Abhilasha Ravichander

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.

Yanai Elazar, Akshita Bhagia, Ian Magnusson et al. · allen-ai, berkeley

Large text corpora are the backbone of language models. However, we have a limited understanding of the content of these corpora, including general statistics, quality, social factors, and inclusion of evaluation data (contamination). In this work, we propose What's In My Big Data? (WIMBD), a platform and a set of sixteen analyses that allow us to reveal and compare the contents of large text corpora. WIMBD builds on two basic capabilities -- count and search -- at scale, which allows us to analyze more than 35 terabytes on a standard compute node. We apply WIMBD to ten different corpora used to train popular language models, including C4, The Pile, and RedPajama. Our analysis uncovers several surprising and previously undocumented findings about these corpora, including the high prevalence of duplicate, synthetic, and low-quality content, personally identifiable information, toxic language, and benchmark contamination. For instance, we find that about 50% of the documents in RedPajama and LAION-2B-en are duplicates. In addition, several datasets used for benchmarking models trained on such corpora are contaminated with respect to important benchmarks, including the Winograd Schema Challenge and parts of GLUE and SuperGLUE. We open-source WIMBD's code and artifacts to provide a standard set of evaluations for new text-based corpora and to encourage more analyses and transparency around them.

Yanai Elazar, Nora Kassner, Shauli Ravfogel et al. · cmu

Large amounts of training data are one of the major reasons for the high performance of state-of-the-art NLP models. But what exactly in the training data causes a model to make a certain prediction? We seek to answer this question by providing a language for describing how training data influences predictions, through a causal framework. Importantly, our framework bypasses the need to retrain expensive models and allows us to estimate causal effects based on observational data alone. Addressing the problem of extracting factual knowledge from pretrained language models (PLMs), we focus on simple data statistics such as co-occurrence counts and show that these statistics do influence the predictions of PLMs, suggesting that such models rely on shallow heuristics. Our causal framework and our results demonstrate the importance of studying datasets and the benefits of causality for understanding NLP models.

Abhilasha Ravichander, Matt Gardner, Ana Marasović · allen-ai, cmu

The full power of human language-based communication cannot be realized without negation. All human languages have some form of negation. Despite this, negation remains a challenging phenomenon for current natural language understanding systems. To facilitate the future development of models that can process negation effectively, we present CONDAQA, the first English reading comprehension dataset which requires reasoning about the implications of negated statements in paragraphs. We collect paragraphs with diverse negation cues, then have crowdworkers ask questions about the implications of the negated statement in the passage. We also have workers make three kinds of edits to the passage -- paraphrasing the negated statement, changing the scope of the negation, and reversing the negation -- resulting in clusters of question-answer pairs that are difficult for models to answer with spurious shortcuts. CONDAQA features 14,182 question-answer pairs with over 200 unique negation cues and is challenging for current state-of-the-art models. The best performing model on CONDAQA (UnifiedQA-v2-3b) achieves only 42% on our consistency metric, well below human performance which is 81%. We release our dataset, along with fully-finetuned, few-shot, and zero-shot evaluations, to facilitate the development of future NLP methods that work on negated language.

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.

Yufei Tian, Abhilasha Ravichander, Lianhui Qin et al. · cmu

We explore the creative problem-solving capabilities of modern LLMs in a novel constrained setting. To this end, we create MACGYVER, an automatically generated dataset consisting of over 1,600 real-world problems deliberately designed to trigger innovative usage of objects and necessitate out-of-the-box thinking. We then present our collection to both LLMs and humans to compare and contrast their problem-solving abilities. MACGYVER is challenging for both groups, but in unique and complementary ways. For instance, humans excel in tasks they are familiar with but struggle with domain-specific knowledge, leading to a higher variance. In contrast, LLMs, exposed to a variety of specialized knowledge, attempt broader problems but fail by proposing physically-infeasible actions. Finally, we provide a detailed error analysis of LLMs, and demonstrate the potential of enhancing their problem-solving ability with novel prompting techniques such as iterative step-wise reflection and divergent-convergent thinking. This work (1) introduces a fresh arena for intelligent agents focusing on intricate aspects of physical reasoning, planning, and unconventional thinking, which supplements the existing spectrum of machine intelligence; and (2) provides insight into the constrained problem-solving capabilities of both humans and AI.

