Khyathi Raghavi Chandu

Yizhong Wang, Hamish Ivison, Pradeep Dasigi et al. · allen-ai, uw

In this work we explore recent advances in instruction-tuning language models on a range of open instruction-following datasets. Despite recent claims that open models can be on par with state-of-the-art proprietary models, these claims are often accompanied by limited evaluation, making it difficult to compare models across the board and determine the utility of various resources. We provide a large set of instruction-tuned models from 6.7B to 65B parameters in size, trained on 12 instruction datasets ranging from manually curated (e.g., OpenAssistant) to synthetic and distilled (e.g., Alpaca) and systematically evaluate them on their factual knowledge, reasoning, multilinguality, coding, and open-ended instruction following abilities through a collection of automatic, model-based, and human-based metrics. We further introduce Tülu, our best performing instruction-tuned model suite finetuned on a combination of high-quality open resources. Our experiments show that different instruction-tuning datasets can uncover or enhance specific skills, while no single dataset (or combination) provides the best performance across all evaluations. Interestingly, we find that model and human preference-based evaluations fail to reflect differences in model capabilities exposed by benchmark-based evaluations, suggesting the need for the type of systemic evaluation performed in this work. Our evaluations show that the best model in any given evaluation reaches on average 87% of ChatGPT performance, and 73% of GPT-4 performance, suggesting that further investment in building better base models and instruction-tuning data is required to close the gap. We release our instruction-tuned models, including a fully finetuned 65B Tülu, along with our code, data, and evaluation framework at https://github.com/allenai/open-instruct to facilitate future research.

Da Yin, Faeze Brahman, Abhilasha Ravichander et al. · allen-ai, cmu

Closed-source agents suffer from several issues such as a lack of affordability, transparency, and reproducibility, particularly on complex interactive tasks. This motivates the development of open-source alternatives. We introduce LUMOS, one of the first frameworks for training open-source LLM-based agents. LUMOS features a learnable, unified, and modular architecture with a planning module that learns high-level subgoal generation, and a grounding module trained to translate these into actions using various tools in the execution module. The design allows for modular upgrades and wider applicability to diverse interactive tasks. To foster generalizable agent learning, we collect large-scale, unified, and high-quality training annotations derived from diverse ground-truth reasoning rationales across various complex interactive tasks. On 9 datasets, LUMOS exhibits several key advantages: (1) LUMOS excels multiple larger open-source agents on the held-out datasets (unused for training) for each task type. LUMOS even surpasses GPT agents on QA and web tasks; (2) LUMOS outperforms open-source agents produced by chain-of-thoughts and unmodularized integrated training; and (3) LUMOS effectively generalizes to unseen tasks, outperforming 33B-scale agents and domain-specific agents.

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Peter West, Ximing Lu, Nouha Dziri et al. · allen-ai, cmu

The recent wave of generative AI has sparked unprecedented global attention, with both excitement and concern over potentially superhuman levels of artificial intelligence: models now take only seconds to produce outputs that would challenge or exceed the capabilities even of expert humans. At the same time, models still show basic errors in understanding that would not be expected even in non-expert humans. This presents us with an apparent paradox: how do we reconcile seemingly superhuman capabilities with the persistence of errors that few humans would make? In this work, we posit that this tension reflects a divergence in the configuration of intelligence in today's generative models relative to intelligence in humans. Specifically, we propose and test the Generative AI Paradox hypothesis: generative models, having been trained directly to reproduce expert-like outputs, acquire generative capabilities that are not contingent upon -- and can therefore exceed -- their ability to understand those same types of outputs. This contrasts with humans, for whom basic understanding almost always precedes the ability to generate expert-level outputs. We test this hypothesis through controlled experiments analyzing generation vs. understanding in generative models, across both language and image modalities. Our results show that although models can outperform humans in generation, they consistently fall short of human capabilities in measures of understanding, as well as weaker correlation between generation and understanding performance, and more brittleness to adversarial inputs. Our findings support the hypothesis that models' generative capability may not be contingent upon understanding capability, and call for caution in interpreting artificial intelligence by analogy to human intelligence.

