As Large Language Models (LLMs) continue to exhibit remarkable performance in natural language understanding tasks, there is a crucial need to measure their ability for human-like multi-step logical reasoning. Existing logical reasoning evaluation benchmarks often focus primarily on simplistic single-step or multi-step reasoning with a limited set of inference rules. Furthermore, the lack of datasets for evaluating non-monotonic reasoning represents a crucial gap since it aligns more closely with human-like reasoning. To address these limitations, we propose Multi-LogiEval, a comprehensive evaluation dataset encompassing multi-step logical reasoning with various inference rules and depths. Multi-LogiEval covers three logic types — propositional, first-order, and non-monotonic consisting of more than 30 inference rules and more than 60 of their combinations with various depths. Leveraging this dataset, we conduct evaluations on a range of LLMs such as GPT-4, ChatGPT, Gemini-Pro, Orca, and Mistral, employing a zero-shot chain-of-thought. Experimental results show that there is a significant drop in the performance of LLMs as the reasoning steps/depth increases (average accuracy of ~68% at depth-1 to ~43% at depth-5). We further conduct a thorough investigation of reasoning chains generated by LLMs which reveals several important findings. We believe that Multi-LogiEval facilitates future research for evaluating and enhancing the logical reasoning ability of LLMs.

Large Language Models (LLMs) have achieved remarkable performance across a wide variety of tasks; however, their large size makes their inference slow and computationally expensive. Focusing on this problem, we study instruction tuning LLMs with additional explicit Losses from the Intermediate layers (LITE) and show that it enables these layers to acquire ‘good’ generation ability without affecting the generation ability of the final layer. We then perform ‘dynamic confidence-based early exiting’ at token level from the intermediate layers which improves the computational efficiency of text generation without sacrificing the quality of the generation. We conduct comprehensive experiments by instruction tuning LLaMA-2 models on the Alpaca dataset and evaluate on four different instruction test sets. We show that dynamic early exiting achieves consistent and considerable inference cost improvements (37.86% for 7B and 46.35% for 13B model) while maintaining the generation quality. We further conduct a thorough analysis of the results and dissect the efficiency improvements which reveals several important findings.

This study explores the sycophantic tendencies of Large Language Models (LLMs), where these models tend to provide answers that match what users want to hear, even if they are not entirely correct. The motivation behind this exploration stems from the common behavior observed in individuals searching the internet for facts with partial or misleading knowledge. Similar to using web search engines, users may recall fragments of misleading keywords and submit them to an LLM, hoping for a comprehensive response. Our empirical analysis of several LLMs shows the potential danger of these models amplifying misinformation when presented with misleading keywords. Additionally, we thoroughly assess four existing hallucination mitigation strategies to reduce LLMs sycophantic behavior. Our experiments demonstrate the effectiveness of these strategies for generating factually correct statements. Furthermore, our analyses delve into knowledge-probing experiments on factual keywords and different categories of sycophancy mitigation.

As Large Language Models (LLMs) play an increasingly pivotal role in natural language processing applications, their safety concerns become critical areas of NLP research. This has resulted in the development of various LLM defense strategies. Unfortunately, despite the shared goal of improving the safety of LLMs, the evaluation suites across various research works are disjoint and lack diverse inputs to ensure accurate and precise evaluation estimates. Furthermore, the important factor of ‘over-defensiveness’ on the safe inputs has largely remained overlooked. Addressing these limitations, this paper presents a systematic evaluation, comparison, and analysis of various LLM defense strategies over both ‘safety’ and ‘over-defensiveness’. To this end, we compile a large and diverse collection of safe and unsafe prompts, design precise evaluation methodology, and study the efficacy of various LLM defense strategies on multiple state-of-the-art LLMs. Our work reveals a number of crucial findings that we believe will pave the way and also facilitate further research in the critical area of improving the safety of LLMs.