Wenting Zhao, Tanya Goyal, Yu Ying Chiu et al. · cmu, uw

While hallucinations of large language models (LLMs) prevail as a major challenge, existing evaluation benchmarks on factuality do not cover the diverse domains of knowledge that the real-world users of LLMs seek information about. To bridge this gap, we introduce WildHallucinations, a benchmark that evaluates factuality. It does so by prompting LLMs to generate information about entities mined from user-chatbot conversations in the wild. These generations are then automatically fact-checked against a systematically curated knowledge source collected from web search. Notably, half of these real-world entities do not have associated Wikipedia pages. We evaluate 118,785 generations from 15 LLMs on 7,919 entities. We find that LLMs consistently hallucinate more on entities without Wikipedia pages and exhibit varying hallucination rates across different domains. Finally, given the same base models, adding a retrieval component only slightly reduces hallucinations but does not eliminate hallucinations.

Mohammad Aflah Khan, Krishna P. Gummadi, Manish Gupta et al. · cmu

Rotary Positional Embedding (RoPE) is a common choice in transformer architectures for encoding relative positional information. Although earlier work has examined omitting RoPE in specific layers, the effect of varying the fraction of hidden dimensions that receive rotary transformations remains largely unexplored. This design choice can yield substantial memory savings, which becomes especially significant at long context lengths. We find up to 10x memory savings over the standard RoPE cache, while achieving comparable final loss. In this work, we present a systematic study examining the impact of partial RoPE on training dynamics and convergence across architectures and datasets. Our findings uncover several notable patterns: (1) applying RoPE to only a small fraction of dimensions (around 10%) achieves convergence comparable to using full RoPE; (2) these trends hold consistently across model size, sequence lengths and datasets of varying quality and architectures, with higher-quality data resulting in lower overall loss and similar benchmark performance; and (3) some models trained with NoPE (No Positional Encoding) showcase unstable learning trajectories, which can be alleviated through minimal RoPE application or QK-Norm which converges to a higher loss. Together, these results offer practical guidance for model designers aiming to balance efficiency and training stability, while emphasizing the previously overlooked importance of partial RoPE.

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.

Mohammad Aflah Khan, Mahsa Amani, Soumi Das et al. · cmu

Agents based on Large Language Models (LLMs) are increasingly being deployed as interfaces to information on online platforms. These agents filter, prioritize, and synthesize information retrieved from the platforms' back-end databases or via web search. In these scenarios, LLM agents govern the information users receive, by drawing users' attention to particular instances of retrieved information at the expense of others. While much prior work has focused on biases in the information LLMs themselves generate, less attention has been paid to the factors that influence what information LLMs select and present to users. We hypothesize that when information is attributed to specific sources (e.g., particular publishers, journals, or platforms), current LLMs exhibit systematic latent source preferences- that is, they prioritize information from some sources over others. Through controlled experiments on twelve LLMs from six model providers, spanning both synthetic and real-world tasks, we find that several models consistently exhibit strong and predictable source preferences. These preferences are sensitive to contextual framing, can outweigh the influence of content itself, and persist despite explicit prompting to avoid them. They also help explain phenomena such as the observed left-leaning skew in news recommendations in prior work. Our findings advocate for deeper investigation into the origins of these preferences, as well as for mechanisms that provide users with transparency and control over the biases guiding LLM-powered agents.

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.

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.

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.