Faeze Brahman, Sachin Kumar, Vidhisha Balachandran et al. · allen-ai, cmu

Chat-based language models are designed to be helpful, yet they should not comply with every user request. While most existing work primarily focuses on refusal of "unsafe" queries, we posit that the scope of noncompliance should be broadened. We introduce a comprehensive taxonomy of contextual noncompliance describing when and how models should not comply with user requests. Our taxonomy spans a wide range of categories including incomplete, unsupported, indeterminate, and humanizing requests (in addition to unsafe requests). To test noncompliance capabilities of language models, we use this taxonomy to develop a new evaluation suite of 1000 noncompliance prompts. We find that most existing models show significantly high compliance rates in certain previously understudied categories with models like GPT-4 incorrectly complying with as many as 30% of requests. To address these gaps, we explore different training strategies using a synthetically-generated training set of requests and expected noncompliant responses. Our experiments demonstrate that while direct finetuning of instruction-tuned models can lead to both over-refusal and a decline in general capabilities, using parameter efficient methods like low rank adapters helps to strike a good balance between appropriate noncompliance and other capabilities.

Lining Zhang, Simon Mille, Yufang Hou et al. · deepmind, uw

To prevent the costly and inefficient use of resources on low-quality annotations, we want a method for creating a pool of dependable annotators who can effectively complete difficult tasks, such as evaluating automatic summarization. Thus, we investigate the recruitment of high-quality Amazon Mechanical Turk workers via a two-step pipeline. We show that we can successfully filter out subpar workers before they carry out the evaluations and obtain high-agreement annotations with similar constraints on resources. Although our workers demonstrate a strong consensus among themselves and CloudResearch workers, their alignment with expert judgments on a subset of the data is not as expected and needs further training in correctness. This paper still serves as a best practice for the recruitment of qualified annotators in other challenging annotation tasks.

Khyathi Raghavi Chandu, Linjie Li, Anas Awadalla et al. · allen-ai, uw

The ability to acknowledge the inevitable uncertainty in their knowledge and reasoning is a prerequisite for AI systems to be truly truthful and reliable. In this paper, we present a taxonomy of uncertainty specific to vision-language AI systems, distinguishing between epistemic uncertainty (arising from a lack of information) and aleatoric uncertainty (due to inherent unpredictability), and further explore finer categories within. Based on this taxonomy, we synthesize a benchmark dataset, CertainlyUncertain, featuring 178K visual question answering (VQA) samples as contrastive pairs. This is achieved by 1) inpainting images to make previously answerable questions into unanswerable ones; and 2) using image captions to prompt large language models for both answerable and unanswerable questions. Additionally, we introduce a new metric confidence-weighted accuracy, that is well correlated with both accuracy and calibration error, to address the shortcomings of existing metrics.

Luca Soldaini, Rodney Kinney, Akshita Bhagia et al. · allen-ai, cmu

Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them. As a result, it is challenging to conduct and advance scientific research on language modeling, such as understanding how training data impacts model capabilities and limitations. To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents. We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices. Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation.

Nathan Lambert, Valentina Pyatkin, Jacob Morrison et al. · allen-ai, cambridge

Reward models (RMs) are at the crux of successfully using RLHF to align pretrained models to human preferences, yet there has been relatively little study that focuses on evaluation of those models. Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models and which values are embedded in them. Resources for reward model training and understanding are sparse in the nascent open-source community around them. To enhance scientific understanding of reward models, we present RewardBench, a benchmark dataset and code-base for evaluation. The RewardBench dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety, to benchmark how reward models perform on challenging, structured and out-of-distribution queries. We create specific comparison datasets for RMs that have subtle, but verifiable reasons (e.g. bugs, incorrect facts) why one answer should be preferred to another. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods, such as the direct MLE training of classifiers and the implicit reward modeling of Direct Preference Optimization (DPO). We present many findings on propensity for refusals, reasoning limitations, and instruction following shortcomings of various reward models towards a better understanding of the RLHF process.

Khyathi Raghavi Chandu, Alborz Geramifard

Contextualizing language technologies beyond a single language kindled embracing multiple modalities and languages. Individually, each of these directions undoubtedly proliferated into several NLP tasks. Despite this momentum, most of the multimodal research is primarily centered around English and multilingual research is primarily centered around contexts from text modality. Challenging this conventional setup, researchers studied the unification of multilingual and multimodal (MultiX) streams. The main goal of this work is to catalogue and characterize these works by charting out the categories of tasks, datasets and methods to address MultiX scenarios. To this end, we review the languages studied, gold or silver data with parallel annotations, and understand how these modalities and languages interact in modeling. We present an account of the modeling approaches along with their strengths and weaknesses to better understand what scenarios they can be used reliably. Following this, we present the high-level trends in the overall paradigm of the field. Finally, we conclude by presenting a road map of challenges and promising research directions.