Recently developed large language models (LLMs) have been shown to perform remarkably well on a wide range of language understanding tasks. But, can they really “reason” over the natural language? This question has been receiving significant research attention and many reasoning skills such as commonsense, numerical, and qualitative have been studied. However, the crucial skill pertaining to ‘logical reasoning’ has remained underexplored. Existing work investigating this reasoning ability of LLMs has focused only on a couple of inference rules (such as modus ponens and modus tollens) of propositional and first-order logic. Addressing the above limitation, we comprehensively evaluate the logical reasoning ability of LLMs on 25 different reasoning patterns spanning over propositional, first-order, and non-monotonic logics. To enable systematic evaluation, we introduce LogicBench, a natural language question-answering dataset focusing on the use of a single inference rule. We conduct detailed analysis with a range of LLMs such as GPT-4, ChatGPT, Gemini, Llama-2, and Mistral using chain-of-thought prompting. Experimental results show that existing LLMs do not fare well on LogicBench; especially, they struggle with instances involving complex reasoning and negations. Furthermore, they sometimes tend to prioritize parametric knowledge over contextual information and overlook the correct reasoning chain. We believe that our work and findings facilitate future research for evaluating and enhancing the logical reasoning ability of LLMs.

Despite remarkable progress made in natural language processing, even the state-of-the-art models often make incorrect predictions. Such predictions hamper the reliability of systems and limit their widespread adoption in real-world applications. ‘Selective prediction’ partly addresses the above concern by enabling models to abstain from answering when their predictions are likely to be incorrect. While selective prediction is advantageous, it leaves us with a pertinent question ‘what to do after abstention’. To this end, we present an explorative study on ‘Post-Abstention’, a task that allows re-attempting the abstained instances with the aim of increasing **coverage** of the system without significantly sacrificing its **accuracy**. We first provide mathematical formulation of this task and then explore several methods to solve it. Comprehensive experiments on 11 QA datasets show that these methods lead to considerable risk improvements –performance metric of the Post-Abstention task– both in the in-domain and the out-of-domain settings. We also conduct a thorough analysis of these results which further leads to several interesting findings. Finally, we believe that our work will encourage and facilitate further research in this important area of addressing the reliability of NLP systems.

State-of-the-art natural language processing models have been shown to achieve remarkable performance in ‘closed-world’ settings where all the labels in the evaluation set are known at training time. However, in real-world settings, ‘novel’ instances that do not belong to any known class are often observed. This renders the ability to deal with novelties crucial. To initiate a systematic research in this important area of ‘dealing with novelties’, we introduce NoveltyTask, a multi-stage task to evaluate a system’s performance on pipelined novelty ‘detection’ and ‘accommodation’ tasks. We provide mathematical formulation of NoveltyTask and instantiate it with the authorship attribution task that pertains to identifying the correct author of a given text. We use amazon reviews corpus and compile a large dataset (consisting of 250k instances across 200 authors/labels) for NoveltyTask. We conduct comprehensive experiments and explore several baseline methods for the task. Our results show that the methods achieve considerably low performance making the task challenging and leaving sufficient room for improvement. Finally, we believe our work will encourage research in this underexplored area of dealing with novelties, an important step en route to developing robust systems.

Recently Large Language Models (LLMs) such as GPT-3, ChatGPT, and FLAN have led to impressive progress in Natural Language Inference (NLI) tasks. However, these models may rely on simple heuristics or artifacts in the evaluation data to achieve their high performance, which suggests that they still suffer from logical inconsistency. To assess the logical consistency of these models, we propose a LogicAttack, a method to attack NLI models using diverse logical forms of premise and hypothesis, providing a more robust evaluation of their performance. Our approach leverages a range of inference rules from propositional logic, such as Modus Tollens and Bidirectional Dilemma, to generate effective adversarial attacks and identify common vulnerabilities across multiple NLI models. We achieve an average ~53% Attack Success Rate (ASR) across multiple logic-based attacks. Moreover, we demonstrate that incorporating generated attack samples into training enhances the logical reasoning ability of the target model and decreases its vulnerability to logic-based attacks. Data and source code are available at https://github.com/msantoshmadhav/LogicAttack.

