Rachel Rudinger

Nishant Balepur, Shramay Palta, Rachel Rudinger · microsoft-research

Chain-of-thought (COT) prompting can help large language models (LLMs) reason toward correct answers, but its efficacy in reasoning toward incorrect answers is unexplored. This process of elimination (PoE), when used with COT, can enhance self-consistency, interpretability, and tasks such as medical diagnoses of exclusion. Thus, we propose PoE with COT, where LLMs must reason toward incorrect options on multiple-choice questions. We evaluate the ability of GPT-3.5, LLaMA-2, and Falcon to perform PoE with COT on a total of four commonsense and scientific reasoning datasets. We find that the strategy of PoE always underperforms the strategy of choosing the correct answer. The agreement of these strategies is also lower than the self-consistency of each strategy. To study these issues further, we conduct error analyses and give suggestions for future work.

Christabel Acquaye, Yi Ting Huang, Marine Carpuat et al.

Standardized math assessments require expensive human pilot studies to establish the difficulty of test items. We investigate the predictive value of open-source large language models (LLMs) for evaluating the difficulty of multiple-choice math questions for real-world students. We show that, while LLMs are poor direct judges of problem difficulty, simulation-based approaches with LLMs yield promising results under the right conditions. Under the proposed approach, we simulate a "classroom" of 4th, 8th, or 12th grade students by prompting the LLM to role-play students of varying proficiency levels. We use the outcomes of these simulations to fit Item Response Theory (IRT) models, comparing learned difficulty parameters for items to their real-world difficulties, as determined by item-level statistics furnished by the National Assessment of Educational Progress (NAEP). We observe correlations as high as 0.75, 0.76, and 0.82 for grades 4, 8, and 12, respectively. In our simulations, we experiment with different "classroom sizes," showing tradeoffs between computation size and accuracy. We find that role-plays with named students improves predictions (compared to student ids), and stratifying names across gender and race further improves predictions. Our results show that LLMs with relatively weaker mathematical abilities (Gemma) actually yield better real-world difficulty predictions than mathematically stronger models (Llama and Qwen), further underscoring the suitability of open-source models for the task.

Yang Trista Cao, Anna Sotnikova, Hal Daumé et al.

NLP models trained on text have been shown to reproduce human stereotypes, which can magnify harms to marginalized groups when systems are deployed at scale. We adapt the Agency-Belief-Communion (ABC) stereotype model of Koch et al. (2016) from social psychology as a framework for the systematic study and discovery of stereotypic group-trait associations in language models (LMs). We introduce the sensitivity test (SeT) for measuring stereotypical associations from language models. To evaluate SeT and other measures using the ABC model, we collect group-trait judgments from U.S.-based subjects to compare with English LM stereotypes. Finally, we extend this framework to measure LM stereotyping of intersectional identities.

Haozhe An, Zongxia Li, Jieyu Zhao et al.

A common limitation of diagnostic tests for detecting social biases in NLP models is that they may only detect stereotypic associations that are pre-specified by the designer of the test. Since enumerating all possible problematic associations is infeasible, it is likely these tests fail to detect biases that are present in a model but not pre-specified by the designer. To address this limitation, we propose SODAPOP (SOcial bias Discovery from Answers about PeOPle) in social commonsense question-answering. Our pipeline generates modified instances from the Social IQa dataset (Sap et al., 2019) by (1) substituting names associated with different demographic groups, and (2) generating many distractor answers from a masked language model. By using a social commonsense model to score the generated distractors, we are able to uncover the model's stereotypic associations between demographic groups and an open set of words. We also test SODAPOP on debiased models and show the limitations of multiple state-of-the-art debiasing algorithms.

Nishant Balepur, Rachel Rudinger

Recent work shows that large language models (LLMs) can answer multiple-choice questions using only the choices, but does this mean that MCQA leaderboard rankings of LLMs are largely influenced by abilities in choices-only settings? To answer this, we use a contrast set that probes if LLMs over-rely on choices-only shortcuts in MCQA. While previous works build contrast sets via expensive human annotations or model-generated data which can be biased, we employ graph mining to extract contrast sets from existing MCQA datasets. We use our method on UnifiedQA, a group of six commonsense reasoning datasets with high choices-only accuracy, to build an 820-question contrast set. After validating our contrast set, we test 12 LLMs, finding that these models do not exhibit reliance on choice-only shortcuts when given both the question and choices. Thus, despite the susceptibility~of MCQA to high choices-only accuracy, we argue that LLMs are not obtaining high ranks on MCQA leaderboards just due to their ability to exploit choices-only shortcuts.

Neha Srikanth, Rupak Sarkar, Heran Mane et al.

Questions posed by information-seeking users often contain implicit false or potentially harmful assumptions. In a high-risk domain such as maternal and infant health, a question-answering system must recognize these pragmatic constraints and go beyond simply answering user questions, examining them in context to respond helpfully. To achieve this, we study assumptions and implications, or pragmatic inferences, made when mothers ask questions about pregnancy and infant care by collecting a dataset of 2,727 inferences from 500 questions across three diverse sources. We study how health experts naturally address these inferences when writing answers, and illustrate that informing existing QA pipelines with pragmatic inferences produces responses that are more complete, mitigating the propagation of harmful beliefs.