Khyathi Raghavi Chandu, Alborz Geramifard

Pre-trained models with dual and cross encoders have shown remarkable success in propelling the landscape of several tasks in vision and language in Visual Question Answering (VQA). However, since they are limited by the requirements of gold annotated data, most of these advancements do not see the light of day in other languages beyond English. We aim to address this problem by introducing a curriculum based on the source and target language translations to finetune the pre-trained models for the downstream task. Experimental results demonstrate that script plays a vital role in the performance of these models. Specifically, we show that target languages that share the same script perform better (~6%) than other languages and mixed-script code-switched languages perform better than their counterparts (~5-12%).

Tejas Srinivasan, Jack Hessel, Tanmay Gupta et al. · allen-ai

Selective prediction minimizes incorrect predictions from vision-language models (VLMs) by allowing them to abstain from answering when uncertain. However, when deploying a vision-language system with low tolerance for inaccurate predictions, selective prediction may be over-cautious and abstain too frequently, even on many correct predictions. We introduce ReCoVERR, an inference-time algorithm to reduce the over-abstention of a selective vision-language system without increasing the error rate of the system's predictions. When the VLM makes a low-confidence prediction, instead of abstaining ReCoVERR tries to find relevant clues in the image that provide additional evidence for the prediction. ReCoVERR uses an LLM to pose related questions to the VLM, collects high-confidence evidences, and if enough evidence confirms the prediction the system makes a prediction instead of abstaining. ReCoVERR enables three VLMs (BLIP2, InstructBLIP, and LLaVA-1.5) to answer up to 20% more questions on the VQAv2 and A-OKVQA tasks without decreasing system accuracy, thus improving overall system reliability. Our code is available at https://github.com/tejas1995/ReCoVERR.

Dirk Groeneveld, Iz Beltagy, Pete Walsh et al. · allen-ai, cmu

Language models (LMs) have become ubiquitous in both NLP research and in commercial product offerings. As their commercial importance has surged, the most powerful models have become closed off, gated behind proprietary interfaces, with important details of their training data, architectures, and development undisclosed. Given the importance of these details in scientifically studying these models, including their biases and potential risks, we believe it is essential for the research community to have access to powerful, truly open LMs. To this end, we have built OLMo, a competitive, truly Open Language Model, to enable the scientific study of language models. Unlike most prior efforts that have only released model weights and inference code, we release OLMo alongside open training data and training and evaluation code. We hope this release will empower the open research community and inspire a new wave of innovation.

Anthony Sicilia, Hyunwoo Kim, Khyathi Raghavi Chandu et al. · allen-ai, nvidia

Effective interlocutors account for the uncertain goals, beliefs, and emotions of others. But even the best human conversationalist cannot perfectly anticipate the trajectory of a dialogue. How well can language models represent inherent uncertainty in conversations? We propose FortUne Dial, an expansion of the long-standing "conversation forecasting" task: instead of just accuracy, evaluation is conducted with uncertainty-aware metrics, effectively enabling abstention on individual instances. We study two ways in which language models potentially represent outcome uncertainty (internally, using scores and directly, using tokens) and propose fine-tuning strategies to improve calibration of both representations. Experiments on eight difficult negotiation corpora demonstrate that our proposed fine-tuning strategies (a traditional supervision strategy and an off-policy reinforcement learning strategy) can calibrate smaller open-source models to compete with pre-trained models 10x their size.

Bill Yuchen Lin, Abhilasha Ravichander, Ximing Lu et al. · allen-ai, cmu

The alignment tuning process of large language models (LLMs) typically involves instruction learning through supervised fine-tuning (SFT) and preference tuning via reinforcement learning from human feedback (RLHF). A recent study, LIMA (Zhou et al. 2023), shows that using merely 1K examples for SFT can achieve significant alignment performance as well, suggesting that the effect of alignment tuning might be "superficial." This raises questions about how exactly the alignment tuning transforms a base LLM. We analyze the effect of alignment tuning by examining the token distribution shift between base LLMs and their aligned counterpart. Our findings reveal that base LLMs and their alignment-tuned versions perform nearly identically in decoding on the majority of token positions. Most distribution shifts occur with stylistic tokens. These direct evidence strongly supports the Superficial Alignment Hypothesis suggested by LIMA. Based on these findings, we rethink the alignment of LLMs by posing the research question: how effectively can we align base LLMs without SFT or RLHF? To address this, we introduce a simple, tuning-free alignment method, URIAL. URIAL achieves effective alignment purely through in-context learning (ICL) with base LLMs, requiring as few as three constant stylistic examples and a system prompt. We conduct a fine-grained and interpretable evaluation on a diverse set of examples, named JUST-EVAL-INSTRUCT. Results demonstrate that base LLMs with URIAL can match or even surpass the performance of LLMs aligned with SFT or SFT+RLHF. We show that the gap between tuning-free and tuning-based alignment methods can be significantly reduced through strategic prompting and ICL. Our findings on the superficial nature of alignment tuning and results with URIAL suggest that deeper analysis and theoretical understanding of alignment is crucial to future LLM research.

