LLM-as-a-judge approaches are a practical and effective way of assessing a range of text tasks. However, when using pairwise comparisons to rank a set of candidates, the computational cost scales quadratically with the number of candidates, which has practical limitations. This paper introduces a Product of Expert (PoE) framework for efficient LLM Comparative Assessment. Here individual comparisons are considered experts that provide information on a pair’s score difference. The PoE framework combines the information from these experts to yield an expression that can be maximized with respect to the underlying set of candidates, and is highly flexible where any form of expert can be assumed. When Gaussian experts are used one can derive simple closed-form solutions for the optimal candidate ranking, as well as expressions for selecting which comparisons should be made to maximize the probability of this ranking. Our approach enables efficient comparative assessment, where by using only a small subset of the possible comparisons, one can generate score predictions that correlate well with human judgements. We evaluate the approach on multiple NLG tasks and demonstrate that our framework can yield considerable computational savings when performing pairwise comparative assessment. With many candidate texts, using as few as 2% of comparisons the PoE solution can achieve similar performance to when all comparisons are used.

Large Language Models (LLMs) are powerful zero-shot assessors used in real-world situations such as assessing written exams and benchmarking systems. Despite these critical applications, no existing work has analyzed the vulnerability of judge-LLMs to adversarial manipulation. This work presents the first study on the adversarial robustness of assessment LLMs, where we demonstrate that short universal adversarial phrases can be concatenated to deceive judge LLMs to predict inflated scores. Since adversaries may not know or have access to the judge-LLMs, we propose a simple surrogate attack where a surrogate model is first attacked, and the learned attack phrase then transferred to unknown judge-LLMs. We propose a practical algorithm to determine the short universal attack phrases and demonstrate that when transferred to unseen models, scores can be drastically inflated such that irrespective of the assessed text, maximum scores are predicted. It is found that judge-LLMs are significantly more susceptible to these adversarial attacks when used for absolute scoring, as opposed to comparative assessment. Our findings raise concerns on the reliability of LLM-as-a-judge methods, and emphasize the importance of addressing vulnerabilities in LLM assessment methods before deployment in high-stakes real-world scenarios.

Watermarking generative-AI systems, such as LLMs, has gained considerable interest, driven by their enhanced capabilities across a wide range of tasks. Although current approaches have demonstrated that small, context-dependent shifts in the word distributions can be used to apply and detect watermarks, there has been little work in analyzing the impact that these perturbations have on the quality of generated texts. Balancing high detectability with minimal performance degradation is crucial in terms of selecting the appropriate watermarking setting; therefore this paper proposes a simple analysis framework where comparative assessment, a flexible NLG evaluation framework, is used to assess the quality degradation caused by a particular watermark setting. We demonstrate that our framework provides easy visualization of the quality-detection trade-off of watermark settings, enabling a simple solution to find an LLM watermark operating point that provides a well-balanced performance. This approach is applied to two different summarization systems and a translation system, enabling cross-model analysis for a task, and cross-task analysis.

Large Language Models (LLMs) have demonstrated impressive zero-shot capabilities and versatility in NLP tasks, however they sometimes fail to maintain crucial invariances for specific tasks. One example is permutation sensitivity, where LLMs’ outputs may significantly vary depending on the order of the input options. While debiasing techniques can mitigate these issues, and yield better performance and reliability, they often come with a high computational cost at inference. This paper addresses this inefficiency at inference time. The aim is to distill the capabilities of a computationally intensive, debiased, teacher model into a more compact student model. We explore two variants of student models: one based on pure distillation, and the other on an error-correction approach for more complex tasks, where the student corrects a single biased decision from the teacher to achieve a debiased output. Our approach is general and can be applied to both black-box and white-box LLMs. Furthermore, we demonstrate that our compact, encoder-only student models can outperform their larger, biased teacher counterparts, achieving better results with significantly fewer parameters.

Text and vision foundation models can perform many tasks in a zero-shot setting, a desirable property that enables these systems to be applied in general and low-resource settings. There has been far less work, however, on the zero-shot abilities of ASR foundation models, with these systems typically fine-tuned to specific tasks or constrained to applications that match their training criterion and data annotation. In this work we investigate the ability of Whisper and MMS, ASR foundation models trained primarily for speech recognition, to perform zero-shot audio classification. We use simple template-based text prompts at the decoder and use the resulting decoding probabilities to generate zero-shot predictions. Without training the model on extra data or adding any new parameters, we demonstrate that Whisper shows promising zero-shot classification performance on a range of 8 audio-classification datasets, outperforming the accuracy of existing state-of-the-art zero-shot baselines by an average of 9%. One important step to unlock the emergent ability is debiasing, where a simple unsupervised reweighting method of the class probabilities yields consistent significant performance gains. We further show that performance increases with model size, implying that as ASR foundation models scale up, they may exhibit improved zero-shot performance.

Current developments in large language models (LLMs) have enabled impressive zero-shot capabilities across various natural language tasks. An interesting application of these systems is in the automated assessment of natural language generation (NLG), a highly challenging area with great practical benefit. In this paper, we explore two options for exploiting the emergent abilities of LLMs for zero-shot NLG assessment: absolute score prediction, and comparative assessment which uses relative comparisons between pairs of candidates. Though comparative assessment has not been extensively studied in NLG assessment, we note that humans often find it more intuitive to compare two options rather than scoring each one independently. This work examines comparative assessment from multiple perspectives: performance compared to absolute grading; positional biases in the prompt; and efficient ranking in terms of the number of comparisons. We illustrate that LLM comparative assessment is a simple, general and effective approach for NLG assessment. For moderate-sized open-source LLMs, such as FlanT5 and Llama2-chat, comparative assessment is superior to prompt scoring, and in many cases can achieve performance competitive with state-of-the-art methods. Additionally, we demonstrate that LLMs often exhibit strong positional biases when making pairwise comparisons, and we propose debiasing methods that can further improve performance.