Abhilasha Sancheti, Aparna Garimella, Balaji Vasan Srinivasan et al.

Legal documents are typically long and written in legalese, which makes it particularly difficult for laypeople to understand their rights and duties. While natural language understanding technologies can be valuable in supporting such understanding in the legal domain, the limited availability of datasets annotated for deontic modalities in the legal domain, due to the cost of hiring experts and privacy issues, is a bottleneck. To this end, we introduce, LEXDEMOD, a corpus of English contracts annotated with deontic modality expressed with respect to a contracting party or agent along with the modal triggers. We benchmark this dataset on two tasks: (i) agent-specific multi-label deontic modality classification, and (ii) agent-specific deontic modality and trigger span detection using Transformer-based (Vaswani et al., 2017) language models. Transfer learning experiments show that the linguistic diversity of modal expressions in LEXDEMOD generalizes reasonably from lease to employment and rental agreements. A small case study indicates that a model trained on LEXDEMOD can detect red flags with high recall. We believe our work offers a new research direction for deontic modality detection in the legal domain.

Abhilasha Sancheti, Aparna Garimella, Balaji Vasan Srinivasan et al.

Reviewing and comprehending key obligations, entitlements, and prohibitions in legal contracts can be a tedious task due to their length and domain-specificity. Furthermore, the key rights and duties requiring review vary for each contracting party. In this work, we propose a new task of party-specific extractive summarization for legal contracts to facilitate faster reviewing and improved comprehension of rights and duties. To facilitate this, we curate a dataset comprising of party-specific pairwise importance comparisons annotated by legal experts, covering ~293K sentence pairs that include obligations, entitlements, and prohibitions extracted from lease agreements. Using this dataset, we train a pairwise importance ranker and propose a pipeline-based extractive summarization system that generates a party-specific contract summary. We establish the need for incorporating domain-specific notion of importance during summarization by comparing our system against various baselines using both automatic and human evaluation methods

Neha Srikanth, Rachel Rudinger

When strong partial-input baselines reveal artifacts in crowdsourced NLI datasets, the performance of full-input models trained on such datasets is often dismissed as reliance on spurious correlations. We investigate whether state-of-the-art NLI models are capable of overriding default inferences made by a partial-input baseline. We introduce an evaluation set of 600 examples consisting of perturbed premises to examine a RoBERTa model's sensitivity to edited contexts. Our results indicate that NLI models are still capable of learning to condition on context--a necessary component of inferential reasoning--despite being trained on artifact-ridden datasets.

Abhilasha Sancheti, Balaji Vasan Srinivasan, Rachel Rudinger

We introduce a new task of entailment relation aware paraphrase generation which aims at generating a paraphrase conforming to a given entailment relation (e.g. equivalent, forward entailing, or reverse entailing) with respect to a given input. We propose a reinforcement learning-based weakly-supervised paraphrasing system, ERAP, that can be trained using existing paraphrase and natural language inference (NLI) corpora without an explicit task-specific corpus. A combination of automated and human evaluations show that ERAP generates paraphrases conforming to the specified entailment relation and are of good quality as compared to the baselines and uncontrolled paraphrasing systems. Using ERAP for augmenting training data for downstream textual entailment task improves performance over an uncontrolled paraphrasing system, and introduces fewer training artifacts, indicating the benefit of explicit control during paraphrasing.

Nishant Balepur, Bhavya Rajasekaran, Jane Oh et al.

Multiple-choice question answering (MCQA) is standard in NLP, but benchmarks lack rigorous quality control. We present BenchMarker, an education-inspired toolkit using LLM judges to flag three common MCQ flaws: 1) contamination - items appearing exactly online; 2) shortcuts - cues in the choices that enable guessing; and 3) writing errors - structural/grammatical issues based on a 19-rule education rubric. We validate BenchMarker with human annotations, then run the tool to audit 12 benchmarks, revealing: 2) contaminated MCQs tend to inflate accuracy, while writing errors tend to lower it and change rankings beyond random; and 3) prior benchmark repairs address their targeted issues (i.e., lowering accuracy with LLM-written distractors), but inadvertently add new flaws (i.e. implausible distractors, many correct answers). Overall, flaws in MCQs degrade NLP evaluation, but education research offers a path forward. We release BenchMarker to bridge the fields and improve MCQA benchmark design.

Nishant Balepur, Malachi Hamada, Varsha Kishore et al.

