Vishakh Padmakumar

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Vishakh Padmakumar, Lujain Ibrahim, Zora Zhiruo Wang et al.

AI tools are increasingly integrated into real-world workflows. However, existing measures of reliance on these tools focus on AI output adoption or on self-reported indicators, rather than how task effort is distributed between users and tools. Here, we introduce offloading score, a measure of reliance that quantifies the fraction of cognitive effort offloaded to an AI tool. Offloading Score is simulation-based -- we construct a counterfactual workflow by estimating how the user would have completed the task without the tool, and then computing the fraction of steps saved by using the tool. We validate offloading score through intrinsic evaluations of metric validity, and a controlled user study ($n=40$) with developers performing programming tasks using AI tools. We vary time pressure to test whether reliance measures capture the known increase in reliance under time pressure. We show that offloading score detects significantly higher reliance in time-constrained settings ($+43\%$, $p=0.018$), while usage-based and self-reported baseline measures of reliance do not distinguish the conditions. We complement this with descriptive insights showing that higher reliance manifests as greater delegation of subtasks to the tool and more direct reuse of AI outputs. Finally, we demonstrate an approach of using offloading score in combination with target outcomes of a task (e.g., code understanding) to identify when reliance may be (in)appropriate. Our framework offers two contributions: an instrument users can apply to measure and reflect on their own reliance, and a quantitative signal that agent designers can utilize to mitigate overreliance.

Vishakh Padmakumar, Richard Yuanzhe Pang, He He et al.

We study the problem of extrapolative controlled generation, i.e., generating sequences with attribute values beyond the range seen in training. This task is of significant importance in automated design, especially drug discovery, where the goal is to design novel proteins that are \textit{better} (e.g., more stable) than existing sequences. Thus, by definition, the target sequences and their attribute values are out of the training distribution, posing challenges to existing methods that aim to directly generate the target sequence. Instead, in this work, we propose Iterative Controlled Extrapolation (ICE) which iteratively makes local edits to a sequence to enable extrapolation. We train the model on synthetically generated sequence pairs that demonstrate small improvement in the attribute value. Results on one natural language task (sentiment analysis) and two protein engineering tasks (ACE2 stability and AAV fitness) show that ICE considerably outperforms state-of-the-art approaches despite its simplicity. Our code and models are available at: https://github.com/vishakhpk/iter-extrapolation.

Vishakh Padmakumar, Leonard Lausen, Miguel Ballesteros et al.

Recent work has found that multi-task training with a large number of diverse tasks can uniformly improve downstream performance on unseen target tasks. In contrast, literature on task transferability has established that the choice of intermediate tasks can heavily affect downstream task performance. In this work, we aim to disentangle the effect of scale and relatedness of tasks in multi-task representation learning. We find that, on average, increasing the scale of multi-task learning, in terms of the number of tasks, indeed results in better learned representations than smaller multi-task setups. However, if the target tasks are known ahead of time, then training on a smaller set of related tasks is competitive to the large-scale multi-task training at a reduced computational cost.

Vishakh Padmakumar, He He

Large language models (LLMs) have led to a surge in collaborative writing with model assistance. As different users incorporate suggestions from the same model, there is a risk of decreased diversity in the produced content, potentially limiting diverse perspectives in public discourse. In this work, we measure the impact of co-writing on diversity via a controlled experiment, where users write argumentative essays in three setups -- using a base LLM (GPT3), a feedback-tuned LLM (InstructGPT), and writing without model help. We develop a set of diversity metrics and find that writing with InstructGPT (but not the GPT3) results in a statistically significant reduction in diversity. Specifically, it increases the similarity between the writings of different authors and reduces the overall lexical and content diversity. We additionally find that this effect is mainly attributable to InstructGPT contributing less diverse text to co-written essays. In contrast, the user-contributed text remains unaffected by model collaboration. This suggests that the recent improvement in generation quality from adapting models to human feedback might come at the cost of more homogeneous and less diverse content.

Tuhin Chakrabarty, Vishakh Padmakumar, Faeze Brahman et al.

The development of large language models (LLMs) capable of following instructions and engaging in conversational interactions sparked increased interest in their utilization across various support tools. We investigate the utility of modern LLMs in assisting professional writers via an empirical user study (n=30). The design of our collaborative writing interface is grounded in the cognitive process model of writing that views writing as a goal-oriented thinking process encompassing non-linear cognitive activities: planning, translating, and reviewing. Participants are asked to submit a post-completion survey to provide feedback on the potential and pitfalls of LLMs as writing collaborators. Upon analyzing the writer-LLM interactions, we find that while writers seek LLM's help across all three types of cognitive activities, they find LLMs more helpful in translation and reviewing. Our findings from analyzing both the interactions and the survey responses highlight future research directions in creative writing assistance using LLMs.