Abhilasha Ravichander, Shrusti Ghela, David Wadden et al. · allen-ai, cmu

Despite their impressive ability to generate high-quality and fluent text, generative large language models (LLMs) also produce hallucinations: statements that are misaligned with established world knowledge or provided input context. However, measuring hallucination can be challenging, as having humans verify model generations on-the-fly is both expensive and time-consuming. In this work, we release HALoGEN, a comprehensive hallucination benchmark consisting of: (1) 10,923 prompts for generative models spanning nine domains including programming, scientific attribution, and summarization, and (2) automatic high-precision verifiers for each use case that decompose LLM generations into atomic units, and verify each unit against a high-quality knowledge source. We use this framework to evaluate ~150,000 generations from 14 language models, finding that even the best-performing models are riddled with hallucinations (sometimes up to 86% of generated atomic facts depending on the domain). We further define a novel error classification for LLM hallucinations based on whether they likely stem from incorrect recollection of training data (Type A errors), or incorrect knowledge in training data (Type B errors), or are fabrication (Type C errors). We hope our framework provides a foundation to enable the principled study of why generative models hallucinate, and advances the development of trustworthy large language models.

Qinyuan Wu, Soumi Das, Mahsa Amani et al.

Agentic AI architectures augment LLMs with external tools, unlocking strong capabilities. However, tool use is not always beneficial; some calls may be redundant or even harmful. Effective tool use, therefore, hinges on a core LLM decision: whether to call or not call a tool, when performing a task. This decision is particularly challenging for web search tools, where the benefits of external information depend on the model's internal knowledge and its ability to integrate potentially noisy tool responses. We introduce a principled framework inspired by decision-making theory to evaluate web search tool-use decisions along three key factors: necessity, utility, and affordability. Our analysis combines two complementary lenses: a normative perspective that infers true need and utility from an optimal allocation of tool calls, and a descriptive perspective that infers the model's self-perceived need and utility from their observed behaviors. We find that models' perceived need and utility of tool calls are often misaligned with their true need and utility. Building on this framework, we train lightweight estimators of need and utility based on models' hidden states. Our estimators enable simple controllers that can improve decision quality and lead to stronger task performance than the self-perceived set up across three tasks and six models.

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.

Yiyou Sun, Yu Gai, Lijie Chen et al. · cmu

Large language models (LLMs) frequently generate hallucinations-content that deviates from factual accuracy or provided context-posing challenges for diagnosis due to the complex interplay of underlying causes. This paper introduces a subsequence association framework to systematically trace and understand hallucinations. Our key insight is that hallucinations arise when dominant hallucinatory associations outweigh faithful ones. Through theoretical and empirical analyses, we demonstrate that decoder-only transformers effectively function as subsequence embedding models, with linear layers encoding input-output associations. We propose a tracing algorithm that identifies causal subsequences by analyzing hallucination probabilities across randomized input contexts. Experiments show our method outperforms standard attribution techniques in identifying hallucination causes and aligns with evidence from the model's training corpus. This work provides a unified perspective on hallucinations and a robust framework for their tracing and analysis.

Abhilasha Ravichander, Jillian Fisher, Taylor Sorensen et al. · cmu, uw

High-quality training data has proven crucial for developing performant large language models (LLMs). However, commercial LLM providers disclose few, if any, details about the data used for training. This lack of transparency creates multiple challenges: it limits external oversight and inspection of LLMs for issues such as copyright infringement, it undermines the agency of data authors, and it hinders scientific research on critical issues such as data contamination and data selection. How can we recover what training data is known to LLMs? In this work, we demonstrate a new method to identify training data known to proprietary LLMs like GPT-4 without requiring any access to model weights or token probabilities, by using information-guided probes. Our work builds on a key observation: text passages with high surprisal are good search material for memorization probes. By evaluating a model's ability to successfully reconstruct high-surprisal tokens in text, we can identify a surprising number of texts memorized by LLMs.