Scientific Deep Research (DR) agents answer user queries by synthesizing research papers into multi-section reports. User feedback can improve their utility, but existing protocols only score the final report, making it hard to study and learn which intermediate actions DR agents should take to improve reports. We collect DRACULA, the first dataset with user feedback on intermediate actions for DR. Over five weeks, nineteen expert CS researchers ask queries to a DR system that proposes actions (e.g., "Add a section on datasets"). Our users select actions they prefer, then judge whether an output report applied their selections successfully, yielding 8,103 action preferences and 5,230 execution judgments. After confirming a DR agent can execute DRACULA's actions, we study the predictability of user-preferred actions via simulation-how well LLMs predict the actions users select-a step toward learning to generate useful actions. We discover: (1) LLM judges initially struggle to predict action selections, but improve most when using a user's full selection history, rather than self-reported or extrapolated user context signals; (2) Users' selections for the same query differ based on unstated goals, bottlenecking simulation and motivating affordances that let users steer reports; and (3) Our simulation results inform an online intervention that generates new actions based on the user's past interactions, which users pick most often in follow-up studies. Overall, while work extensively studies execution, DRACULA reveals a key challenge is deciding which actions to execute in the first place. We open-source DRACULA's study design, user feedback, and simulation tasks to spur future work on action feedback for long-horizon agents.

Neha Srikanth, Jordan Boyd-Graber, Rachel Rudinger

We introduce DiscoTrace, a method to identify the rhetorical strategies that answerers use when responding to information-seeking questions. DiscoTrace represents answers as a sequence of question-related discourse acts paired with interpretations of the original question, annotated on top of rhetorical structure theory parses. Applying DiscoTrace to answers from nine different human communities reveals that communities have diverse preferences for answer construction. In contrast, LLMs do not exhibit rhetorical diversity in their answers, even when prompted to mimic specific human community answering guidelines. LLMs also systematically opt for breadth, addressing interpretations of questions that human answerers choose not to address. Our findings can guide the development of pragmatic LLM answerers that consider a range of strategies informed by context in QA.

Huy Nghiem, Phuong-Anh Nguyen-Le, John Prindle et al.

Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially sensitive decisions remains underexplored. We present a large-scale audit of LLMs' treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveal that LLMs consistently favor low-SES applicants -- even when controlling for academic performance -- and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs' reasoning behaviors in sensitive applications.

Nishant Balepur, Abhilasha Ravichander, Rachel Rudinger · cmu

Multiple-choice question answering (MCQA) is often used to evaluate large language models (LLMs). To see if MCQA assesses LLMs as intended, we probe if LLMs can perform MCQA with choices-only prompts, where models must select the correct answer only from the choices. In three MCQA datasets and four LLMs, this prompt bests a majority baseline in 11/12 cases, with up to 0.33 accuracy gain. To help explain this behavior, we conduct an in-depth, black-box analysis on memorization, choice dynamics, and question inference. Our key findings are threefold. First, we find no evidence that the choices-only accuracy stems from memorization alone. Second, priors over individual choices do not fully explain choices-only accuracy, hinting that LLMs use the group dynamics of choices. Third, LLMs have some ability to infer a relevant question from choices, and surprisingly can sometimes even match the original question. Inferring the original question is an impressive reasoning strategy, but it cannot fully explain the high choices-only accuracy of LLMs in MCQA. Thus, while LLMs are not fully incapable of reasoning in MCQA, we still advocate for the use of stronger baselines in MCQA benchmarks, the design of robust MCQA datasets for fair evaluations, and further efforts to explain LLM decision-making.

Nishant Balepur, Rachel Rudinger, Jordan Lee Boyd-Graber

Multiple choice question answering (MCQA) is popular for LLM evaluation due to its simplicity and human-like testing, but we argue for its reform. We first reveal flaws in MCQA's format, as it struggles to: 1) test generation/subjectivity; 2) match LLM use cases; and 3) fully test knowledge. We instead advocate for generative formats based on human testing, where LLMs construct and explain answers, better capturing user needs and knowledge while remaining easy to score. We then show even when MCQA is a useful format, its datasets suffer from: leakage; unanswerability; shortcuts; and saturation. In each issue, we give fixes from education, like rubrics to guide MCQ writing; scoring methods to bridle guessing; and Item Response Theory to build harder MCQs. Lastly, we discuss LLM errors in MCQA, robustness, biases, and unfaithful explanations, showing how our prior solutions better measure or address these issues. While we do not need to desert MCQA, we encourage more efforts in refining the task based on educational testing, advancing evaluations.

Yang Trista Cao, Anna Sotnikova, Jieyu Zhao et al.

Multilingual large language models have gained prominence for their proficiency in processing and generating text across languages. Like their monolingual counterparts, multilingual models are likely to pick up on stereotypes and other social biases present in their training data. In this paper, we study a phenomenon we term stereotype leakage, which refers to how training a model multilingually may lead to stereotypes expressed in one language showing up in the models' behaviour in another. We propose a measurement framework for stereotype leakage and investigate its effect across English, Russian, Chinese, and Hindi and with GPT-3.5, mT5, and mBERT. Our findings show a noticeable leakage of positive, negative, and non-polar associations across all languages. We find that of these models, GPT-3.5 exhibits the most stereotype leakage, and Hindi is the most susceptible to leakage effects. WARNING: This paper contains model outputs which could be offensive in nature.

