Chao Yu

Shusheng Xu, Wei Fu, Jiaxuan Gao et al. · tsinghua

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions. Our code is publicly available at https://github.com/openpsi-project/ReaLHF.

Zihan Zhou, Chong Li, Xinyi Chen et al.

Enlarging the context window of large language models (LLMs) has become a crucial research area, particularly for applications involving extremely long texts. In this work, we propose a novel training-free framework for processing long texts, utilizing a divide-and-conquer strategy to achieve comprehensive document understanding. The proposed LLM$\times$MapReduce framework splits the entire document into several chunks for LLMs to read and then aggregates the intermediate answers to produce the final output. The main challenge for divide-and-conquer long text processing frameworks lies in the risk of losing essential long-range information when splitting the document, which can lead the model to produce incomplete or incorrect answers based on the segmented texts. Disrupted long-range information can be classified into two categories: inter-chunk dependency and inter-chunk conflict. We design a structured information protocol to better cope with inter-chunk dependency and an in-context confidence calibration mechanism to resolve inter-chunk conflicts. Experimental results demonstrate that LLM$\times$MapReduce can outperform representative open-source and commercial long-context LLMs, and is applicable to several different models.

Yue Mao, Yi Shen, Chao Yu et al.

Aspect based sentiment analysis (ABSA) involves three fundamental subtasks: aspect term extraction, opinion term extraction, and aspect-level sentiment classification. Early works only focused on solving one of these subtasks individually. Some recent work focused on solving a combination of two subtasks, e.g., extracting aspect terms along with sentiment polarities or extracting the aspect and opinion terms pair-wisely. More recently, the triple extraction task has been proposed, i.e., extracting the (aspect term, opinion term, sentiment polarity) triples from a sentence. However, previous approaches fail to solve all subtasks in a unified end-to-end framework. In this paper, we propose a complete solution for ABSA. We construct two machine reading comprehension (MRC) problems and solve all subtasks by joint training two BERT-MRC models with parameters sharing. We conduct experiments on these subtasks, and results on several benchmark datasets demonstrate the effectiveness of our proposed framework, which significantly outperforms existing state-of-the-art methods.

Chao Yu, Zhiguo Su

In this paper, a novel neural network activation function, called Symmetrical Gaussian Error Linear Unit (SGELU), is proposed to obtain high performance. It is achieved by effectively integrating the property of the stochastic regularizer in the Gaussian Error Linear Unit (GELU) with the symmetrical characteristics. Combining with these two merits, the proposed unit introduces the capability of the bidirection convergence to successfully optimize the network without the gradient diminishing problem. The evaluations of SGELU against GELU and Linearly Scaled Hyperbolic Tangent (LiSHT) have been carried out on MNIST classification and MNIST auto-encoder, which provide great validations in terms of the performance, the convergence rate among these applications.