Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning Large Language Models (LLMs) with human values. However, RLHF has been continuously challenged by its high complexity in implementation and computation consumption. Even with recent simplifications, such as Direct Preference Optimization (DPO) and Advantage Leftover Lunch (A-LoL), the problems of over-fitting and training instability remain hindering the alignment process from the expected optimal performance. To address the existing challenges, we propose a novel simplification of RLHF from the perspective of variational inference, called $\textbf{V}$ariational $\textbf{A}$lignment with $\textbf{R}$e-weighting ($\textbf{VAR}$). More specifically, by directly minimizing the distribution gap between the learning LLM policy and the optimal solution of RLHF, we transform the alignment objective into a reward-driven re-weighted supervised fine-tuning (SFT) form, which only requires minor adjustment on the SFT loss to obtain noticeable improvement on training stability and effectiveness. On comprehensive alignment and generation benchmarks, our VAR method has numerically achieved competitive performance in LLM alignment helpfulness and harmlessness.

Human preference alignment is essential to improve the interaction quality of large language models (LLMs). Existing alignment methods depend on manually annotated preference data to guide the LLM optimization directions. However, continuously updating LLMs for alignment raises a distribution gap between model-generated samples and human-annotated responses, hindering training effectiveness. To mitigate this issue, previous methods require additional preference annotation on newly generated samples to adapt to the shifted distribution, which consumes a large amount of annotation resources. Targeting more efficient human preference optimization, we propose an Adversarial Preference Optimization (APO) framework, in which the LLM and the reward model update alternatively via a min-max game. Through adversarial training, the reward model can adapt to the shifted generation distribution of the LLM without any additional annotation. With comprehensive experiments, we find the proposed adversarial training framework further enhances existing alignment baselines in terms of LLM helpfulness and harmlessness. The code is at https://github.com/Linear95/APO.

We explore the self-play training procedure of large language models (LLMs) in a two-player adversarial language game called Adversarial Taboo. In this game, an attacker and a defender communicate around a target word only visible to the attacker. The attacker aims to induce the defender to speak the target word unconsciously, while the defender tries to infer the target word from the attacker's utterances. To win the game, both players should have sufficient knowledge about the target word and high-level reasoning ability to infer and express in this information-reserved conversation. Hence, we are curious about whether LLMs' reasoning ability can be further enhanced by self-play in this adversarial language game (SPAG). With this goal, we select several open-source LLMs and let each act as the attacker and play with a copy of itself as the defender on an extensive range of target words. Through reinforcement learning on the game outcomes, we observe that the LLMs' performances uniformly improve on a broad range of reasoning benchmarks. Furthermore, iteratively adopting this self-play process can continuously promote LLMs' reasoning abilities. The code is at https://github.com/Linear95/SPAG.

Aligning large language models (LLMs) with human preferences has been recognized as the key to improving LLMs' interaction quality. However, in this pluralistic world, human preferences can be diversified by people's different tastes, which hinders the effectiveness of LLM alignment methods. In this paper, we provide the first quantitative analysis to verify the existence of diversified preferences in commonly used human feedback datasets. To mitigate the alignment ineffectiveness caused by diversified preferences, we propose a novel \textbf{M}ulti-\textbf{O}bjective \textbf{Re}ward learning method (MORE), which can automatically adjust the learning gradients across different preference data sources. In experiments, we evaluate MORE with the Pythia-1.4B model on five mixed human preference datasets, on which our method achieves superior performance compared with other baselines in terms of preference accuracy and prediction calibration.

Reward models (RMs) are essential for aligning large language models (LLMs) with human preferences to improve interaction quality. However, the real world is pluralistic, which leads to diversified human preferences with respect to different religions, politics, cultures, etc. Moreover, each individual can have their unique preferences on various topics. Neglecting the diversity of human preferences, current human feedback aligning methods only consider a general reward model, which is below satisfaction for customized or personalized application scenarios. To explore customized preference learning, we collect a domain-specific preference (DSP) dataset, which includes preferred responses for each given query from four practical domains. Besides, from the perspective of data efficiency, we propose a three-stage customized RM learning scheme, then empirically verify its effectiveness on both general preference datasets and our DSP set. Furthermore, we test multiple training and data strategies on the three learning stages. We find several ways to better preserve the general preferring ability while training the customized RMs, especially general preference enrichment, and customized preference imitation learning. The DSP dataset and code are available at https://github.com/Linear95/DSP.