Nishant Balepur, Feng Gu, Abhilasha Ravichander et al. · cmu

Question answering (QA), giving correct answers to questions, is a popular task, but we test reverse question answering (RQA): for an input answer, give a question with that answer. Past work tests QA and RQA separately, but we test them jointly, comparing their difficulty, aiding benchmark design, and checking reasoning consistency. We run 16 LLMs on QA and RQA with trivia questions/answers, revealing: 1) Versus QA, LLMs are much less accurate in RQA for numerical answers, but slightly more accurate in RQA for textual answers; 2) LLMs often answer their own invalid questions from RQA accurately in QA, so RQA errors are not from knowledge gaps alone; 3) RQA errors correlate with question difficulty and inversely correlate with answer frequencies in the Dolma corpus; and 4) LLMs struggle to provide valid multi-hop questions. By finding question and answer types that lead to RQA errors, we suggest improvements for LLM reasoning.

Dayeon Ki, Yu Hou, Rachel Rudinger et al.

Neologisms and emerging slang are central to daily conversation, yet challenging for non-native speakers (NNS) to interpret and use appropriately in cross-cultural communication with native speakers (NS). NNS increasingly make use of Artificial Intelligence (AI) tools to learn these words. We study the utility of such tools in mediating an informal communication scenario through a human-subjects study (N=234): NNS participants learn English neologisms with AI support, write messages using the learned word to an NS friend, and judge contextual appropriateness of the neologism in two provided writing samples. Using both NS evaluator-rated communicative competence of NNS-produced writing and NNS' contextual appropriateness judgments, we compare three AI-based support conditions: AI Definition, AI Rewrite into simpler English, AI Explanation of meaning and usage, and Non-AI Dictionary for comparison. We show that AI Explanation yields the largest gains over no support in NS-rated competence, while contextual appropriateness judgments show indifference across support. NNS participants' self-reported perceptions tend to overestimate NS ratings, revealing a mismatch between perceived and actual competence. We further observe a significant gap between NNS- and NS-produced writing, highlighting the limitations of current AI tools and informing design for future tools.

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.

Dayeon Ki, Rachel Rudinger, Tianyi Zhou et al.

Large Language Models (LLMs) need to adapt their predictions to diverse cultural contexts to benefit diverse communities across the world. While previous efforts have focused on single-LLM, single-turn approaches, we propose to exploit the complementary strengths of multiple LLMs to promote cultural adaptability. We introduce a Multi-Agent Debate framework, where two LLM-based agents debate over a cultural scenario and collaboratively reach a final decision. We propose two variants: one where either LLM agents exclusively debate and another where they dynamically choose between self-reflection and debate during their turns. We evaluate these approaches on 7 open-weight LLMs (and 21 LLM combinations) using the NormAd-ETI benchmark for social etiquette norms in 75 countries. Experiments show that debate improves both overall accuracy and cultural group parity over single-LLM baselines. Notably, multi-agent debate enables relatively small LLMs (7-9B) to achieve accuracies comparable to that of a much larger model (27B parameters).

Christabel Acquaye, Haozhe An, Rachel Rudinger

Recent work has highlighted the culturally-contingent nature of commonsense knowledge. We introduce AMAMMER$ε$, a test set of 525 multiple-choice questions designed to evaluate the commonsense knowledge of English LLMs, relative to the cultural contexts of Ghana and the United States. To create AMAMMER$ε$, we select a set of multiple-choice questions (MCQs) from existing commonsense datasets and rewrite them in a multi-stage process involving surveys of Ghanaian and U.S. participants. In three rounds of surveys, participants from both pools are solicited to (1) write correct and incorrect answer choices, (2) rate individual answer choices on a 5-point Likert scale, and (3) select the best answer choice from the newly-constructed MCQ items, in a final validation step. By engaging participants at multiple stages, our procedure ensures that participant perspectives are incorporated both in the creation and validation of test items, resulting in high levels of agreement within each pool. We evaluate several off-the-shelf English LLMs on AMAMMER$ε$. Uniformly, models prefer answers choices that align with the preferences of U.S. annotators over Ghanaian annotators. Additionally, when test items specify a cultural context (Ghana or the U.S.), models exhibit some ability to adapt, but performance is consistently better in U.S. contexts than Ghanaian. As large resources are devoted to the advancement of English LLMs, our findings underscore the need for culturally adaptable models and evaluations to meet the needs of diverse English-speaking populations around the world.

Yu Hou, Hal Daumé, Rachel Rudinger

Studies of human psychology have demonstrated that people are more motivated to extend empathy to in-group members than out-group members (Cikara et al., 2011). In this study, we investigate how this aspect of intergroup relations in humans is replicated by LLMs in an emotion intensity prediction task. In this task, the LLM is given a short description of an experience a person had that caused them to feel a particular emotion; the LLM is then prompted to predict the intensity of the emotion the person experienced on a numerical scale. By manipulating the group identities assigned to the LLM's persona (the "perceiver") and the person in the narrative (the "experiencer"), we measure how predicted emotion intensities differ between in-group and out-group settings. We observe that LLMs assign higher emotion intensity scores to in-group members than out-group members. This pattern holds across all three types of social groupings we tested: race/ethnicity, nationality, and religion. We perform an in-depth analysis on Llama-3.1-8B, the model which exhibited strongest intergroup bias among those tested.

