Lu Wang

Chenghua Huang, Zhizhen Fan, Lu Wang et al.

Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human preferences, playing a pivotal role in the success of conversational models like GPT-4, ChatGPT, and Llama 2. A core challenge in employing RLHF lies in training a reliable reward model (RM), which relies on high-quality labels typically provided by human experts or advanced AI system. These methods can be costly and may introduce biases that affect the language model's responses. As language models improve, human input may become less effective in further enhancing their performance. In this paper, we propose Self-Evolved Reward Learning (SER), a novel approach where the RM generates additional training data to iteratively improve itself. We conducted extensive experiments on multiple datasets such as HH-RLHF and UltraFeedback, using models like Mistral and Llama 3, and compare SER against various baselines. Our results demonstrate that even with limited human-annotated data, learning from self-feedback can robustly enhance RM performance, thereby boosting the capabilities of large language models (LLMs). Resources of this paper can be found at https://aka.ms/ser

Yudi Zhang, Lu Wang, Meng Fang et al.

Distilling large language models (LLMs) typically involves transferring the teacher model's responses through supervised fine-tuning (SFT). However, this approach neglects the potential to distill both data (output content) and reward signals (quality evaluations). Extracting reliable reward signals directly from teacher models is challenging, as LLMs are optimized for generation rather than evaluation, often resulting in biased or inconsistent assessments. To address this limitation, we propose a novel distillation pipeline that transfers both responses and rewards. Our method generates pseudo-rewards through a self-supervised mechanism that leverages the inherent structure of both teacher and student responses, enabling reward learning without explicit external evaluation. The reward model subsequently guides reinforcement learning (RL), allowing iterative refinement of the student model after an SFT warm-up phase. Experiments on GSM8K and MMLU-PRO demonstrate that our method consistently outperforms traditional SFT-based approaches, enabling student models to surpass the performance of their teachers. This work highlights the potential for scalable, efficient distillation through structured self-supervised reward learning, reducing dependence on external reward supervision.

Shuyang Cao, Lu Wang

We study generating abstractive summaries that are faithful and factually consistent with the given articles. A novel contrastive learning formulation is presented, which leverages both reference summaries, as positive training data, and automatically generated erroneous summaries, as negative training data, to train summarization systems that are better at distinguishing between them. We further design four types of strategies for creating negative samples, to resemble errors made commonly by two state-of-the-art models, BART and PEGASUS, found in our new human annotations of summary errors. Experiments on XSum and CNN/Daily Mail show that our contrastive learning framework is robust across datasets and models. It consistently produces more factual summaries than strong comparisons with post error correction, entailment-based reranking, and unlikelihood training, according to QA-based factuality evaluation. Human judges echo the observation and find that our model summaries correct more errors.

Hou Pong Chan, Lu Wang, Irwin King

We study controllable text summarization which allows users to gain control on a particular attribute (e.g., length limit) of the generated summaries. In this work, we propose a novel training framework based on Constrained Markov Decision Process (CMDP), which conveniently includes a reward function along with a set of constraints, to facilitate better summarization control. The reward function encourages the generation to resemble the human-written reference, while the constraints are used to explicitly prevent the generated summaries from violating user-imposed requirements. Our framework can be applied to control important attributes of summarization, including length, covered entities, and abstractiveness, as we devise specific constraints for each of these aspects. Extensive experiments on popular benchmarks show that our CMDP framework helps generate informative summaries while complying with a given attribute's requirement.

Lu Wang, Munif Ishad Mujib, Jake Williams et al.

With the advent of off-the-shelf intelligent home products and broader internet adoption, researchers increasingly explore smart computing applications that provide easier access to health and wellness resources. AI-based systems like chatbots have the potential to provide services that could provide mental health support. However, existing therapy chatbots are often retrieval-based, requiring users to respond with a constrained set of answers, which may not be appropriate given that such pre-determined inquiries may not reflect each patient's unique circumstances. Generative-based approaches, such as the OpenAI GPT models, could allow for more dynamic conversations in therapy chatbot contexts than previous approaches. To investigate the generative-based model's potential in therapy chatbot contexts, we built a chatbot using the GPT-2 model. We fine-tuned it with 306 therapy session transcripts between family caregivers of individuals with dementia and therapists conducting Problem Solving Therapy. We then evaluated the model's pre-trained and the fine-tuned model in terms of basic qualities using three meta-information measurements: the proportion of non-word outputs, the length of response, and sentiment components. Results showed that: (1) the fine-tuned model created more non-word outputs than the pre-trained model; (2) the fine-tuned model generated outputs whose length was more similar to that of the therapists compared to the pre-trained model; (3) both the pre-trained model and fine-tuned model were likely to generate more negative and fewer positive outputs than the therapists. We discuss potential reasons for the problem, the implications, and solutions for developing therapy chatbots and call for investigations of the AI-based system application.