Richard Yuanzhe Pang, Vishakh Padmakumar, Thibault Sellam et al.

To align conditional text generation model outputs with desired behaviors, there has been an increasing focus on training the model using reinforcement learning (RL) with reward functions learned from human annotations. Under this framework, we identify three common cases where high rewards are incorrectly assigned to undesirable patterns: noise-induced spurious correlation, naturally occurring spurious correlation, and covariate shift. We show that even though learned metrics achieve high performance on the distribution of the data used to train the reward function, the undesirable patterns may be amplified during RL training of the text generation model. While there has been discussion about reward gaming in the RL or safety community, in this discussion piece, we would like to highlight reward gaming in the natural language generation (NLG) community using concrete conditional text generation examples and discuss potential fixes and areas for future work.

Tuhin Chakrabarty, Vishakh Padmakumar, He He

Recent work in training large language models (LLMs) to follow natural language instructions has opened up exciting opportunities for natural language interface design. Building on the prior success of LLMs in the realm of computer-assisted creativity, we aim to study if LLMs can improve the quality of user-generated content through collaboration. We present CoPoet, a collaborative poetry writing system. In contrast to auto-completing a user's text, CoPoet is controlled by user instructions that specify the attributes of the desired text, such as Write a sentence about `love' or Write a sentence ending in `fly'. The core component of our system is a language model fine-tuned on a diverse collection of instructions for poetry writing. Our model is not only competitive with publicly available LLMs trained on instructions (InstructGPT), but is also capable of satisfying unseen compositional instructions. A study with 15 qualified crowdworkers shows that users successfully write poems with CoPoet on diverse topics ranging from Monarchy to Climate change. Further, the collaboratively written poems are preferred by third-party evaluators over those written without the system.

David Anugraha, Vishakh Padmakumar, Diyi Yang

Qualitative insights from user experiences are critical for informing product and policy decisions, but collecting such data at scale is constrained by the time and availability of experts to conduct semi-structured interviews. Recent work has explored using large language models (LLMs) to automate interviewing, yet existing systems lack a principled mechanism for balancing systematic coverage of predefined topics with adaptive exploration, or the ability to pursue follow-ups, deep dives, and emergent themes that arise organically during conversation. In this work, we formulate adaptive semi-structured interviewing as an optimization problem over the interviewer's behavior. We define interview utility as a trade-off between coverage of a predefined interview topic guide, discovery of relevant emergent themes, and interview cost measured by length. Based on this formulation, we introduce SparkMe, a multi-agent LLM interviewer that performs deliberative planning via simulated conversation rollouts to select questions with high expected utility. We evaluate SparkMe through controlled experiments with LLM-based interviewees, showing that it achieves higher interview utility, improving topic guide coverage (+4.7% over the best baseline) and eliciting richer emergent insights while using fewer conversational turns than prior LLM interviewing approaches. We further validate SparkMe in a user study with 70 participants across 7 professions on the impact of AI on their workflows. Domain experts rate SparkMe as producing high-quality adaptive interviews that surface helpful profession-specific insights not captured by prior approaches. The code, datasets, and evaluation protocols for SparkMe are available as open-source at https://github.com/SALT-NLP/SparkMe.

Alicia Parrish, Harsh Trivedi, Nikita Nangia et al.

The use of language-model-based question-answering systems to aid humans in completing difficult tasks is limited, in part, by the unreliability of the text these systems generate. Using hard multiple-choice reading comprehension questions as a testbed, we assess whether presenting humans with arguments for two competing answer options, where one is correct and the other is incorrect, allows human judges to perform more accurately, even when one of the arguments is unreliable and deceptive. If this is helpful, we may be able to increase our justified trust in language-model-based systems by asking them to produce these arguments where needed. Previous research has shown that just a single turn of arguments in this format is not helpful to humans. However, as debate settings are characterized by a back-and-forth dialogue, we follow up on previous results to test whether adding a second round of counter-arguments is helpful to humans. We find that, regardless of whether they have access to arguments or not, humans perform similarly on our task. These findings suggest that, in the case of answering reading comprehension questions, debate is not a helpful format.

