Abstract:Human label variation (HLV) is a valuable source of information that arises when multiple human annotators provide different labels for valid reasons. In Natural Language Inference (NLI) earlier approaches to capturing HLV involve either collecting annotations from many crowd workers to represent human judgment distribution (HJD) or use expert linguists to provide detailed explanations for their chosen labels. While the former method provides denser HJD information, obtaining it is resource-intensive. In contrast, the latter offers richer textual information but it is challenging to scale up to many human judges. Besides, large language models (LLMs) are increasingly used as evaluators (``LLM judges'') but with mixed results, and few works aim to study HJDs. This study proposes to exploit LLMs to approximate HJDs using a small number of expert labels and explanations. Our experiments show that a few explanations significantly improve LLMs' ability to approximate HJDs with and without explicit labels, thereby providing a solution to scale up annotations for HJD. However, fine-tuning smaller soft-label aware models with the LLM-generated model judgment distributions (MJDs) presents partially inconsistent results: while similar in distance, their resulting fine-tuned models and visualized distributions differ substantially. We show the importance of complementing instance-level distance measures with a global-level shape metric and visualization to more effectively evaluate MJDs against human judgment distributions.
Abstract:Recent advances in Large Language Models (LLMs) have sparked wide interest in validating and comprehending the human-like cognitive-behavioral traits LLMs may have. These cognitive-behavioral traits include typically Attitudes, Opinions, Values (AOV). However, measuring AOV embedded within LLMs remains opaque, and different evaluation methods may yield different results. This has led to a lack of clarity on how different studies are related to each other and how they can be interpreted. This paper aims to bridge this gap by providing an overview of recent works on the evaluation of AOV in LLMs. Moreover, we survey related approaches in different stages of the evaluation pipeline in these works. By doing so, we address the potential and challenges with respect to understanding the model, human-AI alignment, and downstream application in social sciences. Finally, we provide practical insights into evaluation methods, model enhancement, and interdisciplinary collaboration, thereby contributing to the evolving landscape of evaluating AOV in LLMs.
Abstract:Overparametrized transformer networks are the state-of-the-art architecture for Large Language Models (LLMs). However, such models contain billions of parameters making large compute a necessity, while raising environmental concerns. To address these issues, we propose FinerCut, a new form of fine-grained layer pruning, which in contrast to prior work at the transformer block level, considers all self-attention and feed-forward network (FFN) layers within blocks as individual pruning candidates. FinerCut prunes layers whose removal causes minimal alternation to the model's output -- contributing to a new, lean, interpretable, and task-agnostic pruning method. Tested across 9 benchmarks, our approach retains 90% performance of Llama3-8B with 25% layers removed, and 95% performance of Llama3-70B with 30% layers removed, all without fine-tuning or post-pruning reconstruction. Strikingly, we observe intriguing results with FinerCut: 42% (34 out of 80) of the self-attention layers in Llama3-70B can be removed while preserving 99% of its performance -- without additional fine-tuning after removal. Moreover, FinerCut provides a tool to inspect the types and locations of pruned layers, allowing to observe interesting pruning behaviors. For instance, we observe a preference for pruning self-attention layers, often at deeper consecutive decoder layers. We hope our insights inspire future efficient LLM architecture designs.
Abstract:Multiple choice questions (MCQs) are commonly used to evaluate the capabilities of large language models (LLMs). One common way to evaluate the model response is to rank the candidate answers based on the log probability of the first token prediction. An alternative way is to examine the text output. Prior work has shown that first token probabilities lack robustness to changes in MCQ phrasing, and that first token probabilities do not match text answers for instruction-tuned models. Therefore, in this paper, we investigate the robustness of text answers. We show that the text answers are more robust to question perturbations than the first token probabilities, when the first token answers mismatch the text answers. The difference in robustness increases as the mismatch rate becomes greater. As the mismatch reaches over 50\%, the text answer is more robust to option order changes than the debiased first token probabilities using state-of-the-art debiasing methods such as PriDe. Our findings provide further evidence for the benefits of text answer evaluation over first token probability evaluation.
Abstract:Big models have achieved revolutionary breakthroughs in the field of AI, but they might also pose potential concerns. Addressing such concerns, alignment technologies were introduced to make these models conform to human preferences and values. Despite considerable advancements in the past year, various challenges lie in establishing the optimal alignment strategy, such as data cost and scalable oversight, and how to align remains an open question. In this survey paper, we comprehensively investigate value alignment approaches. We first unpack the historical context of alignment tracing back to the 1920s (where it comes from), then delve into the mathematical essence of alignment (what it is), shedding light on the inherent challenges. Following this foundation, we provide a detailed examination of existing alignment methods, which fall into three categories: Reinforcement Learning, Supervised Fine-Tuning, and In-context Learning, and demonstrate their intrinsic connections, strengths, and limitations, helping readers better understand this research area. In addition, two emerging topics, personal alignment, and multimodal alignment, are also discussed as novel frontiers in this field. Looking forward, we discuss potential alignment paradigms and how they could handle remaining challenges, prospecting where future alignment will go.