Skyler Hallinan, Jaehun Jung, Melanie Sclar et al. · uw

Membership inference attacks serves as useful tool for fair use of language models, such as detecting potential copyright infringement and auditing data leakage. However, many current state-of-the-art attacks require access to models' hidden states or probability distribution, which prevents investigation into more widely-used, API-access only models like GPT-4. In this work, we introduce N-Gram Coverage Attack, a membership inference attack that relies solely on text outputs from the target model, enabling attacks on completely black-box models. We leverage the observation that models are more likely to memorize and subsequently generate text patterns that were commonly observed in their training data. Specifically, to make a prediction on a candidate member, N-Gram Coverage Attack first obtains multiple model generations conditioned on a prefix of the candidate. It then uses n-gram overlap metrics to compute and aggregate the similarities of these outputs with the ground truth suffix; high similarities indicate likely membership. We first demonstrate on a diverse set of existing benchmarks that N-Gram Coverage Attack outperforms other black-box methods while also impressively achieving comparable or even better performance to state-of-the-art white-box attacks - despite having access to only text outputs. Interestingly, we find that the success rate of our method scales with the attack compute budget - as we increase the number of sequences generated from the target model conditioned on the prefix, attack performance tends to improve. Having verified the accuracy of our method, we use it to investigate previously unstudied closed OpenAI models on multiple domains. We find that more recent models, such as GPT-4o, exhibit increased robustness to membership inference, suggesting an evolving trend toward improved privacy protections.

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.

Benjamin Newman, Abhilasha Ravichander, Jaehun Jung et al. · cmu, uw

Language models are prone to hallucination - generating text that is factually incorrect. Finetuning models on high-quality factual information can potentially reduce hallucination, but concerns remain; obtaining factual gold data can be expensive and training on correct but unfamiliar data may potentially lead to even more downstream hallucination. What data should practitioners finetune on to mitigate hallucinations in language models? In this work, we study the relationship between the factuality of finetuning data and the prevalence of hallucinations in long-form generation tasks. Counterintuitively, we find that finetuning on factual gold data is not as helpful as finetuning on model-generated data that models believe to be factual. Next, we evaluate filtering strategies applied on both factual gold data and model-generated data, and find that finetuning on model-generated data that is filtered by models' own internal judgments often leads to better overall factuality compared to other configurations: training on gold data filtered by models' judgments, training on gold data alone, or training on model-generated data that is supported by gold data. These factuality improvements transfer across three domains we study, suggesting that a models' own beliefs can provide a powerful signal for factuality.

Keivan Rezaei, Mehrdad Saberi, Abhilasha Ravichander et al. · cmu

Large language models are trained on massive corpora of web data, which may include private data, copyrighted material, factually inaccurate data, or data that degrades model performance. Eliminating the influence of such problematic datapoints through complete retraining -- by repeatedly pretraining the model on datasets that exclude these specific instances -- is computationally prohibitive. For this reason, unlearning algorithms have emerged that aim to eliminate the influence of particular datapoints, while otherwise preserving the model -- at a low computational cost. However, precisely estimating and undoing the influence of individual datapoints has proved to be challenging. In this work, we propose a new algorithm, MSA, for estimating and undoing the influence of datapoints -- by leveraging model checkpoints i.e. artifacts capturing model states at different stages of pretraining. Our experimental results demonstrate that MSA consistently outperforms existing machine unlearning algorithms across multiple benchmarks, models, and evaluation metrics, suggesting that MSA could be an effective approach towards more flexible large language models that are capable of data erasure.

Alexander Gill, Abhilasha Ravichander, Ana Marasović · cmu

Large language models (LLMs) are increasingly used for data generation. However, creating evaluation benchmarks raises the bar for this emerging paradigm. Benchmarks must target specific phenomena, penalize exploiting shortcuts, and be challenging. Through two case studies, we investigate whether LLMs can meet these demands by generating reasoning over-text benchmarks and comparing them to those created through careful crowdsourcing. Specifically, we evaluate both the validity and difficulty of LLM-generated versions of two high-quality reading comprehension datasets: CondaQA, which evaluates reasoning about negation, and DROP, which targets reasoning about quantities. We find that prompting LLMs can produce variants of these datasets that are often valid according to the annotation guidelines, at a fraction of the cost of the original crowdsourcing effort. However, we show that they are less challenging for LLMs than their human-authored counterparts. This finding sheds light on what may have been lost by generating evaluation data with LLMs, and calls for critically reassessing the immediate use of this increasingly prevalent approach to benchmark creation.

