Beyond Perplexity: Multi-dimensional Safety Evaluation of LLM Compression

Large language models (LLMs) are increasingly deployed in real-world scenarios with the help of recent model compression techniques. Such momentum towards local deployment means the use of compressed LLMs will widely impact a large population. However, prior analysis works often prioritize on preserving perplexity which is a direct analogy to training loss. The impact of compression method on other critical aspects of model behavior, particularly safety, still calls for a systematic assessment. To this end, we investigate the impact of model compression on four dimensions: (1) degeneration harm, i.e., bias and toxicity in generation; (2) representational harm, i.e., biases in discriminative tasks; (3) dialect bias; (4) language modeling and downstream task performance. We cover a wide spectrum of LLM compression techniques, including unstructured pruning, semi-structured pruning and quantization. Our analysis reveals that compression can lead to unexpected consequences. Although compression may unintentionally remedy LLMs' degeneration harm, it can still exacerbate on the representational harm axis. Although compression may unintentionally remedy LLMs' degeneration harm, it can still exacerbate on the representational harm axis. Moreover, there is a divergent impact on different protected groups as the compression rate grows. Finally, different compression methods have drastically different safety impacts, e.g., quantization mostly preserves bias while pruning degrades quickly. Our findings underscore the importance of integrating safety assessments into the development of compressed LLMs to ensure their reliability across real-world applications. Our full results are available here: \url{https://github.com/zhichaoxu-shufe/Beyond-Perplexity-Compression-Safety-Eval}

An Empirical Investigation of Matrix Factorization Methods for Pre-trained Transformers

The increasing size of transformer-based models in NLP makes the question of compressing them important. In this work, we present a comprehensive analysis of factorization based model compression techniques. Specifically, we focus on comparing straightforward low-rank factorization against the recently introduced Monarch factorization, which exhibits impressive performance preservation on the GLUE benchmark. To mitigate stability issues associated with low-rank factorization of the matrices in pre-trained transformers, we introduce a staged factorization approach wherein layers are factorized one by one instead of being factorized simultaneously. Through this strategy we significantly enhance the stability and reliability of the compression process. Further, we introduce a simple block-wise low-rank factorization method, which has a close relationship to Monarch factorization. Our experiments lead to the surprising conclusion that straightforward low-rank factorization consistently outperforms Monarch factorization across both different compression ratios and six different text classification tasks.

In-Context Example Ordering Guided by Label Distributions

By allowing models to predict without task-specific training, in-context learning (ICL) with pretrained LLMs has enormous potential in NLP. However, a number of problems persist in ICL. In particular, its performance is sensitive to the choice and order of in-context examples. Given the same set of in-context examples with different orderings, model performance may vary between near random to near state-of-the-art. In this work, we formulate in-context example ordering as an optimization problem. We examine three problem settings that differ in the assumptions they make about what is known about the task. Inspired by the idea of learning from label proportions, we propose two principles for in-context example ordering guided by model's probability predictions. We apply our proposed principles to thirteen text classification datasets and nine different autoregressive LLMs with 700M to 13B parameters. We demonstrate that our approach outperforms the baselines by improving the classification accuracy, reducing model miscalibration, and also by selecting better in-context examples.

Promptly Predicting Structures: The Return of Inference

Prompt-based methods have been used extensively across NLP to build zero- and few-shot label predictors. Many NLP tasks are naturally structured: that is, their outputs consist of multiple labels which constrain each other. Annotating data for such tasks can be cumbersome. Can the promise of the prompt-based paradigm be extended to such structured outputs? In this paper, we present a framework for constructing zero- and few-shot linguistic structure predictors. Our key insight is that we can use structural constraints -- and combinatorial inference derived from them -- to filter out inconsistent structures predicted by large language models. We instantiated this framework on two structured prediction tasks, and five datasets. Across all cases, our results show that enforcing consistency not only constructs structurally valid outputs, but also improves performance over the unconstrained variants.

