Large Language Models (LLMs) have demonstrated remarkable capabilities in open-ended text generation tasks. However, the inherent open-ended nature of these tasks implies that there is always room for improvement in the quality of model responses. To address this challenge, various approaches have been proposed to enhance the performance of LLMs. There has been a growing focus on enabling LLMs to self-improve their response quality, thereby reducing the reliance on extensive human annotation efforts for collecting diverse and high-quality training data. Recently, prompting-based methods have been widely explored among self-improvement methods owing to their effectiveness, efficiency, and convenience. However, those methods usually require explicitly and thoroughly written rubrics as inputs to LLMs. It is expensive and challenging to manually derive and provide all necessary rubrics with a real-world complex goal for improvement (e.g., being more helpfulness and less harmful). To this end, we propose an imPlicit self-ImprovemenT (PIT) framework that implicitly learns the improvement goal from human preference data. PIT only requires preference data that are used to train reward models with no extra human efforts. Specifically, we reformulate the training objective of reinforcement learning from human feedback (RLHF) -- instead of maximizing response quality for a given input, we maximize the quality gap of the response conditioned on a reference response. In this way, PIT is implicitly trained with the improvement goal of better aligning with human preferences. Experiments on two real-world datasets and one synthetic dataset show that our method significantly outperforms prompting-based methods.

Sparse Mixture-of-Experts (MoE) is a neural architecture design that adds learnable parameters to Large Language Models (LLMs) without increasing computational complexity (FLOPs). Instruction tuning is a technique for training LLMs to follow instructions. We advocate combining these two approaches, as we find that MoE models benefit more from instruction tuning than dense models. In particular, we conduct empirical studies across three experimental setups: (i) Direct finetuning on individual downstream tasks devoid of instruction tuning; (ii) Instruction tuning followed by in-context few-shot or zero-shot generalization on downstream tasks; and (iii) Instruction tuning supplemented by further finetuning on individual downstream tasks. In the first scenario, MoE models overall underperform dense models of identical computational capacity. This narrative, however, dramatically changes with the introduction of instruction tuning (in the second and third scenarios), used independently or in conjunction with task-specific finetuning. Our most powerful model, FLAN-MoE-32B, surpasses the performance of Flan-PaLM-62B on four benchmark tasks, while using only a third of the FLOPs. The advancements embodied by FLAN-MoE inspire a reevaluation of the design principles of large-scale, high-performance language models in the framework of task-agnostic learning.

Finetuning language models on a collection of datasets phrased as instructions has been shown to improve model performance and generalization to unseen tasks. In this paper we explore instruction finetuning with a particular focus on (1) scaling the number of tasks, (2) scaling the model size, and (3) finetuning on chain-of-thought data. We find that instruction finetuning with the above aspects dramatically improves performance on a variety of model classes (PaLM, T5, U-PaLM), prompting setups (zero-shot, few-shot, CoT), and evaluation benchmarks (MMLU, BBH, TyDiQA, MGSM, open-ended generation, RealToxicityPrompts). For instance, Flan-PaLM 540B instruction-finetuned on 1.8K tasks outperforms PaLM 540B by a large margin (+9.4% on average). Flan-PaLM 540B achieves state-of-the-art performance on several benchmarks (at time of release), such as 75.2% on five-shot MMLU. We also publicly release Flan-T5 checkpoints,1 which achieve strong few-shot performance even compared to much larger models, such as PaLM 62B. Overall, instruction finetuning is a general method for improving the performance and usability of pretrained language models.

Large Language Models (LLMs) have achieved excellent performances in various tasks.However, fine-tuning an LLM requires extensive supervision.Human, on the other hand, may improve their reasoning abilities by self-thinking without external inputs.In this work, we demonstrate that an LLM is also capable of self-improving with only unlabeled datasets.We use a pre-trained LLM to generate "highconfidence" rationale-augmented answers for unlabeled questions using Chain-of-Though (CoT) prompting and self-consistency, and finetune the LLM using those self-generated solutions as target outputs.We show that without any ground truth label, our approach significantly improves the general reasoning ability of PaLM 540B model (74.4%→82.1% on GSM8K, 90.0%→94.4% on OpenBookQA, and 63.4%→67.9% on ANLI-A3) and can also be adapted to extreme low-resource cases where even training questions and CoT prompts are limited.We conduct ablation studies and show that fine-tuning on diverse reasoning paths is critical for self-improvement.

Xiaotao Gu, Liyuan Liu, Hongkun Yu, Jing Li, Chen Chen, Jiawei Han. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Natural Language Processing (NLP) has recently achieved great success by using huge pre-trained models with hundreds of millions of parameters.However, these models suffer from heavy model sizes and high latency such that they cannot be deployed to resourcelimited mobile devices.In this paper, we propose MobileBERT for compressing and accelerating the popular BERT model.Like the original BERT, MobileBERT is task-agnostic, that is, it can be generically applied to various downstream NLP tasks via simple fine-tuning.Basically, MobileBERT is a thin version of BERT LARGE , while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks.To train MobileBERT, we first train a specially designed teacher model, an invertedbottleneck incorporated BERT LARGE model.Then, we conduct knowledge transfer from this teacher to MobileBERT.Empirical studies show that MobileBERT is 4.3× smaller and 5.5× faster than BERT BASE while achieving competitive results on well-known benchmarks.On the natural language inference tasks of GLUE, MobileBERT achieves a GLUE score of 77.7 (0.6 lower than BERT BASE ), and 62 ms latency on a Pixel 4 phone.On the SQuAD v1.1/v2.0question answering task, MobileBERT achieves a dev F1 score of 90.0/79.2(1.5/2.1 higher than BERT BASE ).