Neha Srikanth, Rachel Rudinger

Decomposition of text into atomic propositions is a flexible framework allowing for the closer inspection of input and output text. We use atomic decomposition of hypotheses in two natural language reasoning tasks, traditional NLI and defeasible NLI, to form atomic sub-problems, or granular inferences that models must weigh when solving the overall problem. These atomic sub-problems serve as a tool to further understand the structure of both NLI and defeasible reasoning, probe a model's consistency and understanding of different inferences, and measure the diversity of examples in benchmark datasets. Our results indicate that LLMs still struggle with logical consistency on atomic NLI and defeasible NLI sub-problems. Lastly, we identify critical atomic sub-problems of defeasible NLI examples, or those that most contribute to the overall label, and propose a method to measure the inferential consistency of a model, a metric designed to capture the degree to which a model makes consistently correct or incorrect predictions about the same fact under different contexts.

Neha Srikanth, Marine Carpuat, Rachel Rudinger

Large language models have been shown to behave inconsistently in response to meaning-preserving paraphrastic inputs. At the same time, researchers evaluate the knowledge and reasoning abilities of these models with test evaluations that do not disaggregate the effect of paraphrastic variability on performance. We propose a metric for evaluating the paraphrastic consistency of natural language reasoning models based on the probability of a model achieving the same correctness on two paraphrases of the same problem. We mathematically connect this metric to the proportion of a model's variance in correctness attributable to paraphrasing. To estimate paraphrastic consistency, we collect ParaNLU, a dataset of 7,782 human-written and validated paraphrased reasoning problems constructed on top of existing benchmark datasets for defeasible and abductive natural language inference. Using ParaNLU, we measure the paraphrastic consistency of several model classes and show that consistency dramatically increases with pretraining but not finetuning. All models tested exhibited room for improvement in paraphrastic consistency.

Paola Cascante-Bonilla, Yu Hou, Yang Trista Cao et al.

Compositional reasoning in Vision-Language Models (VLMs) remains challenging as these models often struggle to relate objects, attributes, and spatial relationships. Recent methods aim to address these limitations by relying on the semantics of the textual description, using Large Language Models (LLMs) to break them down into subsets of questions and answers. However, these methods primarily operate on the surface level, failing to incorporate deeper lexical understanding while introducing incorrect assumptions generated by the LLM. In response to these issues, we present Caption Expansion with Contradictions and Entailments (CECE), a principled approach that leverages Natural Language Inference (NLI) to generate entailments and contradictions from a given premise. CECE produces lexically diverse sentences while maintaining their core meaning. Through extensive experiments, we show that CECE enhances interpretability and reduces overreliance on biased or superficial features. By balancing CECE along the original premise, we achieve significant improvements over previous methods without requiring additional fine-tuning, producing state-of-the-art results on benchmarks that score agreement with human judgments for image-text alignment, and achieving an increase in performance on Winoground of +19.2% (group score) and +12.9% on EqBen (group score) over the best prior work (finetuned with targeted data).

Haozhe An, Connor Baumler, Abhilasha Sancheti et al.

We examine LLM representations of gender for first names in various occupational contexts to study how occupations and the gender perception of first names in LLMs influence each other mutually. We find that LLMs' first-name gender representations correlate with real-world gender statistics associated with the name, and are influenced by the co-occurrence of stereotypically feminine or masculine occupations. Additionally, we study the influence of first-name gender representations on LLMs in a downstream occupation prediction task and their potential as an internal metric to identify extrinsic model biases. While feminine first-name embeddings often raise the probabilities for female-dominated jobs (and vice versa for male-dominated jobs), reliably using these internal gender representations for bias detection remains challenging.

Rupak Sarkar, Neha Srikanth, Taylor Hudson et al.

While it is commonly accepted that maintaining common ground plays a role in conversational success, little prior research exists connecting conversational grounding to success in task-oriented conversations. We study failures of grounding in the Ubuntu IRC dataset, where participants use text-only communication to resolve technical issues. We find that disruptions in conversational flow often stem from a misalignment in common ground, driven by a divergence in beliefs and assumptions held by participants. These disruptions, which we call conversational friction, significantly correlate with task success. We find that although LLMs can identify overt cases of conversational friction, they struggle with subtler and more context-dependent instances requiring pragmatic or domain-specific reasoning.

Shramay Palta, Nirupama Chandrasekaran, Rachel Rudinger et al. · microsoft-research

Using a sample of 25,000 Bing Copilot conversations, we study how the agent responds to users of varying levels of domain expertise and the resulting impact on user experience along multiple dimensions. Our findings show that across a variety of topical domains, the agent largely responds at proficient or expert levels of expertise (77% of conversations) which correlates with positive user experience regardless of the user's level of expertise. Misalignment, such that the agent responds at a level of expertise below that of the user, has a negative impact on overall user experience, with the impact more profound for more complex tasks. We also show that users engage more, as measured by the number of words in the conversation, when the agent responds at a level of expertise commensurate with that of the user. Our findings underscore the importance of alignment between user and AI when designing human-centered AI systems, to ensure satisfactory and productive interactions.