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.

Dheeraj Rajagopal, Aman Madaan, Niket Tandon et al.

Recently, models have been shown to predict the effects of unexpected situations, e.g., would cloudy skies help or hinder plant growth? Given a context, the goal of such situational reasoning is to elicit the consequences of a new situation (st) that arises in that context. We propose a method to iteratively build a graph of relevant consequences explicitly in a structured situational graph (st-graph) using natural language queries over a finetuned language model (M). Across multiple domains, CURIE generates st-graphs that humans find relevant and meaningful in eliciting the consequences of a new situation. We show that st-graphs generated by CURIE improve a situational reasoning end task (WIQA-QA) by 3 points on accuracy by simply augmenting their input with our generated situational graphs, especially for a hard subset that requires background knowledge and multi-hop reasoning.

Abhilasha Ravichander, Siddharth Dalmia, Maria Ryskina et al.

When Question-Answering (QA) systems are deployed in the real world, users query them through a variety of interfaces, such as speaking to voice assistants, typing questions into a search engine, or even translating questions to languages supported by the QA system. While there has been significant community attention devoted to identifying correct answers in passages assuming a perfectly formed question, we show that components in the pipeline that precede an answering engine can introduce varied and considerable sources of error, and performance can degrade substantially based on these upstream noise sources even for powerful pre-trained QA models. We conclude that there is substantial room for progress before QA systems can be effectively deployed, highlight the need for QA evaluation to expand to consider real-world use, and hope that our findings will spur greater community interest in the issues that arise when our systems actually need to be of utility to humans.

Yanai Elazar, Nora Kassner, Shauli Ravfogel et al.

Consistency of a model -- that is, the invariance of its behavior under meaning-preserving alternations in its input -- is a highly desirable property in natural language processing. In this paper we study the question: Are Pretrained Language Models (PLMs) consistent with respect to factual knowledge? To this end, we create ParaRel, a high-quality resource of cloze-style query English paraphrases. It contains a total of 328 paraphrases for 38 relations. Using ParaRel, we show that the consistency of all PLMs we experiment with is poor -- though with high variance between relations. Our analysis of the representational spaces of PLMs suggests that they have a poor structure and are currently not suitable for representing knowledge robustly. Finally, we propose a method for improving model consistency and experimentally demonstrate its effectiveness.

Aman Madaan, Dheeraj Rajagopal, Yiming Yang et al.

Reasoning about events and tracking their influences is fundamental to understanding processes. In this paper, we present EIGEN - a method to leverage pre-trained language models to generate event influences conditioned on a context, nature of their influence, and the distance in a reasoning chain. We also derive a new dataset for research and evaluation of methods for event influence generation. EIGEN outperforms strong baselines both in terms of automated evaluation metrics (by 10 ROUGE points) and human judgments on closeness to reference and relevance of generations. Furthermore, we show that the event influences generated by EIGEN improve the performance on a "what-if" Question Answering (WIQA) benchmark (over 3% F1), especially for questions that require background knowledge and multi-hop reasoning.

Abhilasha Ravichander, Yonatan Belinkov, Eduard Hovy

Although neural models have achieved impressive results on several NLP benchmarks, little is understood about the mechanisms they use to perform language tasks. Thus, much recent attention has been devoted to analyzing the sentence representations learned by neural encoders, through the lens of `probing' tasks. However, to what extent was the information encoded in sentence representations, as discovered through a probe, actually used by the model to perform its task? In this work, we examine this probing paradigm through a case study in Natural Language Inference, showing that models can learn to encode linguistic properties even if they are not needed for the task on which the model was trained. We further identify that pretrained word embeddings play a considerable role in encoding these properties rather than the training task itself, highlighting the importance of careful controls when designing probing experiments. Finally, through a set of controlled synthetic tasks, we demonstrate models can encode these properties considerably above chance-level even when distributed in the data as random noise, calling into question the interpretation of absolute claims on probing tasks.