Peter West, Ronan Le Bras, Taylor Sorensen et al.

We present NovaCOMET, an open commonsense knowledge model, that combines the best aspects of knowledge and general task models. Compared to previous knowledge models, NovaCOMET allows open-format relations enabling direct application to reasoning tasks; compared to general task models like Flan-T5, it explicitly centers knowledge, enabling superior performance for commonsense reasoning. NovaCOMET leverages the knowledge of opaque proprietary models to create an open knowledge pipeline. First, knowledge is symbolically distilled into NovATOMIC, a publicly-released discrete knowledge graph which can be audited, critiqued, and filtered. Next, we train NovaCOMET on NovATOMIC by fine-tuning an open-source pretrained model. NovaCOMET uses an open-format training objective, replacing the fixed relation sets of past knowledge models, enabling arbitrary structures within the data to serve as inputs or outputs. The resulting generation model, optionally augmented with human annotation, matches or exceeds comparable open task models like Flan-T5 on a range of commonsense generation tasks. NovaCOMET serves as a counterexample to the contemporary focus on instruction tuning only, demonstrating a distinct advantage to explicitly modeling commonsense knowledge as well.

Parul Chopra, Sai Krishna Rallabandi, Alan W Black et al.

Code-switching (CS), a ubiquitous phenomenon due to the ease of communication it offers in multilingual communities still remains an understudied problem in language processing. The primary reasons behind this are: (1) minimal efforts in leveraging large pretrained multilingual models, and (2) the lack of annotated data. The distinguishing case of low performance of multilingual models in CS is the intra-sentence mixing of languages leading to switch points. We first benchmark two sequence labeling tasks -- POS and NER on 4 different language pairs with a suite of pretrained models to identify the problems and select the best performing model, char-BERT, among them (addressing (1)). We then propose a self training method to repurpose the existing pretrained models using a switch-point bias by leveraging unannotated data (addressing (2)). We finally demonstrate that our approach performs well on both tasks by reducing the gap between the switch point performance while retaining the overall performance on two distinct language pairs in both the tasks. Our code is available here: https://github.com/PC09/EMNLP2021-Switch-Point-biased-Self-Training.

Sai Muralidhar Jayanthi, Kavya Nerella, Khyathi Raghavi Chandu et al.

The NLP community has witnessed steep progress in a variety of tasks across the realms of monolingual and multilingual language processing recently. These successes, in conjunction with the proliferating mixed language interactions on social media have boosted interest in modeling code-mixed texts. In this work, we present CodemixedNLP, an open-source library with the goals of bringing together the advances in code-mixed NLP and opening it up to a wider machine learning community. The library consists of tools to develop and benchmark versatile model architectures that are tailored for mixed texts, methods to expand training sets, techniques to quantify mixing styles, and fine-tuned state-of-the-art models for 7 tasks in Hinglish. We believe this work has a potential to foster a distributed yet collaborative and sustainable ecosystem in an otherwise dispersed space of code-mixing research. The toolkit is designed to be simple, easily extensible, and resourceful to both researchers as well as practitioners.

Yutaro Yamada, Khyathi Chandu, Yuchen Lin et al. · allen-ai

Diffusion-based image generation models such as DALL-E 3 and Stable Diffusion-XL demonstrate remarkable capabilities in generating images with realistic and unique compositions. Yet, these models are not robust in precisely reasoning about physical and spatial configurations of objects, especially when instructed with unconventional, thereby out-of-distribution descriptions, such as "a chair with five legs". In this paper, we propose a language agent with chain-of-3D-thoughts (L3GO), an inference-time approach that can reason about part-based 3D mesh generation of unconventional objects that current data-driven diffusion models struggle with. More concretely, we use large language models as agents to compose a desired object via trial-and-error within the 3D simulation environment. To facilitate our investigation, we develop a new benchmark, Unconventionally Feasible Objects (UFO), as well as SimpleBlenv, a wrapper environment built on top of Blender where language agents can build and compose atomic building blocks via API calls. Human and automatic GPT-4V evaluations show that our approach surpasses the standard GPT-4 and other language agents (e.g., ReAct and Reflexion) for 3D mesh generation on ShapeNet. Moreover, when tested on our UFO benchmark, our approach outperforms other state-of-the-art text-to-2D image and text-to-3D models based on human evaluation.

