Measuring Faithfulness of Chains of Thought by Unlearning Reasoning Steps

When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. However, despite much work on CoT prompting, it is unclear if CoT reasoning is faithful to the models' parameteric beliefs. We introduce a framework for measuring parametric faithfulness of generated reasoning, and propose Faithfulness by Unlearning Reasoning steps (FUR), an instance of this framework. FUR erases information contained in reasoning steps from model parameters. We perform experiments unlearning CoTs of four LMs prompted on four multi-choice question answering (MCQA) datasets. Our experiments show that FUR is frequently able to change the underlying models' prediction by unlearning key steps, indicating when a CoT is parametrically faithful. Further analysis shows that CoTs generated by models post-unlearning support different answers, hinting at a deeper effect of unlearning. Importantly, CoT steps identified as important by FUR do not align well with human notions of plausbility, emphasizing the need for specialized alignment

On Evaluating Explanation Utility for Human-AI Decision Making in NLP

Is explainability a false promise? This debate has emerged from the insufficient evidence that explanations aid people in situations they are introduced for. More human-centered, application-grounded evaluations of explanations are needed to settle this. Yet, with no established guidelines for such studies in NLP, researchers accustomed to standardized proxy evaluations must discover appropriate measurements, tasks, datasets, and sensible models for human-AI teams in their studies. To help with this, we first review fitting existing metrics. We then establish requirements for datasets to be suitable for application-grounded evaluations. Among over 50 datasets available for explainability research in NLP, we find that 4 meet our criteria. By finetuning Flan-T5-3B, we demonstrate the importance of reassessing the state of the art to form and study human-AI teams. Finally, we present the exemplar studies of human-AI decision-making for one of the identified suitable tasks -- verifying the correctness of a legal claim given a contract.

Chain-of-Thought Unfaithfulness as Disguised Accuracy

Understanding the extent to which Chain-of-Thought (CoT) generations align with a large language model's (LLM) internal computations is critical for deciding whether to trust an LLM's output. As a proxy for CoT faithfulness, arXiv:2307.13702 propose a metric that measures a model's dependence on its CoT for producing an answer. Within a single family of proprietary models, they find that LLMs exhibit a scaling-then-inverse-scaling relationship between model size and their measure of faithfulness, and that a 13 billion parameter model exhibits increased faithfulness compared to models ranging from 810 million to 175 billion parameters in size. We evaluate whether these results generalize as a property of all LLMs. We replicate their experimental setup with three different families of models and, under specific conditions, successfully reproduce the scaling trends for CoT faithfulness they report. However, we discover that simply changing the order of answer choices in the prompt can reduce the metric by 73 percentage points. The faithfulness metric is also highly correlated ($R^2$ = 0.91) with accuracy, raising doubts about its validity as a construct for evaluating faithfulness.

Whispers of Doubt Amidst Echoes of Triumph in NLP Robustness

Are the longstanding robustness issues in NLP resolved by today's larger and more performant models? To address this question, we conduct a thorough investigation using 19 models of different sizes spanning different architectural choices and pretraining objectives. We conduct evaluations using (a) OOD and challenge test sets, (b) CheckLists, (c) contrast sets, and (d) adversarial inputs. Our analysis reveals that not all OOD tests provide further insight into robustness. Evaluating with CheckLists and contrast sets shows significant gaps in model performance; merely scaling models does not make them sufficiently robust. Finally, we point out that current approaches for adversarial evaluations of models are themselves problematic: they can be easily thwarted, and in their current forms, do not represent a sufficiently deep probe of model robustness. We conclude that not only is the question of robustness in NLP as yet unresolved, but even some of the approaches to measure robustness need to be reassessed.

How Much Consistency Is Your Accuracy Worth?

