2023-11-11 更新
Model-Based Minimum Bayes Risk Decoding
Authors:Yuu Jinnai, Tetsuro Morimura, Ukyo Honda, Kaito Ariu, Kenshi Abe
Minimum Bayes Risk (MBR) decoding has been shown to be a powerful alternative to beam search decoding in a variety of text generation tasks. MBR decoding selects a hypothesis from a pool of hypotheses that has the least expected risk under a probability model according to a given utility function. Since it is impractical to compute the expected risk exactly over all possible hypotheses, two approximations are commonly used in MBR. First, it integrates over a sampled set of hypotheses rather than over all possible hypotheses. Second, it estimates the probability of each hypothesis using a Monte Carlo estimator. While the first approximation is necessary to make it computationally feasible, the second is not essential since we typically have access to the model probability at inference time. We propose Model-Based MBR (MBMBR), a variant of MBR that uses the model probability itself as the estimate of the probability distribution instead of the Monte Carlo estimate. We show analytically and empirically that the model-based estimate is more promising than the Monte Carlo estimate in text generation tasks. Our experiments show that MBMBR outperforms MBR in several text generation tasks, both with encoder-decoder models and with large language models.
PDF
点此查看论文截图
DeeLM: Dependency-enhanced Large Language Model for Sentence Embeddings
Authors:Xianming Li, Jing Li
Recent studies have proposed using large language models (LLMs) for sentence embeddings. However, most existing LLMs are built with an autoregressive architecture that primarily captures forward dependencies while neglecting backward dependencies. Previous work has highlighted the importance of backward dependencies in improving sentence embeddings. To address this issue, in this paper, we first present quantitative evidence demonstrating the limited learning of backward dependencies in LLMs. Then, we propose a novel approach called Dependency-Enhanced Large Language Model (DeeLM) to improve sentence embeddings. Specifically, we found a turning point in LLMs, where surpassing specific LLM layers leads to a significant performance drop in the semantic textual similarity (STS) task. STS is a crucial task for evaluating sentence embeddings. We then extract the layers after the turning point to make them bidirectional, allowing for the learning of backward dependencies. Extensive experiments demonstrate that DeeLM outperforms baselines and achieves state-of-the-art performance across various STS tasks.
PDF Work in Progress
点此查看论文截图
Improving Vision-and-Language Reasoning via Spatial Relations Modeling
Authors:Cheng Yang, Rui Xu, Ye Guo, Peixiang Huang, Yiru Chen, Wenkui Ding, Zhongyuan Wang, Hong Zhou
Visual commonsense reasoning (VCR) is a challenging multi-modal task, which requires high-level cognition and commonsense reasoning ability about the real world. In recent years, large-scale pre-training approaches have been developed and promoted the state-of-the-art performance of VCR. However, the existing approaches almost employ the BERT-like objectives to learn multi-modal representations. These objectives motivated from the text-domain are insufficient for the excavation on the complex scenario of visual modality. Most importantly, the spatial distribution of the visual objects is basically neglected. To address the above issue, we propose to construct the spatial relation graph based on the given visual scenario. Further, we design two pre-training tasks named object position regression (OPR) and spatial relation classification (SRC) to learn to reconstruct the spatial relation graph respectively. Quantitative analysis suggests that the proposed method can guide the representations to maintain more spatial context and facilitate the attention on the essential visual regions for reasoning. We achieve the state-of-the-art results on VCR and two other vision-and-language reasoning tasks VQA, and NLVR.
PDF
点此查看论文截图
u-LLaVA: Unifying Multi-Modal Tasks via Large Language Model
Authors:Jinjin Xu, Liwu Xu, Yuzhe Yang, Xiang Li, Yanchun Xie, Yi-Jie Huang, Yaqian Li
Recent advances such as LLaVA and Mini-GPT4 have successfully integrated visual information into LLMs, yielding inspiring outcomes and giving rise to a new generation of multi-modal LLMs, or MLLMs. Nevertheless, these methods struggle with hallucinations and the mutual interference between tasks. To tackle these problems, we propose an efficient and accurate approach to adapt to downstream tasks by utilizing LLM as a bridge to connect multiple expert models, namely u-LLaVA. Firstly, we incorporate the modality alignment module and multi-task modules into LLM. Then, we reorganize or rebuild multi-type public datasets to enable efficient modality alignment and instruction following. Finally, task-specific information is extracted from the trained LLM and provided to different modules for solving downstream tasks. The overall framework is simple, effective, and achieves state-of-the-art performance across multiple benchmarks. We also release our model, the generated data, and the code base publicly available.