Jae Sung Park, Jack Hessel, Khyathi Raghavi Chandu et al. · allen-ai, uw

Instruction following vision-language (VL) models offer a flexible interface that supports a broad range of multimodal tasks in a zero-shot fashion. However, interfaces that operate on full images do not directly enable the user to "point to" and access specific regions within images. This capability is important not only to support reference-grounded VL benchmarks, but also, for practical applications that require precise within-image reasoning. We build Localized Visual Commonsense models, which allow users to specify (multiple) regions as input. We train our model by sampling localized commonsense knowledge from a large language model (LLM): specifically, we prompt an LLM to collect commonsense knowledge given a global literal image description and a local literal region description automatically generated by a set of VL models. With a separately trained critic model that selects high-quality examples, we find that training on the localized commonsense corpus can successfully distill existing VL models to support a reference-as-input interface. Empirical results and human evaluations in a zero-shot setup demonstrate that our distillation method results in more precise VL models of reasoning compared to a baseline of passing a generated referring expression to an LLM.

Marcel Nawrath, Agnieszka Nowak, Tristan Ratz et al. · amazon-science

At the heart of the Pyramid evaluation method for text summarization lie human written summary content units (SCUs). These SCUs are concise sentences that decompose a summary into small facts. Such SCUs can be used to judge the quality of a candidate summary, possibly partially automated via natural language inference (NLI) systems. Interestingly, with the aim to fully automate the Pyramid evaluation, Zhang and Bansal (2021) show that SCUs can be approximated by automatically generated semantic role triplets (STUs). However, several questions currently lack answers, in particular: i) Are there other ways of approximating SCUs that can offer advantages? ii) Under which conditions are SCUs (or their approximations) offering the most value? In this work, we examine two novel strategies to approximate SCUs: generating SCU approximations from AMR meaning representations (SMUs) and from large language models (SGUs), respectively. We find that while STUs and SMUs are competitive, the best approximation quality is achieved by SGUs. We also show through a simple sentence-decomposition baseline (SSUs) that SCUs (and their approximations) offer the most value when ranking short summaries, but may not help as much when ranking systems or longer summaries.

Jae Sung Park, Zixian Ma, Linjie Li et al. · uw

Reasoning over visual relationships-spatial, functional, interactional, social, etc.-is considered to be a fundamental component of human cognition. Yet, despite the major advances in visual comprehension in multimodal language models (MLMs), precise reasoning over relationships and their generations remains a challenge. We introduce ROBIN: an MLM instruction-tuned with densely annotated relationships capable of constructing high-quality dense scene graphs at scale. To train ROBIN, we curate SVG, a synthetic scene graph dataset by completing the missing relations of selected objects in existing scene graphs using a teacher MLM and a carefully designed filtering process to ensure high-quality. To generate more accurate and rich scene graphs at scale for any image, we introduce SG-EDIT: a self-distillation framework where GPT-4o further refines ROBIN's predicted scene graphs by removing unlikely relations and/or suggesting relevant ones. In total, our dataset contains 146K images and 5.6M relationships for 2.6M objects. Results show that our ROBIN-3B model, despite being trained on less than 3 million instances, outperforms similar-size models trained on over 300 million instances on relationship understanding benchmarks, and even surpasses larger models up to 13B parameters. Notably, it achieves state-of-the-art performance in referring expression comprehension with a score of 88.9, surpassing the previous best of 87.4. Our results suggest that training on the refined scene graph data is crucial to maintaining high performance across diverse visual reasoning task.