Julian Michael, Salsabila Mahdi, David Rein et al.

As AI systems are used to answer more difficult questions and potentially help create new knowledge, judging the truthfulness of their outputs becomes more difficult and more important. How can we supervise unreliable experts, which have access to the truth but may not accurately report it, to give answers that are systematically true and don't just superficially seem true, when the supervisor can't tell the difference between the two on their own? In this work, we show that debate between two unreliable experts can help a non-expert judge more reliably identify the truth. We collect a dataset of human-written debates on hard reading comprehension questions where the judge has not read the source passage, only ever seeing expert arguments and short quotes selectively revealed by 'expert' debaters who have access to the passage. In our debates, one expert argues for the correct answer, and the other for an incorrect answer. Comparing debate to a baseline we call consultancy, where a single expert argues for only one answer which is correct half of the time, we find that debate performs significantly better, with 84% judge accuracy compared to consultancy's 74%. Debates are also more efficient, being 68% of the length of consultancies. By comparing human to AI debaters, we find evidence that with more skilled (in this case, human) debaters, the performance of debate goes up but the performance of consultancy goes down. Our error analysis also supports this trend, with 46% of errors in human debate attributable to mistakes by the honest debater (which should go away with increased skill); whereas 52% of errors in human consultancy are due to debaters obfuscating the relevant evidence from the judge (which should become worse with increased skill). Overall, these results show that debate is a promising approach for supervising increasingly capable but potentially unreliable AI systems.

Joachim Baumann, Vishakh Padmakumar, Xiang Li et al.

AI coding agents are being adopted at scale, yet we lack empirical evidence on how people actually use them and how much of their output is useful in practice. We present SWE-chat, the first large-scale dataset of real coding agent sessions collected from open-source developers in the wild. The dataset currently contains 6,000 sessions, comprising more than 63,000 user prompts and 355,000 agent tool calls. SWE-chat is a living dataset; our collection pipeline automatically and continually discovers and processes sessions from public repositories. Leveraging SWE-chat, we provide an initial empirical characterization of real-world coding agent usage and failure modes. We find that coding patterns are bimodal: in 41% of sessions, agents author virtually all committed code ("vibe coding"), while in 23%, humans write all code themselves. Despite rapidly improving capabilities, coding agents remain inefficient in natural settings. Just 44% of all agent-produced code survives into user commits, and agent-written code introduces more security vulnerabilities than code authored by humans. Furthermore, users push back against agent outputs -- through corrections, failure reports, and interruptions -- in 44% of all turns. By capturing complete interaction traces with human vs. agent code authorship attribution, SWE-chat provides an empirical foundation for moving beyond curated benchmarks towards an evidence-based understanding of how AI agents perform in real developer workflows.

John Joon Young Chung, Vishakh Padmakumar, Melissa Roemmele et al.

People have different creative writing preferences, and large language models (LLMs) for these tasks can benefit from adapting to each user's preferences. However, these models are often trained over a dataset that considers varying personal tastes as a monolith. To facilitate developing personalized creative writing LLMs, we introduce LiteraryTaste, a dataset of reading preferences from 60 people, where each person: 1) self-reported their reading habits and tastes (stated preference), and 2) annotated their preferences over 100 pairs of short creative writing texts (revealed preference). With our dataset, we found that: 1) people diverge on creative writing preferences, 2) finetuning a transformer encoder could achieve 75.8% and 67.7% accuracy when modeling personal and collective revealed preferences, and 3) stated preferences had limited utility in modeling revealed preferences. With an LLM-driven interpretability pipeline, we analyzed how people's preferences vary. We hope our work serves as a cornerstone for personalizing creative writing technologies.

John Joon Young Chung, Vishakh Padmakumar, Melissa Roemmele et al.

As creative writing tasks do not have singular correct answers, large language models (LLMs) trained to perform these tasks should be able to generate diverse valid outputs. However, LLM post-training often focuses on improving generation quality but neglects to facilitate output diversity. Hence, in creative writing generation, we investigate post-training approaches to promote both output diversity and quality. Our core idea is to include deviation -- the degree of difference between a training sample and all other samples with the same prompt -- in the training objective to facilitate learning from rare high-quality instances. By adopting our approach to direct preference optimization (DPO) and odds ratio preference optimization (ORPO), we demonstrate that we can promote the output diversity of trained models while minimally decreasing quality. Our best model with 8B parameters could achieve on-par diversity as a human-created dataset while having output quality similar to the best instruction-tuned models we examined, GPT-4o and DeepSeek-R1. We further validate our approaches with a human evaluation, an ablation, and a comparison to an existing diversification approach, DivPO.