Xinyu Hua, Ashwin Sreevatsa, Lu Wang

We study the task of long-form opinion text generation, which faces at least two distinct challenges. First, existing neural generation models fall short of coherence, thus requiring efficient content planning. Second, diverse types of information are needed to guide the generator to cover both subjective and objective content. To this end, we propose DYPLOC, a generation framework that conducts dynamic planning of content while generating the output based on a novel design of mixed language models. To enrich the generation with diverse content, we further propose to use large pre-trained models to predict relevant concepts and to generate claims. We experiment with two challenging tasks on newly collected datasets: (1) argument generation with Reddit ChangeMyView, and (2) writing articles using New York Times' Opinion section. Automatic evaluation shows that our model significantly outperforms competitive comparisons. Human judges further confirm that our generations are more coherent with richer content.

Luyang Huang, Shuyang Cao, Nikolaus Parulian et al.

The quadratic computational and memory complexities of large Transformers have limited their scalability for long document summarization. In this paper, we propose Hepos, a novel efficient encoder-decoder attention with head-wise positional strides to effectively pinpoint salient information from the source. We further conduct a systematic study of existing efficient self-attentions. Combined with Hepos, we are able to process ten times more tokens than existing models that use full attentions. For evaluation, we present a new dataset, GovReport, with significantly longer documents and summaries. Results show that our models produce significantly higher ROUGE scores than competitive comparisons, including new state-of-the-art results on PubMed. Human evaluation also shows that our models generate more informative summaries with fewer unfaithful errors.

Xinyu Hua, Lu Wang

Pre-trained Transformers have enabled impressive breakthroughs in generating long and fluent text, yet their outputs are often "rambling" without coherently arranged content. In this work, we present a novel content-controlled text generation framework, PAIR, with planning and iterative refinement, which is built upon a large model, BART. We first adapt the BERT model to automatically construct the content plans, consisting of keyphrase assignments and their corresponding sentence-level positions. The BART model is employed for generation without modifying its structure. We then propose a refinement algorithm to gradually enhance the generation quality within the sequence-to-sequence framework. Evaluation with automatic metrics shows that adding planning consistently improves the generation quality on three distinct domains, with an average of 20 BLEU points and 12 METEOR points improvements. In addition, human judges rate our system outputs to be more relevant and coherent than comparisons without planning.

Xinyu Hua, Lu Wang

Building effective text generation systems requires three critical components: content selection, text planning, and surface realization, and traditionally they are tackled as separate problems. Recent all-in-one style neural generation models have made impressive progress, yet they often produce outputs that are incoherent and unfaithful to the input. To address these issues, we present an end-to-end trained two-step generation model, where a sentence-level content planner first decides on the keyphrases to cover as well as a desired language style, followed by a surface realization decoder that generates relevant and coherent text. For experiments, we consider three tasks from domains with diverse topics and varying language styles: persuasive argument construction from Reddit, paragraph generation for normal and simple versions of Wikipedia, and abstract generation for scientific articles. Automatic evaluation shows that our system can significantly outperform competitive comparisons. Human judges further rate our system generated text as more fluent and correct, compared to the generations by its variants that do not consider language style.

Xinyu Hua, Zhe Hu, Lu Wang

Automatic argument generation is an appealing but challenging task. In this paper, we study the specific problem of counter-argument generation, and present a novel framework, CANDELA. It consists of a powerful retrieval system and a novel two-step generation model, where a text planning decoder first decides on the main talking points and a proper language style for each sentence, then a content realization decoder reflects the decisions and constructs an informative paragraph-level argument. Furthermore, our generation model is empowered by a retrieval system indexed with 12 million articles collected from Wikipedia and popular English news media, which provides access to high-quality content with diversity. Automatic evaluation on a large-scale dataset collected from Reddit shows that our model yields significantly higher BLEU, ROUGE, and METEOR scores than the state-of-the-art and non-trivial comparisons. Human evaluation further indicates that our system arguments are more appropriate for refutation and richer in content.

Lu Wang, Wang Ling

We study the problem of generating abstractive summaries for opinionated text. We propose an attention-based neural network model that is able to absorb information from multiple text units to construct informative, concise, and fluent summaries. An importance-based sampling method is designed to allow the encoder to integrate information from an important subset of input. Automatic evaluation indicates that our system outperforms state-of-the-art abstractive and extractive summarization systems on two newly collected datasets of movie reviews and arguments. Our system summaries are also rated as more informative and grammatical in human evaluation.