Keivan Rezaei, Khyathi Chandu, Soheil Feizi et al. · cmu

Large language models trained on web-scale corpora can memorize undesirable data containing misinformation, copyrighted material, or private or sensitive information. Recently, several machine unlearning algorithms have been proposed to eliminate the effect of such datapoints from trained models -- that is, to approximate a model that had never been trained on these datapoints in the first place. However, evaluating the effectiveness of unlearning algorithms remains an open challenge. Previous work has relied on heuristics -- such as verifying that the model can no longer reproduce the specific information targeted for removal while maintaining accuracy on unrelated test data. These approaches inadequately capture the complete effect of reversing the influence of datapoints on a trained model. In this work, we propose the RESTOR framework for machine unlearning evaluation, which assesses the ability of unlearning algorithms for targeted data erasure, by evaluating the ability of models to forget the knowledge introduced in these datapoints, while simultaneously recovering the model's knowledge state had it never encountered these datapoints. RESTOR helps uncover several novel insights about popular unlearning algorithms, and the mechanisms through which they operate -- for instance, identifying that some algorithms merely emphasize forgetting but not recovering knowledge, and that localizing unlearning targets can enhance unlearning performance.

Jeffrey Li, Alex Fang, Georgios Smyrnis et al.

We introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments with the goal of improving language models. As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set. The resulting dataset, DCLM-Baseline enables training a 7B parameter language model from scratch to 64% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6.6 percentage point improvement on MMLU while being trained with 40% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63% & 66%), and performs similarly on an average of 53 natural language understanding tasks while being trained with 6.6x less compute than Llama 3 8B. Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation.

Bill Yuchen Lin, Yuntian Deng, Khyathi Chandu et al.

We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of ``slightly better/worse'' to ``tie'' if the winner response exceeds the loser one by more than $K$ characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard's 0.91 and AlpacaEval2.0's 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates.

Ximing Lu, Faeze Brahman, Peter West et al.

While extreme-scale language models have demonstrated exceptional performance on a variety of language tasks, the degree of control over these language models through pure prompting can often be limited. Directly fine-tuning such language models can be effective for tailoring them, but it can be either extremely costly (e.g., GPT-3) or not even feasible for the broader community (e.g., GPT-4). We propose Inference-time Policy Adapters (IPA), which efficiently tailors a language model such as GPT-3 without fine-tuning it. IPA guides a large base model during decoding time through a lightweight policy adapter trained to optimize an arbitrary user objective with reinforcement learning. On five challenging text generation tasks, such as toxicity reduction and lexically constrained generation, IPA consistently brings significant improvements over off-the-shelf language models. It outperforms competitive baseline methods, sometimes even including expensive fine-tuning. In particular, tailoring GPT-2 with IPA can outperform GPT-3, while tailoring GPT-3 with IPA brings a major performance boost over GPT-3 (and sometimes even over GPT-4). Our promising results highlight the potential of IPA as a lightweight alternative to tailoring extreme-scale language models.

Hyundong Cho, Andrea Madotto, Zhaojiang Lin et al.

Dialogue systems are frequently updated to accommodate new services, but naively updating them by continually training with data for new services in diminishing performance on previously learnt services. Motivated by the insight that dialogue state tracking (DST), a crucial component of dialogue systems that estimates the user's goal as a conversation proceeds, is a simple natural language understanding task, we propose reformulating it as a bundle of granular example-guided question answering tasks to minimize the task shift between services and thus benefit continual learning. Our approach alleviates service-specific memorization and teaches a model to contextualize the given question and example to extract the necessary information from the conversation. We find that a model with just 60M parameters can achieve a significant boost by learning to learn from in-context examples retrieved by a retriever trained to identify turns with similar dialogue state changes. Combining our method with dialogue-level memory replay, our approach attains state of the art performance on DST continual learning metrics without relying on any complex regularization or parameter expansion methods.

Khyathi Raghavi Chandu, Yonatan Bisk, Alan W Black

The NLP community has seen substantial recent interest in grounding to facilitate interaction between language technologies and the world. However, as a community, we use the term broadly to reference any linking of text to data or non-textual modality. In contrast, Cognitive Science more formally defines "grounding" as the process of establishing what mutual information is required for successful communication between two interlocutors -- a definition which might implicitly capture the NLP usage but differs in intent and scope. We investigate the gap between these definitions and seek answers to the following questions: (1) What aspects of grounding are missing from NLP tasks? Here we present the dimensions of coordination, purviews and constraints. (2) How is the term "grounding" used in the current research? We study the trends in datasets, domains, and tasks introduced in recent NLP conferences. And finally, (3) How to advance our current definition to bridge the gap with Cognitive Science? We present ways to both create new tasks or repurpose existing ones to make advancements towards achieving a more complete sense of grounding.

Sebastian Gehrmann, Tosin Adewumi, Karmanya Aggarwal et al.