Measuring and Improving Attentiveness to Partial Inputs with Counterfactuals

The inevitable appearance of spurious correlations in training datasets hurts the generalization of NLP models on unseen data. Previous work has found that datasets with paired inputs are prone to correlations between a specific part of the input (e.g., the hypothesis in NLI) and the label; consequently, models trained only on those outperform chance. Are these correlations picked up by models trained on the full input data? To address this question, we propose a new evaluation method, Counterfactual Attentiveness Test (CAT). CAT uses counterfactuals by replacing part of the input with its counterpart from a different example (subject to some restrictions), expecting an attentive model to change its prediction. Using CAT, we systematically investigate established supervised and in-context learning models on ten datasets spanning four tasks: natural language inference, reading comprehension, paraphrase detection, and visual & language reasoning. CAT reveals that reliance on such correlations is mainly data-dependent. Surprisingly, we find that GPT3 becomes less attentive with an increased number of demonstrations, while its accuracy on the test data improves. Our results demonstrate that augmenting training or demonstration data with counterfactuals is effective in improving models' attentiveness. We show that models' attentiveness measured by CAT reveals different conclusions from solely measuring correlations in data.

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.

TempTabQA: Temporal Question Answering for Semi-Structured Tables

Semi-structured data, such as Infobox tables, often include temporal information about entities, either implicitly or explicitly. Can current NLP systems reason about such information in semi-structured tables? To tackle this question, we introduce the task of temporal question answering on semi-structured tables. We present a dataset, TempTabQA, which comprises 11,454 question-answer pairs extracted from 1,208 Wikipedia Infobox tables spanning more than 90 distinct domains. Using this dataset, we evaluate several state-of-the-art models for temporal reasoning. We observe that even the top-performing LLMs lag behind human performance by more than 13.5 F1 points. Given these results, our dataset has the potential to serve as a challenging benchmark to improve the temporal reasoning capabilities of NLP models.

The Integer Linear Programming Inference Cookbook

Over the years, integer linear programs have been employed to model inference in many natural language processing problems. This survey is meant to guide the reader through the process of framing a new inference problem as an instance of an integer linear program and is structured as a collection of recipes. At the end, we will see two worked examples to illustrate the use of these recipes.

Don't Retrain, Just Rewrite: Countering Adversarial Perturbations by Rewriting Text

Can language models transform inputs to protect text classifiers against adversarial attacks? In this work, we present ATINTER, a model that intercepts and learns to rewrite adversarial inputs to make them non-adversarial for a downstream text classifier. Our experiments on four datasets and five attack mechanisms reveal that ATINTER is effective at providing better adversarial robustness than existing defense approaches, without compromising task accuracy. For example, on sentiment classification using the SST-2 dataset, our method improves the adversarial accuracy over the best existing defense approach by more than 4% with a smaller decrease in task accuracy (0.5% vs 2.5%). Moreover, we show that ATINTER generalizes across multiple downstream tasks and classifiers without having to explicitly retrain it for those settings. Specifically, we find that when ATINTER is trained to remove adversarial perturbations for the sentiment classification task on the SST-2 dataset, it even transfers to a semantically different task of news classification (on AGNews) and improves the adversarial robustness by more than 10%.

Learning Semantic Role Labeling from Compatible Label Sequences

This paper addresses the question of how to efficiently learn from disjoint, compatible label sequences. We argue that the compatible structures between disjoint label sets help model learning and inference. We verify this hypothesis on the task of semantic role labeling (SRL), specifically, tagging a sentence with two role sequences: VerbNet arguments and PropBank arguments. Prior work has shown that cross-task interaction improves performance. However, the two tasks are still separately decoded, running the risk of generating structurally inconsistent label sequences (as per lexicons like SEMLINK). To eliminate this issue, we first propose a simple and effective setup that jointly handles VerbNet and PropBank labels as one sequence. With this setup, we show that enforcing SEMLINK constraints during decoding constantly improves the overall F1. With special input constructions, our joint model infers VerbNet arguments from PropBank arguments with over 99% accuracy. We also propose a constrained marginal model that uses SEMLINK information during training to further benefit from the large amounts of PropBank-only data. Our models achieve state-of-the-art F1's on VerbNet and PropBank argument labeling on the CoNLL05 dataset with strong out-of-domain generalization.