Alexander Shypula, Shuo Li, Botong Zhang et al.

Recent work suggests that preference-tuning techniques--including Reinforcement Learning from Human Preferences (RLHF) methods like PPO and GRPO, as well as alternatives like DPO--reduce diversity, creating a dilemma given that such models are widely deployed in applications requiring diverse outputs. To address this, we introduce a framework for measuring effective semantic diversity--diversity among outputs that meet quality thresholds--which better reflects the practical utility of large language models (LLMs). Using open-ended tasks that require no human intervention, we find counterintuitive results: although preference-tuned models--especially those trained via RL--exhibit reduced lexical and syntactic diversity, they produce greater effective semantic diversity than SFT or base models, not from increasing diversity among high-quality outputs, but from generating more high-quality outputs overall. We discover that preference tuning reduces syntactic diversity while preserving semantic diversity--revealing a distinction between diversity in form and diversity in content that traditional metrics often overlook. Our analysis further shows that smaller models are consistently more parameter-efficient at generating unique content within a fixed sampling budget, offering insights into the relationship between model scaling and diversity. These findings have important implications for applications that require diverse yet high-quality outputs, from creative assistance to synthetic data generation.

Abulhair Saparov, Srushti Pawar, Shreyas Pimpalgaonkar et al.

Search is an ability foundational in many important tasks, and recent studies have shown that large language models (LLMs) struggle to perform search robustly. It is unknown whether this inability is due to a lack of data, insufficient model parameters, or fundamental limitations of the transformer architecture. In this work, we use the foundational graph connectivity problem as a testbed to generate effectively limitless high-coverage data to train small transformers and test whether they can learn to perform search. We find that, when given the right training distribution, the transformer is able to learn to search. We analyze the algorithm that the transformer has learned through a novel mechanistic interpretability technique that enables us to extract the computation graph from the trained model. We find that transformers perform search at every vertex in parallel: For each vertex in the input graph, transformers compute the set of vertices reachable from that vertex. Each layer then progressively expands these sets, allowing the model to search over a number of vertices exponential in $n_{\text{layers}}$. However, we find that as the input graph size increases, the transformer has greater difficulty in learning the task. This difficulty is not resolved even as the number of parameters is increased, suggesting that increasing model scale will not lead to robust search abilities. We also find that performing search in-context (i.e., chain-of-thought) does not resolve this inability to learn to search on larger graphs.

Nishant Balepur, Vishakh Padmakumar, Fumeng Yang et al.

LLMs are aligned to follow input instructions by learning which of two responses users prefer for a prompt. However, such preference data do not convey why users prefer responses that are chosen or rejected, so LLMs trained on these datasets cannot tailor responses to varied user needs. To surface these parameters of personalization, we apply abductive reasoning to preference data, inferring needs and interests of users, i.e., personas, that may prefer either response. We test this idea in two steps: Persona Inference (PI), abductively inferring personas of users who prefer chosen or rejected outputs, and Persona Tailoring (PT), training models to tailor outputs to personas from PI. We show: 1) LLMs infer personas accurately explaining why different users may prefer both chosen or rejected outputs; 2) Training on preference data augmented with PI personas via PT boosts personalization and generalizes to supporting user-written personas; and 3) Rejected response personas form harder personalization evaluations, showing PT better aids users with uncommon preferences versus typical alignment methods. We argue for an abductive view of preferences for personalization, asking not only which response is better but when, why, and for whom.

Vishakh Padmakumar, Chuanyang Jin, Hannah Rose Kirk et al.

Large language models (LLMs) are increasingly deployed via public-facing interfaces to interact with millions of users, each with diverse preferences. Despite this, preference tuning of LLMs predominantly relies on reward models trained using binary judgments where annotators select the preferred choice out of pairs of model outputs. In this work, we argue that this reliance on binary choices does not capture the broader, aggregate preferences of the target user in real-world tasks. We propose a taxonomy that identifies two dimensions of subjectivity where different users disagree on the preferred output-namely, the Plurality of Responses to Prompts, where prompts allow for multiple correct answers, and the Indistinguishability of Responses, where candidate outputs are paraphrases of each other. We show that reward models correlate weakly with user preferences in these cases. As a first step to address this issue, we introduce a simple yet effective method that augments existing binary preference datasets with synthetic preference judgments to estimate potential user disagreement. Incorporating these via a margin term as a form of regularization during model training yields predictions that better align with the aggregate user preferences.