We introduce GEM, a living benchmark for natural language Generation (NLG), its Evaluation, and Metrics. Measuring progress in NLG relies on a constantly evolving ecosystem of automated metrics, datasets, and human evaluation standards. Due to this moving target, new models often still evaluate on divergent anglo-centric corpora with well-established, but flawed, metrics. This disconnect makes it challenging to identify the limitations of current models and opportunities for progress. Addressing this limitation, GEM provides an environment in which models can easily be applied to a wide set of tasks and in which evaluation strategies can be tested. Regular updates to the benchmark will help NLG research become more multilingual and evolve the challenge alongside models. This paper serves as the description of the data for which we are organizing a shared task at our ACL 2021 Workshop and to which we invite the entire NLG community to participate.

Khyathi Raghavi Chandu, Ruo-Ping Dong, Alan Black

Generating long form narratives such as stories and procedures from multiple modalities has been a long standing dream for artificial intelligence. In this regard, there is often crucial subtext that is derived from the surrounding contexts. The general seq2seq training methods render the models shorthanded while attempting to bridge the gap between these neighbouring contexts. In this paper, we tackle this problem by using \textit{infilling} techniques involving prediction of missing steps in a narrative while generating textual descriptions from a sequence of images. We also present a new large scale \textit{visual procedure telling} (ViPT) dataset with a total of 46,200 procedures and around 340k pairwise images and textual descriptions that is rich in such contextual dependencies. Generating steps using infilling technique demonstrates the effectiveness in visual procedures with more coherent texts. We conclusively show a METEOR score of 27.51 on procedures which is higher than the state-of-the-art on visual storytelling. We also demonstrate the effects of interposing new text with missing images during inference. The code and the dataset will be publicly available at https://visual-narratives.github.io/Visual-Narratives/.

Khyathi Raghavi Chandu, Alan W Black

Neural text generation metamorphosed into several critical natural language applications ranging from text completion to free form narrative generation. In order to progress research in text generation, it is critical to absorb the existing research works and position ourselves in this massively growing field. Specifically, this paper surveys the fundamental components of modeling approaches relaying task agnostic impacts across various generation tasks such as storytelling, summarization, translation etc., In this context, we present an abstraction of the imperative techniques with respect to learning paradigms, pretraining, modeling approaches, decoding and the key challenges outstanding in the field in each of them. Thereby, we deliver a one-stop destination for researchers in the field to facilitate a perspective on where to situate their work and how it impacts other closely related generation tasks.

Khyathi Raghavi Chandu, Piyush Sharma, Soravit Changpinyo et al.

Training large-scale image captioning (IC) models demands access to a rich and diverse set of training examples, gathered from the wild, often from noisy alt-text data. However, recent modeling approaches to IC often fall short in terms of performance in this case, because they assume a clean annotated dataset (as opposed to the noisier alt-text--based annotations), and employ an end-to-end generation approach, which often lacks both controllability and interpretability. We address these problems by breaking down the task into two simpler, more controllable tasks -- skeleton prediction and skeleton-based caption generation. Specifically, we show that selecting content words as skeletons} helps in generating improved and denoised captions when leveraging rich yet noisy alt-text--based uncurated datasets. We also show that the predicted English skeletons can be further cross-lingually leveraged to generate non-English captions, and present experimental results covering caption generation in French, Italian, German, Spanish and Hindi. We also show that skeleton-based prediction allows for better control of certain caption properties, such as length, content, and gender expression, providing a handle to perform human-in-the-loop semi-automatic corrections.

Khyathi Raghavi Chandu, Alan W Black

Code-Switching (CS) is a common phenomenon observed in several bilingual and multilingual communities, thereby attaining prevalence in digital and social media platforms. This increasing prominence demands the need to model CS languages for critical downstream tasks. A major problem in this domain is the dearth of annotated data and a substantial corpora to train large scale neural models. Generating vast amounts of quality text assists several down stream tasks that heavily rely on language modeling such as speech recognition, text-to-speech synthesis etc,. We present a novel vantage point of CS to be style variations between both the participating languages. Our approach does not need any external annotations such as lexical language ids. It mainly relies on easily obtainable monolingual corpora without any parallel alignment and a limited set of naturally CS sentences. We propose a two-stage generative adversarial training approach where the first stage generates competitive negative examples for CS and the second stage generates more realistic CS sentences. We present our experiments on the following pairs of languages: Spanish-English, Mandarin-English, Hindi-English and Arabic-French. We show that the trends in metrics for generated CS move closer to real CS data in each of the above language pairs through the dual stage training process. We believe this viewpoint of CS as style variations opens new perspectives for modeling various tasks in CS text.