Shramay Palta, Peter Rankel, Sarah Wiegreffe et al. · microsoft-research

We investigate the degree to which human plausibility judgments of multiple-choice commonsense benchmark answers are subject to influence by (im)plausibility arguments for or against an answer, in particular, using rationales generated by LLMs. We collect 3,000 plausibility judgments from humans and another 13,600 judgments from LLMs. Overall, we observe increases and decreases in mean human plausibility ratings in the presence of LLM-generated PRO and CON rationales, respectively, suggesting that, on the whole, human judges find these rationales convincing. Experiments with LLMs reveal similar patterns of influence. Our findings demonstrate a novel use of LLMs for studying aspects of human cognition, while also raising practical concerns that, even in domains where humans are ``experts'' (i.e., common sense), LLMs have the potential to exert considerable influence on people's beliefs.

Nishant Balepur, Atrey Desai, Rachel Rudinger

Large language models (LLMs) now give reasoning before answering, excelling in tasks like multiple-choice question answering (MCQA). Yet, a concern is that LLMs do not solve MCQs as intended, as work finds LLMs sans reasoning succeed in MCQA without using the question, i.e., choices-only. Such partial-input success is often deemed problematic, but reasoning traces could reveal if these strategies are truly shallow in choices-only settings. To study these strategies, reasoning LLMs solve MCQs in full and choices-only inputs; test-time reasoning often boosts accuracy on full and in choices-only half the time. While possibly due to shallow shortcuts, choices-only success is barely affected by the length of reasoning traces, and after finding traces pass faithfulness tests, we show they use less problematic strategies like inferring missing questions. In all, we challenge claims that partial-input success is always a flaw, so we discuss how reasoning traces could separate problematic data from less problematic reasoning.

Nishant Balepur, Matthew Shu, Yoo Yeon Sung et al. · allen-ai, oxford

To assist users in complex tasks, LLMs generate plans: step-by-step instructions towards a goal. While alignment methods aim to ensure LLM plans are helpful, they train (RLHF) or evaluate (ChatbotArena) on what users prefer, assuming this reflects what helps them. We test this with Planorama: an interface where 126 users answer 300 multi-step questions with LLM plans. We get 4388 plan executions and 5584 comparisons to measure plan helpfulness (QA success) and user preferences on plans, and recreate the setup in agents and reward models to see if they simulate or prefer what helps users. We expose: 1) user/model preferences and agent success do not accurately predict which plans help users, so common alignment feedback can misalign with helpfulness; 2) this gap is not due to user-specific preferences, as users are similarly successful when using plans they prefer/disprefer; 3) surface-level cues like brevity and question similarity strongly link to preferences, but such biases fail to predict helpfulness. In all, we argue aligning helpful LLMs needs feedback from real user interactions, not just preferences of what looks helpful, so we discuss the plan NLP researchers can execute to solve this problem.

Mollie Shichman, Claire Bonial, Austin Blodgett et al.

During Human Robot Interactions in disaster relief scenarios, Large Language Models (LLMs) have the potential for substantial physical reasoning to assist in mission objectives. However, these reasoning capabilities are often found only in larger models, which are not currently reasonable to deploy on robotic systems due to size constraints. To meet our problem space requirements, we introduce a dataset and pipeline to create Field Reasoning and Instruction Decoding Agent (FRIDA) models. In our pipeline, domain experts and linguists combine their knowledge to make high-quality, few-shot prompts used to generate synthetic data for fine-tuning. We hand-curate datasets for this few-shot prompting and for evaluation to improve LLM reasoning on both general and disaster-specific objects. We concurrently run an ablation study to understand which kinds of synthetic data most affect performance. We fine-tune several small instruction-tuned models and find that ablated FRIDA models only trained on objects' physical state and function data outperformed both the FRIDA models trained on all synthetic data and the base models in our evaluation. We demonstrate that the FRIDA pipeline is capable of instilling physical common sense with minimal data.

Haozhe An, Christabel Acquaye, Colin Wang et al.

We examine whether large language models (LLMs) exhibit race- and gender-based name discrimination in hiring decisions, similar to classic findings in the social sciences (Bertrand and Mullainathan, 2004). We design a series of templatic prompts to LLMs to write an email to a named job applicant informing them of a hiring decision. By manipulating the applicant's first name, we measure the effect of perceived race, ethnicity, and gender on the probability that the LLM generates an acceptance or rejection email. We find that the hiring decisions of LLMs in many settings are more likely to favor White applicants over Hispanic applicants. In aggregate, the groups with the highest and lowest acceptance rates respectively are masculine White names and masculine Hispanic names. However, the comparative acceptance rates by group vary under different templatic settings, suggesting that LLMs' race- and gender-sensitivity may be idiosyncratic and prompt-sensitive.

Haozhe An, Rachel Rudinger

Through the use of first name substitution experiments, prior research has demonstrated the tendency of social commonsense reasoning models to systematically exhibit social biases along the dimensions of race, ethnicity, and gender (An et al., 2023). Demographic attributes of first names, however, are strongly correlated with corpus frequency and tokenization length, which may influence model behavior independent of or in addition to demographic factors. In this paper, we conduct a new series of first name substitution experiments that measures the influence of these factors while controlling for the others. We find that demographic attributes of a name (race, ethnicity, and gender) and name tokenization length are both factors that systematically affect the behavior of social commonsense reasoning models.