Yuhan Liu, Fangyuan Xu, Vishakh Padmakumar et al.

When posed with prompts that permit a large number of valid answers, comprehensively generating them is the first step towards satisfying a wide range of users. In this paper, we study methods to elicit a comprehensive set of valid responses. To evaluate this, we introduce \textbf{diversity coverage}, a metric that measures the total quality scores assigned to each \textbf{unique} answer in the predicted answer set relative to the best possible answer set with the same number of answers. Using this metric, we evaluate 18 LLMs, finding no single model dominates at generating diverse responses to a wide range of open-ended prompts. Yet, per each prompt, there exists a model that outperforms all other models significantly at generating a diverse answer set. Motivated by this finding, we introduce a router that predicts the best model for each query. On NB-Wildchat, our trained router outperforms the single best model baseline (26.3% vs $23.8%). We further show generalization to an out-of-domain dataset (NB-Curated) as well as different answer-generation prompting strategies. Our work lays foundation for studying generating comprehensive answers when we have access to a suite of models.

Vishakh Padmakumar, Chen Yueh-Han, Jane Pan et al. · berkeley, cmu

As large language models (LLMs) are increasingly used for ideation and scientific discovery, it is important to evaluate their ability to generate novel output. Prior work evaluates novelty as originality with respect to model training data, but original outputs may be of low quality. In contrast, non-expert judges more reliably score quality but may favor memorized outputs, limiting the reliability of human preference as a metric. We introduce a new novelty metric for LLM generations that balances originality and quality -- the harmonic mean of the fraction of \ngrams unseen during training and a task-specific quality score. Using this framework, we identify trends that affect the novelty of generations from three families of open-data models (OLMo, OLMo-2, and Pythia) on three creative tasks: story completion, poetry writing, and creative tool use. We find that model-generated text from some base LLMs is less novel than human-written text from the internet. However, increasing model scale and post-training reliably improves novelty due to improvements in output quality. We also find that improving the base model at the same scale (\eg OLMo 7B to OLMo-2 7B) leads to higher novelty due to higher originality. Finally, we observe that inference-time methods, such as prompting and providing novel in-context examples, have a much smaller effect on novelty, often increasing originality at the expense of quality. This highlights the need for further research into more effective elicitation strategies as we use models for creative applications.

Vishakh Padmakumar, Joseph Chee Chang, Kyle Lo et al.

The increasing volume of academic literature makes it essential for researchers to organize, compare, and contrast collections of documents. Large language models (LLMs) can support this process by generating schemas defining shared aspects along which to compare papers. However, progress on schema generation has been slow due to: (i) ambiguity in reference-based evaluations, and (ii) lack of editing/refinement methods. Our work is the first to address both issues. First, we present an approach for augmenting unannotated table corpora with \emph{synthesized intents}, and apply it to create a dataset for studying schema generation conditioned on a given information need, thus reducing ambiguity. With this dataset, we show how incorporating table intents significantly improves baseline performance in reconstructing reference schemas. We start by comprehensively benchmarking several single-shot schema generation methods, including prompted LLM workflows and fine-tuned models, showing that smaller, open-weight models can be fine-tuned to be competitive with state-of-the-art prompted LLMs. Next, we propose several LLM-based schema refinement techniques and show that these can further improve schemas generated by these methods.

Vishakh Padmakumar, Zichao Wang, David Arbour et al.

While large language models (LLMs) are increasingly capable of handling longer contexts, recent work has demonstrated that they exhibit the "lost in the middle" phenomenon (Liu et al., 2024) of unevenly attending to different parts of the provided context. This hinders their ability to cover diverse source material in multi-document summarization, as noted in the DiverseSumm benchmark (Huang et al., 2024). In this work, we contend that principled content selection is a simple way to increase source coverage on this task. As opposed to prompting an LLM to perform the summarization in a single step, we explicitly divide the task into three steps -- (1) reducing document collections to atomic key points, (2) using determinantal point processes (DPP) to perform select key points that prioritize diverse content, and (3) rewriting to the final summary. By combining prompting steps, for extraction and rewriting, with principled techniques, for content selection, we consistently improve source coverage on the DiverseSumm benchmark across various LLMs. Finally, we also show that by incorporating relevance to a provided user intent into the DPP kernel, we can generate personalized summaries that cover relevant source information while retaining coverage.

