As a relative quality comparison of model responses, human and Large Language Model (LLM) preferences serve as common alignment goals in model fine-tuning and criteria in evaluation. Yet, these preferences merely reflect broad tendencies, resulting in less explainable and controllable models with potential safety risks. In this work, we dissect the preferences of human and 32 different LLMs to understand their quantitative composition, using annotations from real-world user-model conversations for a fine-grained, scenario-wise analysis. We find that humans are less sensitive to errors, favor responses that support their stances, and show clear dislike when models admit their limits. On the contrary, advanced LLMs like GPT-4-Turbo emphasize correctness, clarity, and harmlessness more. Additionally, LLMs of similar sizes tend to exhibit similar preferences, regardless of their training methods, and fine-tuning for alignment does not significantly alter the preferences of pretrained-only LLMs. Finally, we show that preference-based evaluation can be intentionally manipulated. In both training-free and training-based settings, aligning a model with the preferences of judges boosts scores, while injecting the least preferred properties lowers them. This results in notable score shifts: up to 0.59 on MT-Bench (1-10 scale) and 31.94 on AlpacaEval 2.0 (0-100 scale), highlighting the significant impact of this strategic adaptation. Interactive Demo: https://huggingface.co/spaces/GAIR/Preference-Dissection-Visualization Dataset: https://huggingface.co/datasets/GAIR/preference-dissection Code: https://github.com/GAIR-NLP/Preference-Dissection
Critique, as a natural language description for assessing the quality of model-generated content, has been proven to play an essential role in the training, evaluation, and refinement of Large Language Models (LLMs). However, there is a lack of principled understanding in evaluating the quality of the critique itself. In this paper, we pioneer the critique of critique, termed MetaCritique, which is a framework to evaluate the critique from two aspects, i.e., factuality as precision score and comprehensiveness as recall score. We calculate the harmonic mean of precision and recall as the overall rating called F1 score. To obtain a reliable evaluation outcome, we propose Atomic Information Units (AIUs), which describe the critique in a more fine-grained manner. MetaCritique takes each AIU into account and aggregates each AIU's judgment for the overall score. Moreover, given the evaluation process involves intricate reasoning, our MetaCritique provides a natural language rationale to support each judgment. We construct a meta-evaluation dataset containing 300 critiques (2653 AIUs) across four tasks (question answering, reasoning, entailment, and summarization), and we conduct a comparative study to demonstrate the feasibility and effectiveness. Experiments also show superior critique judged by MetaCritique leads to better refinement, indicating generative artificial intelligence indeed has the potential to be significantly advanced with our MetaCritique. We will release relevant code and meta-evaluation datasets at https://github.com/GAIR-NLP/MetaCritique.
With the rapid development of large language models, AI assistants like ChatGPT have widely entered people's works and lives. In this paper, we present an evolving large language model assistant that utilizes verbal long-term memory. It focuses on preserving the knowledge and experience from the history dialogue between the user and AI assistant, which can be applied to future dialogue for generating a better response. The model generates a set of records for each finished dialogue and stores them in the memory. In later usage, given a new user input, the model uses it to retrieve its related memory to improve the quality of the response. To find the best form of memory, we explore different ways of constructing the memory and propose a new memorizing mechanism called conditional memory to solve the problems in previous methods. We also investigate the retrieval and usage of memory in the generation process. The assistant uses GPT-4 as the backbone and we evaluate it on three constructed test datasets focusing on different abilities required by an AI assistant with long-term memory.
In the quest to advance human-centric natural language generation (NLG) systems, ensuring alignment between NLG models and human preferences is crucial. For this alignment, current popular methods leverage a reinforcement learning (RL) approach with a reward model trained on feedback from humans. However, inherent disagreements due to the subjective nature of human preferences pose a significant challenge for training the reward model, resulting in a deterioration of the NLG performance. To tackle this issue, previous approaches typically rely on majority voting or averaging to consolidate multiple inconsistent preferences into a merged one. Although straightforward to understand and execute, such methods suffer from an inability to capture the nuanced degrees of disaggregation among humans and may only represent a specialized subset of individuals, thereby lacking the ability to quantitatively disclose the universality of human preferences. To address this challenge, this paper proposes a novel approach, which employs a Bayesian framework to account for the distribution of disagreements among human preferences as training a preference model, and names it as d-PM. Besides, considering the RL strategy's inefficient and complex training process over the training efficiency, we further propose utilizing the contrastive learning strategy to train the NLG model with the preference scores derived from the d-PM model. Extensive experiments on two human-centric NLG tasks, i.e., emotional support conversation and integrity "Rule-of-Thumb" generation, show that our method consistently exceeds previous SOTA models in both automatic and human evaluations.
The rapid development of Large Language Models (LLMs) has substantially expanded the range of tasks they can address. In the field of Natural Language Processing (NLP), researchers have shifted their focus from conventional NLP tasks (e.g., sequence tagging and parsing) towards tasks that revolve around aligning with human needs (e.g., brainstorming and email writing). This shift in task distribution imposes new requirements on evaluating these aligned models regarding generality (i.e., assessing performance across diverse scenarios), flexibility (i.e., examining under different protocols), and interpretability (i.e., scrutinizing models with explanations). In this paper, we propose a generative judge with 13B parameters, Auto-J, designed to address these challenges. Our model is trained on user queries and LLM-generated responses under massive real-world scenarios and accommodates diverse evaluation protocols (e.g., pairwise response comparison and single-response evaluation) with well-structured natural language critiques. To demonstrate the efficacy of our approach, we construct a new testbed covering 58 different scenarios. Experimentally, Auto-J outperforms a series of strong competitors, including both open-source and closed-source models, by a large margin. We also provide detailed analysis and case studies to further reveal the potential of our method and make a variety of resources public at https://github.com/GAIR-NLP/auto-j.
Unsupervised extractive summarization aims to extract salient sentences from a document as the summary without labeled data. Recent literatures mostly research how to leverage sentence similarity to rank sentences in the order of salience. However, sentence similarity estimation using pre-trained language models mostly takes little account of document-level information and has a weak correlation with sentence salience ranking. In this paper, we proposed two novel strategies to improve sentence similarity estimation for unsupervised extractive summarization. We use contrastive learning to optimize a document-level objective that sentences from the same document are more similar than those from different documents. Moreover, we use mutual learning to enhance the relationship between sentence similarity estimation and sentence salience ranking, where an extra signal amplifier is used to refine the pivotal information. Experimental results demonstrate the effectiveness of our strategies.
Encoder-decoder models have achieved remarkable success in abstractive text summarization, which aims to compress one or more documents into a shorter version without the loss of the essential content. Unfortunately, these models mostly suffer a discrepancy between training and inference, i.e., the exposure bias problem. During the training stage, with teacher forcing these models are optimized to maximize the likelihood of the gold summary given the gold summary tokens as input to the decoder, while at inference the given tokens are replaced by the generated tokens. Consequently, low-quality summaries are very likely to be generated. To remedy this problem, we propose to leverage contrastive learning to decrease the likelihood of these low-quality summaries, and meanwhile increase the likelihood of the gold summary. Since our solution expands the states that the model perceives during training, we expect that the exposure bias problem can be alleviated. We experimentally demonstrate that our method effectively improves the performance of the state-of-the-art model on different datasets.