In current NLP research, large-scale language models and their abilities are widely being discussed. Some recent works have also found notable failures of these models. Often these failure examples involve complex reasoning abilities. This work focuses on a simple commonsense ability, reasoning about when an action (or its effect) is feasible. To this end, we introduce FeasibilityQA, a question-answering dataset involving binary classification (BCQ) and multi-choice multi-correct questions (MCQ) that test understanding of feasibility. We show that even state-of-the-art models such as GPT-3, GPT-2, and T5 struggle to answer the feasibility questions correctly. Specifically, on (MCQ, BCQ) questions, GPT-3 achieves accuracy of just (19%, 62%) and (25%, 64%) in zero-shot and few-shot settings, respectively. We also evaluate models by providing relevant knowledge statements required to answer the question and find that the additional knowledge leads to a 7% gain in performance, but the overall performance still remains low. These results make one wonder how much commonsense knowledge about action feasibility is encoded in state-of-the-art models and how well they can reason about it.

It’s better to say “I can’t answer” than to answer incorrectly. This selective prediction ability is crucial for NLP systems to be reliably deployed in real-world applications. Prior work has shown that existing selective prediction techniques fail to perform well, especially in the out-of-domain setting. In this work, we propose a method that improves probability estimates of models by calibrating them using prediction confidence and difficulty score of instances. Using these two signals, we first annotate held-out instances and then train a calibrator to predict the likelihood of correctness of the model’s prediction. We instantiate our method with Natural Language Inference (NLI) and Duplicate Detection (DD) tasks and evaluate it in both In-Domain (IID) and Out-of-Domain (OOD) settings. In (IID, OOD) settings, we show that the representations learned by our calibrator result in an improvement of (15.81%, 5.64%) and (6.19%, 13.9%) over ‘MaxProb’ -a selective prediction baseline- on NLI and DD tasks respectively.

Knowledge of difficulty level of questions helps a teacher in several ways, such as estimating students’ potential quickly by asking carefully selected questions and improving quality of examination by modifying trivial and hard questions. Can we extract such benefits of instance difficulty in Natural Language Processing? To this end, we conduct Instance-Level Difficulty Analysis of Evaluation data (ILDAE) in a large-scale setup of 23 datasets and demonstrate its five novel applications: 1) conducting efficient-yet-accurate evaluations with fewer instances saving computational cost and time, 2) improving quality of existing evaluation datasets by repairing erroneous and trivial instances, 3) selecting the best model based on application requirements, 4) analyzing dataset characteristics for guiding future data creation, 5) estimating Out-of-Domain performance reliably. Comprehensive experiments for these applications lead to several interesting results, such as evaluation using just 5% instances (selected via ILDAE) achieves as high as 0.93 Kendall correlation with evaluation using complete dataset and computing weighted accuracy using difficulty scores leads to 5.2% higher correlation with Out-of-Domain performance. We release the difficulty scores and hope our work will encourage research in this important yet understudied field of leveraging instance difficulty in evaluations.

Given the ubiquitous nature of numbers in text, reasoning with numbers to perform simple calculations is an important skill of AI systems. While many datasets and models have been developed to this end, state-of-the-art AI systems are brittle; failing to perform the underlying mathematical reasoning when they appear in a slightly different scenario. Drawing inspiration from GLUE that was proposed in the context of natural language understanding, we propose NumGLUE, a multi-task benchmark that evaluates the performance of AI systems on eight different tasks, that at their core require simple arithmetic understanding. We show that this benchmark is far from being solved with neural models including state-of-the-art large-scale language models performing significantly worse than humans (lower by 46.4 %). Further, NumGLUE promotes sharing knowledge across tasks, especially those with limited training data as evidenced by the superior performance (average gain of 3.4 % on each task) when a model is jointly trained on all the tasks as opposed to task-specific modeling. Finally, we hope that NumGLUE will encourage systems that perform robust and general arithmetic reasoning within language, a first step towards being able to perform more complex mathematical reasoning.