Abulhair Saparov, Richard Yuanzhe Pang, Vishakh Padmakumar et al.

Given the intractably large size of the space of proofs, any model that is capable of general deductive reasoning must generalize to proofs of greater complexity. Recent studies have shown that large language models (LLMs) possess some abstract deductive reasoning ability given chain-of-thought prompts. However, they have primarily been tested on proofs using modus ponens or of a specific size, and from the same distribution as the in-context examples. To measure the general deductive reasoning ability of LLMs, we test on a broad set of deduction rules and measure their ability to generalize to more complex proofs from simpler demonstrations from multiple angles: depth-, width-, and compositional generalization. To facilitate systematic exploration, we construct a new synthetic and programmable reasoning dataset that enables control over deduction rules and proof complexity. Our experiments on four LLMs of various sizes and training objectives show that they are able to generalize to compositional proofs. However, they have difficulty generalizing to longer proofs, and they require explicit demonstrations to produce hypothetical subproofs, specifically in proof by cases and proof by contradiction.

Angelica Chen, Jason Phang, Alicia Parrish et al.

Large language models (LLMs) have achieved widespread success on a variety of in-context few-shot tasks, but this success is typically evaluated via correctness rather than consistency. We argue that self-consistency is an important criteria for valid multi-step reasoning in tasks where the solution is composed of the answers to multiple sub-steps. We propose two types of self-consistency that are particularly important for multi-step reasoning -- hypothetical consistency (a model's ability to predict what its output would be in a hypothetical other context) and compositional consistency (consistency of a model's final outputs when intermediate sub-steps are replaced with the model's outputs for those steps). We demonstrate that multiple variants of the GPT-3/-4 models exhibit poor consistency rates across both types of consistency on a variety of tasks.

Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi et al.

To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than relying on summaries or excerpts. In addition, only half of the questions are answerable by annotators working under tight time constraints, indicating that skimming and simple search are not enough to consistently perform well. Our baseline models perform poorly on this task (55.4%) and significantly lag behind human performance (93.5%).

Vishakh Padmakumar, He He

Machine-in-the-loop writing aims to enable humans to collaborate with models to complete their writing tasks more effectively. Prior work has found that providing humans a machine-written draft or sentence-level continuations has limited success since the generated text tends to deviate from humans' intention. To allow the user to retain control over the content, we train a rewriting model that, when prompted, modifies specified spans of text within the user's original draft to introduce descriptive and figurative elements locally in the text. We evaluate the model on its ability to collaborate with humans on the task of creative image captioning. On a user study through Amazon Mechanical Turk, our model is rated to be more helpful than a baseline infilling language model. In addition, third-party evaluation shows that users write more descriptive and figurative captions when collaborating with our model compared to completing the task alone.

Alicia Parrish, Angelica Chen, Nikita Nangia et al.

It is well documented that NLP models learn social biases, but little work has been done on how these biases manifest in model outputs for applied tasks like question answering (QA). We introduce the Bias Benchmark for QA (BBQ), a dataset of question sets constructed by the authors that highlight attested social biases against people belonging to protected classes along nine social dimensions relevant for U.S. English-speaking contexts. Our task evaluates model responses at two levels: (i) given an under-informative context, we test how strongly responses reflect social biases, and (ii) given an adequately informative context, we test whether the model's biases override a correct answer choice. We find that models often rely on stereotypes when the context is under-informative, meaning the model's outputs consistently reproduce harmful biases in this setting. Though models are more accurate when the context provides an informative answer, they still rely on stereotypes and average up to 3.4 percentage points higher accuracy when the correct answer aligns with a social bias than when it conflicts, with this difference widening to over 5 points on examples targeting gender for most models tested.

Vishakh Padmakumar, He He

Unsupervised approaches to extractive summarization usually rely on a notion of sentence importance defined by the semantic similarity between a sentence and the document. We propose new metrics of relevance and redundancy using pointwise mutual information (PMI) between sentences, which can be easily computed by a pre-trained language model. Intuitively, a relevant sentence allows readers to infer the document content (high PMI with the document), and a redundant sentence can be inferred from the summary (high PMI with the summary). We then develop a greedy sentence selection algorithm to maximize relevance and minimize redundancy of extracted sentences. We show that our method outperforms similarity-based methods on datasets in a range of domains including news, medical journal articles, and personal anecdotes.