Contrast set consistency is a robustness measurement that evaluates the rate at which a model correctly responds to all instances in a bundle of minimally different examples relying on the same knowledge. To draw additional insights, we propose to complement consistency with relative consistency -- the probability that an equally accurate model would surpass the consistency of the proposed model, given a distribution over possible consistencies. Models with 100% relative consistency have reached a consistency peak for their accuracy. We reflect on prior work that reports consistency in contrast sets and observe that relative consistency can alter the assessment of a model's consistency compared to another. We anticipate that our proposed measurement and insights will influence future studies aiming to promote consistent behavior in models.

CONDAQA: A Contrastive Reading Comprehension Dataset for Reasoning about Negation

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.

Does Self-Rationalization Improve Robustness to Spurious Correlations?

Rationalization is fundamental to human reasoning and learning. NLP models trained to produce rationales along with predictions, called self-rationalization models, have been investigated for their interpretability and utility to end-users. However, the extent to which training with human-written rationales facilitates learning remains an under-explored question. We ask whether training models to self-rationalize can aid in their learning to solve tasks for the right reasons. Specifically, we evaluate how training self-rationalization models with free-text rationales affects robustness to spurious correlations in fine-tuned encoder-decoder and decoder-only models of six different sizes. We evaluate robustness to spurious correlations by measuring performance on 1) manually annotated challenge datasets and 2) subsets of original test sets where reliance on spurious correlations would fail to produce correct answers. We find that while self-rationalization can improve robustness to spurious correlations in low-resource settings, it tends to hurt robustness in higher-resource settings. Furthermore, these effects depend on model family and size, as well as on rationale content. Together, our results suggest that explainability can come at the cost of robustness; thus, appropriate care should be taken when training self-rationalizing models with the goal of creating more trustworthy models.

Do Androids Laugh at Electric Sheep? Humor "Understanding" Benchmarks from The New Yorker Caption Contest

We challenge AI models to "demonstrate understanding" of the sophisticated multimodal humor of The New Yorker Caption Contest. Concretely, we develop three carefully circumscribed tasks for which it suffices (but is not necessary) to grasp potentially complex and unexpected relationships between image and caption, and similarly complex and unexpected allusions to the wide varieties of human experience; these are the hallmarks of a New Yorker-caliber cartoon. We investigate vision-and-language models that take as input the cartoon pixels and caption directly, as well as language-only models for which we circumvent image-processing by providing textual descriptions of the image. Even with the rich multifaceted annotations we provide for the cartoon images, we identify performance gaps between high-quality machine learning models (e.g., a fine-tuned, 175B parameter language model) and humans. We publicly release our corpora including annotations describing the image's locations/entities, what's unusual about the scene, and an explanation of the joke.

Few-Shot Self-Rationalization with Natural Language Prompts

Self-rationalization models that predict task labels and generate free-text elaborations for their predictions could enable more intuitive interaction with NLP systems. These models are, however, currently trained with a large amount of human-written free-text explanations for each task which hinders their broader usage. We propose to study a more realistic setting of self-rationalization using few training examples. We present FEB -- a standardized collection of four existing English-language datasets and associated metrics. We identify the right prompting approach by extensively exploring natural language prompts on FEB. Then, by using this prompt and scaling the model size, we demonstrate that making progress on few-shot self-rationalization is possible. We show there is still ample room for improvement in this task: the average plausibility of generated explanations assessed by human annotators is at most 51%, while plausibility of human explanations is 76%. We hope that FEB together with our proposed approach will spur the community to take on the few-shot self-rationalization challenge.

Effective Attention Sheds Light On Interpretability

An attention matrix of a transformer self-attention sublayer can provably be decomposed into two components and only one of them (effective attention) contributes to the model output. This leads us to ask whether visualizing effective attention gives different conclusions than interpretation of standard attention. Using a subset of the GLUE tasks and BERT, we carry out an analysis to compare the two attention matrices, and show that their interpretations differ. Effective attention is less associated with the features related to the language modeling pretraining such as the separator token, and it has more potential to illustrate linguistic features captured by the model for solving the end-task. Given the found differences, we recommend using effective attention for studying a transformer's behavior since it is more pertinent to the model output by design.