Abhilasha Sancheti, Rachel Rudinger

Script Knowledge (Schank and Abelson, 1975) has long been recognized as crucial for language understanding as it can help in filling in unstated information in a narrative. However, such knowledge is expensive to produce manually and difficult to induce from text due to reporting bias (Gordon and Van Durme, 2013). In this work, we are interested in the scientific question of whether explicit script knowledge is present and accessible through pre-trained generative language models (LMs). To this end, we introduce the task of generating full event sequence descriptions (ESDs) given a scenario in the form of natural language prompts. In zero-shot probing experiments, we find that generative LMs produce poor ESDs with mostly omitted, irrelevant, repeated or misordered events. To address this, we propose a pipeline-based script induction framework (SIF) which can generate good quality ESDs for unseen scenarios (e.g., bake a cake). SIF is a two-staged framework that fine-tunes LM on a small set of ESD examples in the first stage. In the second stage, ESD generated for an unseen scenario is post-processed using RoBERTa-based models to filter irrelevant events, remove repetitions, and reorder the temporally misordered events. Through automatic and manual evaluations, we demonstrate that SIF yields substantial improvements ($1$-$3$ BLUE points) over a fine-tuned LM. However, manual analysis shows that there is great room for improvement, offering a new research direction for inducing script knowledge.

Christine Herlihy, Rachel Rudinger

Crowdworker-constructed natural language inference (NLI) datasets have been found to contain statistical artifacts associated with the annotation process that allow hypothesis-only classifiers to achieve better-than-random performance (Poliak et al., 2018; Gururanganet et al., 2018; Tsuchiya, 2018). We investigate whether MedNLI, a physician-annotated dataset with premises extracted from clinical notes, contains such artifacts (Romanov and Shivade, 2018). We find that entailed hypotheses contain generic versions of specific concepts in the premise, as well as modifiers related to responsiveness, duration, and probability. Neutral hypotheses feature conditions and behaviors that co-occur with, or cause, the condition(s) in the premise. Contradiction hypotheses feature explicit negation of the premise and implicit negation via assertion of good health. Adversarial filtering demonstrates that performance degrades when evaluated on the difficult subset. We provide partition information and recommendations for alternative dataset construction strategies for knowledge-intensive domains.

Noah Weber, Anton Belyy, Nils Holzenberger et al.

Event schemas are structured knowledge sources defining typical real-world scenarios (e.g., going to an airport). We present a framework for efficient human-in-the-loop construction of a schema library, based on a novel script induction system and a well-crafted interface that allows non-experts to "program" complex event structures. Associated with this work we release a schema library: a machine readable resource of 232 detailed event schemas, each of which describe a distinct typical scenario in terms of its relevant sub-event structure (what happens in the scenario), participants (who plays a role in the scenario), fine-grained typing of each participant, and the implied relational constraints between them. We make our schema library and the SchemaBlocks interface available online.

Faeze Brahman, Vered Shwartz, Rachel Rudinger et al.

The black-box nature of neural models has motivated a line of research that aims to generate natural language rationales to explain why a model made certain predictions. Such rationale generation models, to date, have been trained on dataset-specific crowdsourced rationales, but this approach is costly and is not generalizable to new tasks and domains. In this paper, we investigate the extent to which neural models can reason about natural language rationales that explain model predictions, relying only on distant supervision with no additional annotation cost for human-written rationales. We investigate multiple ways to automatically generate rationales using pre-trained language models, neural knowledge models, and distant supervision from related tasks, and train generative models capable of composing explanatory rationales for unseen instances. We demonstrate our approach on the defeasible inference task, a nonmonotonic reasoning task in which an inference may be strengthened or weakened when new information (an update) is introduced. Our model shows promises at generating post-hoc rationales explaining why an inference is more or less likely given the additional information, however, it mostly generates trivial rationales reflecting the fundamental limitations of neural language models. Conversely, the more realistic setup of jointly predicting the update or its type and generating rationale is more challenging, suggesting an important future direction.

Vered Shwartz, Rachel Rudinger, Oyvind Tafjord

Pre-trained language models (LMs) may perpetuate biases originating in their training corpus to downstream models. We focus on artifacts associated with the representation of given names (e.g., Donald), which, depending on the corpus, may be associated with specific entities, as indicated by next token prediction (e.g., Trump). While helpful in some contexts, grounding happens also in under-specified or inappropriate contexts. For example, endings generated for `Donald is a' substantially differ from those of other names, and often have more-than-average negative sentiment. We demonstrate the potential effect on downstream tasks with reading comprehension probes where name perturbation changes the model answers. As a silver lining, our experiments suggest that additional pre-training on different corpora may mitigate this bias.

Noah Weber, Rachel Rudinger, Benjamin Van Durme

When does a sequence of events define an everyday scenario and how can this knowledge be induced from text? Prior works in inducing such scripts have relied on, in one form or another, measures of correlation between instances of events in a corpus. We argue from both a conceptual and practical sense that a purely correlation-based approach is insufficient, and instead propose an approach to script induction based on the causal effect between events, formally defined via interventions. Through both human and automatic evaluations, we show that the output of our method based on causal effects better matches the intuition of what a script represents

Aaron Steven White, Elias Stengel-Eskin, Siddharth Vashishtha et al.

