Lifeng Liu

Aiyuan Yang, Bin Xiao, Bingning Wang et al. · pku

Large language models (LLMs) have demonstrated remarkable performance on a variety of natural language tasks based on just a few examples of natural language instructions, reducing the need for extensive feature engineering. However, most powerful LLMs are closed-source or limited in their capability for languages other than English. In this technical report, we present Baichuan 2, a series of large-scale multilingual language models containing 7 billion and 13 billion parameters, trained from scratch, on 2.6 trillion tokens. Baichuan 2 matches or outperforms other open-source models of similar size on public benchmarks like MMLU, CMMLU, GSM8K, and HumanEval. Furthermore, Baichuan 2 excels in vertical domains such as medicine and law. We will release all pre-training model checkpoints to benefit the research community in better understanding the training dynamics of Baichuan 2.

Sahisnu Mazumder, Bing Liu, Shuai Wang et al.

This paper proposes a new method to drastically speed up deep reinforcement learning (deep RL) training for problems that have the property of state-action permissibility (SAP). Two types of permissibility are defined under SAP. The first type says that after an action $a_t$ is performed in a state $s_t$ and the agent has reached the new state $s_{t+1}$, the agent can decide whether $a_t$ is permissible or not permissible in $s_t$. The second type says that even without performing $a_t$ in $s_t$, the agent can already decide whether $a_t$ is permissible or not in $s_t$. An action is not permissible in a state if the action can never lead to an optimal solution and thus should not be tried (over and over again). We incorporate the proposed SAP property and encode action permissibility knowledge into two state-of-the-art deep RL algorithms to guide their state-action exploration together with a virtual stopping strategy. Results show that the SAP-based guidance can markedly speed up RL training.

Xinbei Ma, Tianjie Ju, Jiyang Qiu et al.

Large language models (LLMs) have played a pivotal role in building communicative AI, yet they encounter the challenge of efficient updates. Model editing enables the manipulation of specific knowledge memories and the behavior of language generation without retraining. However, the robustness of model editing remains an open question. This work seeks to understand the strengths and limitations of editing methods, facilitating practical applications of communicative AI. We focus on three key research questions. RQ1: Can edited LLMs behave consistently resembling communicative AI in realistic situations? RQ2: To what extent does the rephrasing of prompts lead LLMs to deviate from the edited knowledge memory? RQ3: Which knowledge features are correlated with the performance and robustness of editing? Our empirical studies uncover a substantial disparity between existing editing methods and the practical application of LLMs. On rephrased prompts that are flexible but common in realistic applications, the performance of editing experiences a significant decline. Further analysis shows that more popular knowledge is memorized better, easier to recall, and more challenging to edit effectively. Code is publicly available at https://github.com/xbmxb/edit_analysis .

Lifeng Liu, Fengda Zhang, Jun Xiao et al.

Federated learning is proposed as a machine learning setting to enable distributed edge devices, such as mobile phones, to collaboratively learn a shared prediction model while keeping all the training data on device, which can not only take full advantage of data distributed across millions of nodes to train a good model but also protect data privacy. However, learning in scenario above poses new challenges. In fact, data across a massive number of unreliable devices is likely to be non-IID (identically and independently distributed), which may make the performance of models trained by federated learning unstable. In this paper, we introduce a framework designed for large-scale federated learning which consists of approaches to generating dataset and modular evaluation framework. Firstly, we construct a suite of open-source non-IID datasets by providing three respects including covariate shift, prior probability shift, and concept shift, which are grounded in real-world assumptions. In addition, we design several rigorous evaluation metrics including the number of network nodes, the size of datasets, the number of communication rounds and communication resources etc. Finally, we present an open-source benchmark for large-scale federated learning research.

Ang Lv, Yuhan Chen, Kaiyi Zhang et al.

In this paper, we delve into several mechanisms employed by Transformer-based language models (LLMs) for factual recall tasks. We outline a pipeline consisting of three major steps: (1) Given a prompt ``The capital of France is,'' task-specific attention heads extract the topic token, such as ``France,'' from the context and pass it to subsequent MLPs. (2) As attention heads' outputs are aggregated with equal weight and added to the residual stream, the subsequent MLP acts as an ``activation,'' which either erases or amplifies the information originating from individual heads. As a result, the topic token ``France'' stands out in the residual stream. (3) A deep MLP takes ``France'' and generates a component that redirects the residual stream towards the direction of the correct answer, i.e., ``Paris.'' This procedure is akin to applying an implicit function such as ``get\_capital($X$),'' and the argument $X$ is the topic token information passed by attention heads. To achieve the above quantitative and qualitative analysis for MLPs, we proposed a novel analytic method aimed at decomposing the outputs of the MLP into components understandable by humans. Additionally, we observed a universal anti-overconfidence mechanism in the final layer of models, which suppresses correct predictions. We mitigate this suppression by leveraging our interpretation to improve factual recall confidence. The above interpretations are evaluated across diverse tasks spanning various domains of factual knowledge, using various language models from the GPT-2 families, 1.3B OPT, up to 7B Llama-2, and in both zero- and few-shot setups.

Xingzhou Lou, Junge Zhang, Jian Xie et al.

Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.