Abstract:The open-ended nature of language generation makes the evaluation of autoregressive large language models (LLMs) challenging. One common evaluation approach uses multiple-choice questions (MCQ) to limit the response space. The model is then evaluated by ranking the candidate answers by the log probability of the first token prediction. However, first-tokens may not consistently reflect the final response output, due to model's diverse response styles such as starting with "Sure" or refusing to answer. Consequently, MCQ evaluation is not indicative of model behaviour when interacting with users. But by how much? We evaluate how aligned first-token evaluation is with the text output along several dimensions, namely final option choice, refusal rate, choice distribution and robustness under prompt perturbation. Our results show that the two approaches are severely misaligned on all dimensions, reaching mismatch rates over 60%. Models heavily fine-tuned on conversational or safety data are especially impacted. Crucially, models remain misaligned even when we increasingly constrain prompts, i.e., force them to start with an option letter or example template. Our findings i) underscore the importance of inspecting the text output, too and ii) caution against relying solely on first-token evaluation.
Abstract:Warning: this paper includes model outputs showing offensive content. Recent large-scale Visual-Language Generative Models (VLGMs) have achieved unprecedented improvement in multimodal image/text generation. However, these models might also generate toxic content, e.g., offensive text and pornography images, raising significant ethical risks. Despite exhaustive studies on toxic degeneration of language models, this problem remains largely unexplored within the context of visual-language generation. This work delves into the propensity for toxicity generation and susceptibility to toxic data across various VLGMs. For this purpose, we built ToViLaG, a dataset comprising 32K co-toxic/mono-toxic text-image pairs and 1K innocuous but evocative text that tends to stimulate toxicity. Furthermore, we propose WInToRe, a novel toxicity metric tailored to visual-language generation, which theoretically reflects different aspects of toxicity considering both input and output. On such a basis, we benchmarked the toxicity of a diverse spectrum of VLGMs and discovered that some models do more evil than expected while some are more vulnerable to infection, underscoring the necessity of VLGMs detoxification. Therefore, we develop an innovative bottleneck-based detoxification method. Our method could reduce toxicity while maintaining comparable generation quality, providing a promising initial solution to this line of research.
Abstract:Label aggregation such as majority voting is commonly used to resolve annotator disagreement in dataset creation. However, this may disregard minority values and opinions. Recent studies indicate that learning from individual annotations outperforms learning from aggregated labels, though they require a considerable amount of annotation. Active learning, as an annotation cost-saving strategy, has not been fully explored in the context of learning from disagreement. We show that in the active learning setting, a multi-head model performs significantly better than a single-head model in terms of uncertainty estimation. By designing and evaluating acquisition functions with annotator-specific heads on two datasets, we show that group-level entropy works generally well on both datasets. Importantly, it achieves performance in terms of both prediction and uncertainty estimation comparable to full-scale training from disagreement, while saving up to 70% of the annotation budget.
Abstract:Opinion summarization is expected to digest larger review sets and provide summaries from different perspectives. However, most existing solutions are deficient in epitomizing extensive reviews and offering opinion summaries from various angles due to the lack of designs for information selection. To this end, we propose SUBSUMM, a supervised summarization framework for large-scale multi-perspective opinion summarization. SUBSUMM consists of a review sampling strategy set and a two-stage training scheme. The sampling strategies take sentiment orientation and contrastive information value into consideration, with which the review subsets from different perspectives and quality levels can be selected. Subsequently, the summarizer is encouraged to learn from the sub-optimal and optimal subsets successively in order to capitalize on the massive input. Experimental results on AmaSum and Rotten Tomatoes datasets demonstrate that SUBSUMM is adept at generating pros, cons, and verdict summaries from hundreds of input reviews. Furthermore, our in-depth analysis verifies that the advanced selection of review subsets and the two-stage training scheme are vital to boosting the summarization performance.
Abstract:Recently, various intermediate layer distillation (ILD) objectives have been shown to improve compression of BERT models via Knowledge Distillation (KD). However, a comprehensive evaluation of the objectives in both task-specific and task-agnostic settings is lacking. To the best of our knowledge, this is the first work comprehensively evaluating distillation objectives in both settings. We show that attention transfer gives the best performance overall. We also study the impact of layer choice when initializing the student from the teacher layers, finding a significant impact on the performance in task-specific distillation. For vanilla KD and hidden states transfer, initialisation with lower layers of the teacher gives a considerable improvement over higher layers, especially on the task of QNLI (up to an absolute percentage change of 17.8 in accuracy). Attention transfer behaves consistently under different initialisation settings. We release our code as an efficient transformer-based model distillation framework for further studies.