combines the two approaches to avoid both their deficiencies: first retrieve a response and then refine it -- the final sequence generator treating the retrieval as additional context. We show on the recent CONVAI2 challenge task our approach produces responses superior to both standard retrieval and generation models in human evaluations.

Saizheng Zhang, Emily Dinan, Jack Urbanek et al.

Chit-chat models are known to have several problems: they lack specificity, do not display a consistent personality and are often not very captivating. In this work we present the task of making chit-chat more engaging by conditioning on profile information. We collect data and train models to (i) condition on their given profile information; and (ii) information about the person they are talking to, resulting in improved dialogues, as measured by next utterance prediction. Since (ii) is initially unknown our model is trained to engage its partner with personal topics, and we show the resulting dialogue can be used to predict profile information about the interlocutors.