Millions of news articles are published online every day, which can be overwhelming for readers to follow. Grouping articles that are reporting the same event into news stories is a common way of assisting readers in their news consumption. However, it remains a challenging research problem to efficiently and effectively generate a representative headline for each story. Automatic summarization of a document set has been studied for decades, while few studies have focused on generating representative headlines for a set of articles. Unlike summaries, which aim to capture most information with least redundancy, headlines aim to capture information jointly shared by the story articles in short length and exclude information specific to each individual article.

We present the EKNOT system that automatically discovers major events from online news articles, connects each event to its discussion in Twitter, and provides a comprehensive summary of the events from both news media and social media's point of view. EKNOT takes a time period as input and outputs a complete picture of the events within the given time range along with the public opinions. For each event, EKNOT provides multi-dimensional summaries: a) a summary from news for an objective description; b) a summary from tweets containing opinions/sentiments; c) an entity graph which illustrates the major players involved and their correlations; d) the time span of the event; and e) an opinion (sentiment) distribution. Also, if a user is interested in a particular event, he/she can zoom into this event to investigate its aspects (sub-events) summarized in the same manner. EKNOT is built on real-time crawled news articles and tweets, allowing users to explore the dynamics of major events with minimal delays.

Users often write reviews on different themes involving linguistic structures with complex sentiments. The sentiment polarity of a word can be different across themes. Moreover, contextual valence shifters may change sentiment polarity depending on the contexts that they appear in. Both challenges cannot be modeled effectively and explicitly in traditional sentiment analysis. Studying both phenomena requires multi-theme sentiment analysis at the word level, which is very interesting but significantly more challenging than overall polarity classification. To simultaneously resolve the multi-theme and sentiment shifting problems, we propose a data-driven framework to enable both capabilities: (1) polarity predictions of the same word in reviews of different themes, and (2) discovery and quantification of contextual valence shifters. The framework formulates multi-theme sentiment by factorizing the review sentiments with theme/word embeddings and then derives the shifter effect learning problem as a logistic regression. The improvement of sentiment polarity classification accuracy demonstrates not only the importance of multi-theme and sentiment shifting, but also effectiveness of our framework. Human evaluations and case studies further show the success of multi-theme word sentiment predictions and automatic effect quantification of contextual valence shifters.

Typically a user prefers an item (e.g., a movie) because she likes certain features of the item (e.g., director, genre, producer). This observation motivates us to consider a feature-centric recommendation approach to item recommendation: instead of directly predicting the rating on items, we predict the rating on the features of items, and use such ratings to derive the rating on an item. This approach offers several advantages over the traditional item-centric approach: it incorporates more information about why a user chooses an item, it generalizes better due to the denser feature rating data, it explains the prediction of item ratings through the predicted feature ratings. Another contribution is turning a principled item-centric solution into a feature-centric solution, instead of inventing a new algorithm that is feature-centric. This approach maximally leverages previous research. We demonstrate this approach by turning the traditional item-centric latent factor model into a feature-centric solution and demonstrate its superiority over item-centric approaches.

A major event often has repercussions on both news media and microblogging sites such as Twitter. Reports from mainstream news agencies and discussions from Twitter complement each other to form a complete picture. An event can have multiple aspects (sub-events) describing it from multiple angles, each of which attracts opinions/comments posted on Twitter. Mining such reflections is interesting to both policy makers and ordinary people seeking information. In this paper, we propose a unified framework to mine multi-aspect reflections of news events in Twitter. We propose a novel and efficient dynamic hierarchical entity-aware event discovery model to learn news events and their multiple aspects. The aspects of an event are linked to their reflections in Twitter by a bootstrapped dataless classification scheme, which elegantly handles the challenges of selecting informative tweets under overwhelming noise and bridging the vocabularies of news and tweets. In addition, we demonstrate that our framework naturally generates an informative presentation of each event with entity graphs, time spans, news summaries and tweet highlights to facilitate user digestion.

User preferences change over time and capturing such changes is essential for developing accurate recommender systems. Despite its importance, only a few works in collaborative filtering have addressed this issue. In this paper, we consider evolving preferences and we model user dynamics by introducing and learning a transition matrix for each user's latent vectors between consecutive time windows. Intuitively, the transition matrix for a user summarizes the time-invariant pattern of the evolution for the user. We first extend the conventional probabilistic matrix factorization and then improve upon this solution through its fully Bayesian model. These solutions take advantage of the model complexity and scalability of conventional Bayesian matrix factorization, yet adapt dynamically to user's evolving preferences. We evaluate the effectiveness of these solutions through empirical studies on six large-scale real life data sets.