Curriculum learning strategies in prior multitask learning approaches arrange datasets in a difficulty hierarchy either based on human perception or by exhaustively searching the optimal arrangement. However, human perception of difficulty may not always correlate well with machine interpretation leading to poor performance and exhaustive search is computationally expensive. Addressing these concerns, we propose two classes of techniques to arrange training instances into a learning curriculum based on difficulty scores computed via model-based approaches. The two classes i.e Dataset-level and Instance-level differ in granularity of arrangement. Through comprehensive experiments with 12 datasets, we show that instance-level and dataset-level techniques result in strong representations as they lead to an average performance improvement of 4.17% and 3.15% over their respective baselines. Furthermore, we find that most of this improvement comes from correctly answering the difficult instances, implying a greater efficacy of our techniques on difficult tasks

How well can NLP models generalize to a variety of unseen tasks when provided with task instructions? To address this question, we first introduce Super-NaturalInstructions, a benchmark of 1,616 diverse NLP tasks and their expert-written instructions. Our collection covers 76 distinct task types, including but not limited to classification, extraction, infilling, sequence tagging, text rewriting, and text composition. This large and diverse collection of tasks enables rigorous benchmarking of cross-task generalization under instructions—training models to follow instructions on a subset of tasks and evaluating them on the remaining unseen ones.Furthermore, we build Tk-Instruct, a transformer model trained to follow a variety of in-context instructions (plain language task definitions or k-shot examples). Our experiments show that Tk-Instruct outperforms existing instruction-following models such as InstructGPT by over 9% on our benchmark despite being an order of magnitude smaller. We further analyze generalization as a function of various scaling parameters, such as the number of observed tasks, the number of instances per task, and model sizes. We hope our dataset and model facilitate future progress towards more general-purpose NLP models.

Do all instances need inference through the big models for a correct prediction? Perhaps not; some instances are easy and can be answered correctly by even small capacity models. This provides opportunities for improving the computational efficiency of systems. In this work, we present an explorative study on ‘model cascading’, a simple technique that utilizes a collection of models of varying capacities to accurately yet efficiently output predictions. Through comprehensive experiments in multiple task settings that differ in the number of models available for cascading (K value), we show that cascading improves both the computational efficiency and the prediction accuracy. For instance, in K=3 setting, cascading saves up to 88.93% computation cost and consistently achieves superior prediction accuracy with an improvement of up to 2.18%. We also study the impact of introducing additional models in the cascade and show that it further increases the efficiency improvements. Finally, we hope that our work will facilitate development of efficient NLP systems making their widespread adoption in real-world applications possible.

In order to equip NLP systems with ‘selective prediction’ capability, several task-specific approaches have been proposed. However, which approaches work best across tasks or even if they consistently outperform the simplest baseline MaxProb remains to be explored. To this end, we systematically study selective prediction in a large-scale setup of 17 datasets across several NLP tasks. Through comprehensive experiments under in-domain (IID), out-of-domain (OOD), and adversarial (ADV) settings, we show that despite leveraging additional resources (held-out data/computation), none of the existing approaches consistently and considerably outperforms MaxProb in all three settings. Furthermore, their performance does not translate well across tasks. For instance, Monte-Carlo Dropout outperforms all other approaches on Duplicate Detection datasets but does not fare well on NLI datasets, especially in the OOD setting. Thus, we recommend that future selective prediction approaches should be evaluated across tasks and settings for reliable estimation of their capabilities.

Transformer-based models achieve impressive performance on numerous Natural Language Inference (NLI) benchmarks when trained on respective training datasets. However, in certain cases, training samples may not be available or collecting them could be time-consuming and resource-intensive. In this work, we address the above challenge and present an explorative study on unsupervised NLI, a paradigm in which no human-annotated training samples are available. We investigate it under three settings: PH, P, and NPH that differ in the extent of unlabeled data available for learning. As a solution, we propose a procedural data generation approach that leverages a set of sentence transformations to collect PHL (Premise, Hypothesis, Label) triplets for training NLI models, bypassing the need for human-annotated training data. Comprehensive experiments with several NLI datasets show that the proposed approach results in accuracies of up to 66.75%, 65.9%, 65.39% in PH, P, and NPH settings respectively, outperforming all existing unsupervised baselines. Furthermore, fine-tuning our model with as little as ~0.1% of the human-annotated training dataset (500 instances) leads to 12.2% higher accuracy than the model trained from scratch on the same 500 instances. Supported by this superior performance, we conclude with a recommendation for collecting high-quality task-specific data.