We present the Universal Decompositional Semantics (UDS) dataset (v1.0), which is bundled with the Decomp toolkit (v0.1). UDS1.0 unifies five high-quality, decompositional semantics-aligned annotation sets within a single semantic graph specification---with graph structures defined by the predicative patterns produced by the PredPatt tool and real-valued node and edge attributes constructed using sophisticated normalization procedures. The Decomp toolkit provides a suite of Python 3 tools for querying UDS graphs using SPARQL. Both UDS1.0 and Decomp0.1 are publicly available at http://decomp.io.

Chandler May, Alex Wang, Shikha Bordia et al.

The Word Embedding Association Test shows that GloVe and word2vec word embeddings exhibit human-like implicit biases based on gender, race, and other social constructs (Caliskan et al., 2017). Meanwhile, research on learning reusable text representations has begun to explore sentence-level texts, with some sentence encoders seeing enthusiastic adoption. Accordingly, we extend the Word Embedding Association Test to measure bias in sentence encoders. We then test several sentence encoders, including state-of-the-art methods such as ELMo and BERT, for the social biases studied in prior work and two important biases that are difficult or impossible to test at the word level. We observe mixed results including suspicious patterns of sensitivity that suggest the test's assumptions may not hold in general. We conclude by proposing directions for future work on measuring bias in sentence encoders.

J. Edward Hu, Rachel Rudinger, Matt Post et al.

We present ParaBank, a large-scale English paraphrase dataset that surpasses prior work in both quantity and quality. Following the approach of ParaNMT, we train a Czech-English neural machine translation (NMT) system to generate novel paraphrases of English reference sentences. By adding lexical constraints to the NMT decoding procedure, however, we are able to produce multiple high-quality sentential paraphrases per source sentence, yielding an English paraphrase resource with more than 4 billion generated tokens and exhibiting greater lexical diversity. Using human judgments, we also demonstrate that ParaBank's paraphrases improve over ParaNMT on both semantic similarity and fluency. Finally, we use ParaBank to train a monolingual NMT model with the same support for lexically-constrained decoding for sentence rewriting tasks.

Aaron Steven White, Rachel Rudinger, Kyle Rawlins et al.

We investigate neural models' ability to capture lexicosyntactic inferences: inferences triggered by the interaction of lexical and syntactic information. We take the task of event factuality prediction as a case study and build a factuality judgment dataset for all English clause-embedding verbs in various syntactic contexts. We use this dataset, which we make publicly available, to probe the behavior of current state-of-the-art neural systems, showing that these systems make certain systematic errors that are clearly visible through the lens of factuality prediction.

Adam Poliak, Jason Naradowsky, Aparajita Haldar et al.

We propose a hypothesis only baseline for diagnosing Natural Language Inference (NLI). Especially when an NLI dataset assumes inference is occurring based purely on the relationship between a context and a hypothesis, it follows that assessing entailment relations while ignoring the provided context is a degenerate solution. Yet, through experiments on ten distinct NLI datasets, we find that this approach, which we refer to as a hypothesis-only model, is able to significantly outperform a majority class baseline across a number of NLI datasets. Our analysis suggests that statistical irregularities may allow a model to perform NLI in some datasets beyond what should be achievable without access to the context.

Rachel Rudinger, Jason Naradowsky, Brian Leonard et al.

We present an empirical study of gender bias in coreference resolution systems. We first introduce a novel, Winograd schema-style set of minimal pair sentences that differ only by pronoun gender. With these "Winogender schemas," we evaluate and confirm systematic gender bias in three publicly-available coreference resolution systems, and correlate this bias with real-world and textual gender statistics.

Adam Poliak, Aparajita Haldar, Rachel Rudinger et al.

We present a large-scale collection of diverse natural language inference (NLI) datasets that help provide insight into how well a sentence representation captures distinct types of reasoning. The collection results from recasting 13 existing datasets from 7 semantic phenomena into a common NLI structure, resulting in over half a million labeled context-hypothesis pairs in total. We refer to our collection as the DNC: Diverse Natural Language Inference Collection. The DNC is available online at https://www.decomp.net, and will grow over time as additional resources are recast and added from novel sources.

Rachel Rudinger, Adam Teichert, Ryan Culkin et al.

We present a model for semantic proto-role labeling (SPRL) using an adapted bidirectional LSTM encoding strategy that we call "Neural-Davidsonian": predicate-argument structure is represented as pairs of hidden states corresponding to predicate and argument head tokens of the input sequence. We demonstrate: (1) state-of-the-art results in SPRL, and (2) that our network naturally shares parameters between attributes, allowing for learning new attribute types with limited added supervision.

Rachel Rudinger, Aaron Steven White, Benjamin Van Durme

We present two neural models for event factuality prediction, which yield significant performance gains over previous models on three event factuality datasets: FactBank, UW, and MEANTIME. We also present a substantial expansion of the It Happened portion of the Universal Decompositional Semantics dataset, yielding the largest event factuality dataset to date. We report model results on this extended factuality dataset as well.