Although dominant in natural language processing, transformer-based models remain challenged by the task of long-sequence processing, because the computational cost of self-attention operations in transformers swells quadratically with the input sequence length. To alleviate the complexity of long-sequence processing, we propose a simple framework to enable the offthe-shelf pre-trained transformers to process much longer sequences, while the computation and memory costs remain growing linearly with the input sequence lengths. More specifically, our method divides each long-sequence input into a batch of chunks, then aligns the interchunk information during the encoding steps, and finally selects the most representative hidden states from the encoder for the decoding process. To extract inter-chunk semantic information, we align the start and end token embeddings among chunks in each encoding transformer block. To learn an effective hidden selection policy, we design a dual updating scheme inspired by reinforcement learning, which regards the decoders of transformers as environments, and the downstream performance metrics as the rewards to evaluate the hidden selection actions. Our empirical results on real-world long-text summarization and reading comprehension tasks demonstrate effective improvements compared to prior longsequence processing baselines.

Pretrained language models (PLMs), such as GPT2, have achieved remarkable empirical performance in text generation tasks. However, pretrained on large-scale natural language corpora, the generated text from PLMs may exhibit social bias against disadvantaged demographic groups. To improve the fairness of PLMs in text generation, we propose to minimize the mutual information between the semantics in the generated text sentences and their demographic polarity, i.e., the demographic group to which the sentence is referring. In this way, the mentioning of a demographic group (e.g., male or female) is encouraged to be independent from how it is described in the generated text, thus effectively alleviating the social bias. Moreover, we propose to efficiently estimate the upper bound of the above mutual information via importance sampling, leveraging a natural language corpus. We also propose a distillation mechanism that preserves the language modeling ability of the PLMs after debiasing. Empirical results on real-world benchmarks demonstrate that the proposed method yields superior performance in term of both fairness and language modeling ability.

In domains where data are sensitive or private, there is great value in methods that can learn in a distributed manner without the data ever leaving the local devices. In light of this need, federated learning has emerged as a popular training paradigm. However, many federated learning approaches trade transmitting data for communicating updated weight parameters for each local device. Therefore, a successful breach that would have otherwise directly compromised the data instead grants whitebox access to the local model, which opens the door to a number of attacks, including exposing the very data federated learning seeks to protect. Additionally, in distributed scenarios, individual client devices commonly exhibit high statistical heterogeneity. Many common federated approaches learn a single global model; while this may do well on average, performance degrades when the i.i.d. assumption is violated, underfitting individuals further from the mean, and raising questions of fairness. To address these issues, we propose Weight Anonymized Factorization for Federated Learning (WAFFLe), an approach that combines the Indian Buffet Process with a shared dictionary of weight factors for neural networks. Experiments on MNIST, FashionMNIST, and CIFAR-10 demonstrate WAFFLe's significant improvement to local test performance and fairness while simultaneously providing an extra layer of security.

Mutual information (MI) minimization has gained considerable interests in various machine learning tasks. However, estimating and minimizing MI in high-dimensional spaces remains a challenging problem, especially when only samples, rather than distribution forms, are accessible. Previous works mainly focus on MI lower bound approximation, which is not applicable to MI minimization problems. In this paper, we propose a novel Contrastive Log-ratio Upper Bound (CLUB) of mutual information. We provide a theoretical analysis of the properties of CLUB and its variational approximation. Based on this upper bound, we introduce a MI minimization training scheme and further accelerate it with a negative sampling strategy. Simulation studies on Gaussian distributions show the reliable estimation ability of CLUB. Real-world MI minimization experiments, including domain adaptation and information bottleneck, demonstrate the effectiveness of the proposed method. The code is at https://github.com/Linear95/CLUB.

Learning disentangled representations of natural language is essential for many NLP tasks, e.g., conditional text generation, style transfer, personalized dialogue systems, etc. Similar problems have been studied extensively for other forms of data, such as images and videos. However, the discrete nature of natural language makes the disentangling of textual representations more challenging (e.g., the manipulation over the data space cannot be easily achieved). Inspired by information theory, we propose a novel method that effectively manifests disentangled representations of text, without any supervision on semantics. A new mutual information upper bound is derived and leveraged to measure dependence between style and content. By minimizing this upper bound, the proposed method induces style and content embeddings into two independent low-dimensional spaces. Experiments on both conditional text generation and text-style transfer demonstrate the high quality of our disentangled representation in terms of content and style preservation.