Sequence-to-sequence models often require an expensive autoregressive decoding process. However, for some downstream tasks such as out-of-distribution (OOD) detection and resource allocation, the actual decoding output is not needed, just a scalar attribute of this sequence. In such scenarios, where knowing the quality of a system’s output to predict poor performance prevails over knowing the output itself, is it possible to bypass the autoregressive decoding? We propose Non-Autoregressive Proxy (NAP) models that can efficiently predict scalar-valued sequence-level attributes. Importantly, NAPs predict these metrics directly from the encodings, avoiding the expensive decoding stage. We consider two sequence tasks: Machine Translation (MT) and Automatic Speech Recognition (ASR). In OOD for MT, NAPs outperform ensembles while being significantly faster. NAPs are also proven capable of predicting metrics such as BERTScore (MT) or word error rate (ASR). For downstream tasks, such as data filtering and resource optimization, NAPs generate performance predictions that outperform predictive uncertainty while being highly inference efficient.

In this paper, we consider the challenge of summarizing patients medical progress notes in a limited data setting. For the Problem List Summarization (shared task 1A) at the BioNLP Workshop 2023, we demonstrate that ClinicalT5 fine-tuned to 765 medical clinic notes outperforms other extractive, abstractive and zero-shot baselines, yielding reasonable baseline systems for medical note summarization. Further, we introduce Hierarchical Ensemble of Summarization Models (HESM), consisting of token-level ensembles of diverse fine-tuned ClinicalT5 models, followed by Minimum Bayes Risk (MBR) decoding. Our HESM approach lead to a considerable summarization performance boost, and when evaluated on held-out challenge data achieved a ROUGE-L of 32.77, which was the best-performing system at the top of the shared task leaderboard.

Recently it has been shown that without any access to the contextual passage, multiple choice reading comprehension (MCRC) systems are able to answer questions significantly better than random on average. These systems use their accumulated “world knowledge” to directly answer questions, rather than using information from the passage. This paper examines the possibility of exploiting this observation as a tool for test designers to ensure that the form of “world knowledge” is acceptable for a particular set of questions. We propose information-theory based metrics that enable the level of “world knowledge” exploited by systems to be assessed. Two metrics are described: the expected number of options, which measures whether a passage-free system can identify the answer a question using world knowledge; and the contextual mutual information, which measures the importance of context for a given question. We demonstrate that questions with low expected number of options, and hence answerable by the shortcut system, are often similarly answerable by humans without context. This highlights that the general knowledge ‘shortcuts’ could be equally used by exam candidates, and that our proposed metrics may be helpful for future test designers to monitor the quality of questions.

Generative Large Language Models (LLMs) such as GPT-3 are capable of generating highly fluent responses to a wide variety of user prompts. However, LLMs are known to hallucinate facts and make non-factual statements which can undermine trust in their output. Existing fact-checking approaches either require access to the output probability distribution (which may not be available for systems such as ChatGPT) or external databases that are interfaced via separate, often complex, modules. In this work, we propose “SelfCheckGPT”, a simple sampling-based approach that can be used to fact-check the responses of black-box models in a zero-resource fashion, i.e. without an external database. SelfCheckGPT leverages the simple idea that if an LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another. We investigate this approach by using GPT-3 to generate passages about individuals from the WikiBio dataset, and manually annotate the factuality of the generated passages. We demonstrate that SelfCheckGPT can: i) detect non-factual and factual sentences; and ii) rank passages in terms of factuality. We compare our approach to several baselines and show that our approach has considerably higher AUC-PR scores in sentence-level hallucination detection and higher correlation scores in passage-level factuality assessment compared to grey-box methods.

Multiple-choice tests are a common approach for assessing candidates’ comprehension skills. Standard multiple-choice reading comprehension exams require candidates to select the correct answer option from a discrete set based on a question in relation to a contextual passage. For appropriate assessment, the distractor answer options must by definition be incorrect but plausible and diverse. However, generating good quality distractors satisfying these criteria is a challenging task for content creators. We propose automated assessment metrics for the quality of distractors in multiple-choice reading comprehension tests. Specifically, we define quality in terms of the incorrectness, plausibility and diversity of the distractor options. We assess incorrectness using the classification ability of a binary multiple-choice reading comprehension system. Plausibility is assessed by considering the distractor confidence - the probability mass associated with the distractor options for a standard multi-class multiple-choice reading comprehension system. Diversity is assessed by pairwise comparison of an embedding-based equivalence metric between the distractors of a question. To further validate the plausibility metric we compare against candidate distributions over multiple-choice questions and agreement with a ChatGPT model’s interpretation of distractor plausibility and diversity.

Machine learning systems have shown impressive performance across a range of natural language tasks. However, it has been hypothesized that these systems are prone to learning spurious correlations that may be present in the training data. Though these correlations will not impact in-domain performance, they are unlikely to generalize well to out-of-domain data, limiting the applicability of systems. This work examines this phenomenon on text classification tasks. Rather than artificially injecting features into the data, we demonstrate that real spurious correlations can be exploited by current state-of-the-art deep-learning systems. Specifically, we show that even when only ‘stop’ words are available at the input stage, it is possible to predict the class significantly better than random. Though it is shown that these stop words are not required for good in-domain performance, they can degrade the ability of the system to generalize well to out-of-domain data.