Ruo-Ping Dong, Khyathi Raghavi Chandu, Alan W Black

We are enveloped by stories of visual interpretations in our everyday lives. The way we narrate a story often comprises of two stages, which are, forming a central mind map of entities and then weaving a story around them. A contributing factor to coherence is not just basing the story on these entities but also, referring to them using appropriate terms to avoid repetition. In this paper, we address these two stages of introducing the right entities at seemingly reasonable junctures and also referring them coherently in the context of visual storytelling. The building blocks of the central mind map, also known as entity skeleton are entity chains including nominal and coreference expressions. This entity skeleton is also represented in different levels of abstractions to compose a generalized frame to weave the story. We build upon an encoder-decoder framework to penalize the model when the decoded story does not adhere to this entity skeleton. We establish a strong baseline for skeleton informed generation and then extend this to have the capability of multitasking by predicting the skeleton in addition to generating the story. Finally, we build upon this model and propose a glocal hierarchical attention model that attends to the skeleton both at the sentence (local) and the story (global) levels. We observe that our proposed models outperform the baseline in terms of automatic evaluation metric, METEOR. We perform various analysis targeted to evaluate the performance of our task of enforcing the entity skeleton such as the number and diversity of the entities generated. We also conduct human evaluation from which it is concluded that the visual stories generated by our model are preferred 82% of the times. In addition, we show that our glocal hierarchical attention model improves coherence by introducing more pronouns as required by the presence of nouns.

Shrimai Prabhumoye, Khyathi Raghavi Chandu, Ruslan Salakhutdinov et al.

Visual storytelling is the task of generating stories based on a sequence of images. Inspired by the recent works in neural generation focusing on controlling the form of text, this paper explores the idea of generating these stories in different personas. However, one of the main challenges of performing this task is the lack of a dataset of visual stories in different personas. Having said that, there are independent datasets for both visual storytelling and annotated sentences for various persona. In this paper we describe an approach to overcome this by getting labelled persona data from a different task and leveraging those annotations to perform persona based story generation. We inspect various ways of incorporating personality in both the encoder and the decoder representations to steer the generation in the target direction. To this end, we propose five models which are incremental extensions to the baseline model to perform the task at hand. In our experiments we use five different personas to guide the generation process. We find that the models based on our hypotheses perform better at capturing words while generating stories in the target persona.

Sunayana Sitaram, Khyathi Raghavi Chandu, Sai Krishna Rallabandi et al.

Code-switching, the alternation of languages within a conversation or utterance, is a common communicative phenomenon that occurs in multilingual communities across the world. This survey reviews computational approaches for code-switched Speech and Natural Language Processing. We motivate why processing code-switched text and speech is essential for building intelligent agents and systems that interact with users in multilingual communities. As code-switching data and resources are scarce, we list what is available in various code-switched language pairs with the language processing tasks they can be used for. We review code-switching research in various Speech and NLP applications, including language processing tools and end-to-end systems. We conclude with future directions and open problems in the field.

Khyathi Raghavi Chandu, Mary Arpita Pyreddy, Matthieu Felix et al.

Problems at the intersection of language and vision, like visual question answering, have recently been gaining a lot of attention in the field of multi-modal machine learning as computer vision research moves beyond traditional recognition tasks. There has been recent success in visual question answering using deep neural network models which use the linguistic structure of the questions to dynamically instantiate network layouts. In the process of converting the question to a network layout, the question is simplified, which results in loss of information in the model. In this paper, we enrich the image information with textual data using image captions and external knowledge bases to generate more coherent answers. We achieve 57.1% overall accuracy on the test-dev open-ended questions from the visual question answering (VQA 1.0) real image dataset.

Soumya Wadhwa, Khyathi Raghavi Chandu, Eric Nyberg

The task of Question Answering has gained prominence in the past few decades for testing the ability of machines to understand natural language. Large datasets for Machine Reading have led to the development of neural models that cater to deeper language understanding compared to information retrieval tasks. Different components in these neural architectures are intended to tackle different challenges. As a first step towards achieving generalization across multiple domains, we attempt to understand and compare the peculiarities of existing end-to-end neural models on the Stanford Question Answering Dataset (SQuAD) by performing quantitative as well as qualitative analysis of the results attained by each of them. We observed that prediction errors reflect certain model-specific biases, which we further discuss in this paper.