PDF
点此查看论文截图
LLaVA-Plus: Learning to Use Tools for Creating Multimodal Agents
Authors:Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang, Jianfeng Gao, Chunyuan Li
LLaVA-Plus is a general-purpose multimodal assistant that expands the capabilities of large multimodal models. It maintains a skill repository of pre-trained vision and vision-language models and can activate relevant tools based on users’ inputs to fulfill real-world tasks. LLaVA-Plus is trained on multimodal instruction-following data to acquire the ability to use tools, covering visual understanding, generation, external knowledge retrieval, and compositions. Empirical results show that LLaVA-Plus outperforms LLaVA in existing capabilities and exhibits new ones. It is distinct in that the image query is directly grounded and actively engaged throughout the entire human-AI interaction sessions, significantly improving tool use performance and enabling new scenarios.
PDF 25 pages, 25M file size. Project Page: https://llava-vl.github.io/llava-plus/
点此查看论文截图
Zero-Shot Goal-Directed Dialogue via RL on Imagined Conversations
Authors:Joey Hong, Sergey Levine, Anca Dragan
Large language models (LLMs) have emerged as powerful and general solutions to many natural language tasks. However, many of the most important applications of language generation are interactive, where an agent has to talk to a person to reach a desired outcome. For example, a teacher might try to understand their student’s current comprehension level to tailor their instruction accordingly, and a travel agent might ask questions of their customer to understand their preferences in order to recommend activities they might enjoy. LLMs trained with supervised fine-tuning or “single-step” RL, as with standard RLHF, might struggle which tasks that require such goal-directed behavior, since they are not trained to optimize for overall conversational outcomes after multiple turns of interaction. In this work, we explore a new method for adapting LLMs with RL for such goal-directed dialogue. Our key insight is that, though LLMs might not effectively solve goal-directed dialogue tasks out of the box, they can provide useful data for solving such tasks by simulating suboptimal but human-like behaviors. Given a textual description of a goal-directed dialogue task, we leverage LLMs to sample diverse synthetic rollouts of hypothetical in-domain human-human interactions. Our algorithm then utilizes this dataset with offline reinforcement learning to train an interactive conversational agent that can optimize goal-directed objectives over multiple turns. In effect, the LLM produces examples of possible interactions, and RL then processes these examples to learn to perform more optimal interactions. Empirically, we show that our proposed approach achieves state-of-the-art performance in various goal-directed dialogue tasks that include teaching and preference elicitation.
PDF 25 pages, 6 figures
点此查看论文截图
Efficient Parallelization Layouts for Large-Scale Distributed Model Training
Authors:Johannes Hagemann, Samuel Weinbach, Konstantin Dobler, Maximilian Schall, Gerard de Melo
Efficiently training large language models requires parallelizing across hundreds of hardware accelerators and invoking various compute and memory optimizations. When combined, many of these strategies have complex interactions regarding the final training efficiency. Prior work tackling this problem did not have access to the latest set of optimizations, such as FlashAttention or sequence parallelism. In this work, we conduct a comprehensive ablation study of possible training configurations for large language models. We distill this large study into several key recommendations for the most efficient training. For instance, we find that using a micro-batch size of 1 usually enables the most efficient training layouts. Larger micro-batch sizes necessitate activation checkpointing or higher degrees of model parallelism and also lead to larger pipeline bubbles. Our most efficient configurations enable us to achieve state-of-the-art training efficiency results over a range of model sizes, most notably a Model FLOPs utilization of 70.5% when training a 13B model.
PDF
点此查看论文截图