Abhilasha Ravichander, Alan W Black, Shomir Wilson et al.

Privacy policies are long and complex documents that are difficult for users to read and understand, and yet, they have legal effects on how user data is collected, managed and used. Ideally, we would like to empower users to inform themselves about issues that matter to them, and enable them to selectively explore those issues. We present PrivacyQA, a corpus consisting of 1750 questions about the privacy policies of mobile applications, and over 3500 expert annotations of relevant answers. We observe that a strong neural baseline underperforms human performance by almost 0.3 F1 on PrivacyQA, suggesting considerable room for improvement for future systems. Further, we use this dataset to shed light on challenges to question answerability, with domain-general implications for any question answering system. The PrivacyQA corpus offers a challenging corpus for question answering, with genuine real-world utility.

Abhilasha Ravichander, Aakanksha Naik, Carolyn Rose et al.

Quantitative reasoning is a higher-order reasoning skill that any intelligent natural language understanding system can reasonably be expected to handle. We present EQUATE (Evaluating Quantitative Understanding Aptitude in Textual Entailment), a new framework for quantitative reasoning in textual entailment. We benchmark the performance of 9 published NLI models on EQUATE, and find that on average, state-of-the-art methods do not achieve an absolute improvement over a majority-class baseline, suggesting that they do not implicitly learn to reason with quantities. We establish a new baseline Q-REAS that manipulates quantities symbolically. In comparison to the best performing NLI model, it achieves success on numerical reasoning tests (+24.2%), but has limited verbal reasoning capabilities (-8.1%). We hope our evaluation framework will support the development of models of quantitative reasoning in language understanding.

Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh et al.

Natural language inference (NLI) is the task of determining if a natural language hypothesis can be inferred from a given premise in a justifiable manner. NLI was proposed as a benchmark task for natural language understanding. Existing models perform well at standard datasets for NLI, achieving impressive results across different genres of text. However, the extent to which these models understand the semantic content of sentences is unclear. In this work, we propose an evaluation methodology consisting of automatically constructed "stress tests" that allow us to examine whether systems have the ability to make real inferential decisions. Our evaluation of six sentence-encoder models on these stress tests reveals strengths and weaknesses of these models with respect to challenging linguistic phenomena, and suggests important directions for future work in this area.

Abhilasha Ravichander, Shruti Rijhwani, Rajat Kulshreshtha et al.

Hierarchical models are utilized in a wide variety of problems which are characterized by task hierarchies, where predictions on smaller subtasks are useful for trying to predict a final task. Typically, neural networks are first trained for the subtasks, and the predictions of these networks are subsequently used as additional features when training a model and doing inference for a final task. In this work, we focus on improving learning for such hierarchical models and demonstrate our method on the task of speaker trait prediction. Speaker trait prediction aims to computationally identify which personality traits a speaker might be perceived to have, and has been of great interest to both the Artificial Intelligence and Social Science communities. Persuasiveness prediction in particular has been of interest, as persuasive speakers have a large amount of influence on our thoughts, opinions and beliefs. In this work, we examine how leveraging the relationship between related speaker traits in a hierarchical structure can help improve our ability to predict how persuasive a speaker is. We present a novel algorithm that allows us to backpropagate through this hierarchy. This hierarchical model achieves a 25% relative error reduction in classification accuracy over current state-of-the art methods on the publicly available POM dataset.

Paul Michel, Abhilasha Ravichander, Shruti Rijhwani

We investigate the pertinence of methods from algebraic topology for text data analysis. These methods enable the development of mathematically-principled isometric-invariant mappings from a set of vectors to a document embedding, which is stable with respect to the geometry of the document in the selected metric space. In this work, we evaluate the utility of these topology-based document representations in traditional NLP tasks, specifically document clustering and sentiment classification. We find that the embeddings do not benefit text analysis. In fact, performance is worse than simple techniques like $\textit{tf-idf}$, indicating that the geometry of the document does not provide enough variability for classification on the basis of topic or sentiment in the chosen datasets.