Hui Wang

Yichong Huang, Xiaocheng Feng, Baohang Li et al.

Large language models (LLMs) exhibit complementary strengths in various tasks, motivating the research of LLM ensembling. However, existing work focuses on training an extra reward model or fusion model to select or combine all candidate answers, posing a great challenge to the generalization on unseen data distributions. Besides, prior methods use textual responses as communication media, ignoring the valuable information in the internal representations. In this work, we propose a training-free ensemble framework DeePEn, fusing the informative probability distributions yielded by different LLMs at each decoding step. Unfortunately, the vocabulary discrepancy between heterogeneous LLMs directly makes averaging the distributions unfeasible due to the token misalignment. To address this challenge, DeePEn maps the probability distribution of each model from its own probability space to a universal relative space based on the relative representation theory, and performs aggregation. Next, we devise a search-based inverse transformation to transform the aggregated result back to the probability space of one of the ensembling LLMs (main model), in order to determine the next token. We conduct extensive experiments on ensembles of different number of LLMs, ensembles of LLMs with different architectures, and ensembles between the LLM and the specialist model. Experimental results show that (i) DeePEn achieves consistent improvements across six benchmarks covering subject examination, reasoning, and knowledge, (ii) a well-performing specialist model can benefit from a less effective LLM through distribution fusion, and (iii) DeePEn has complementary strengths with other ensemble methods such as voting.

Han Zhang, Lin Gui, Yu Lei et al.

Reinforcement Learning from Human Feedback (RLHF) is commonly utilized to improve the alignment of Large Language Models (LLMs) with human preferences. Given the evolving nature of human preferences, continual alignment becomes more crucial and practical in comparison to traditional static alignment. Nevertheless, making RLHF compatible with Continual Learning (CL) is challenging due to its complex process. Meanwhile, directly learning new human preferences may lead to Catastrophic Forgetting (CF) of historical preferences, resulting in helpless or harmful outputs. To overcome these challenges, we propose the Continual Optimal Policy Regularization (COPR) method, which draws inspiration from the optimal policy theory. COPR utilizes a sampling distribution as a demonstration and regularization constraints for CL. It adopts the Lagrangian Duality (LD) method to dynamically regularize the current policy based on the historically optimal policy, which prevents CF and avoids over-emphasizing unbalanced objectives. We also provide formal proof for the learnability of COPR. The experimental results show that COPR outperforms strong CL baselines on our proposed benchmark, in terms of reward-based, GPT-4 evaluations and human assessment. Furthermore, we validate the robustness of COPR under various CL settings, including different backbones, replay memory sizes, and learning orders.

Chengpeng Fu, Xiaocheng Feng, Yichong Huang et al.

Leveraging large language models for machine translation has demonstrated promising results. However, it does require the large language models to possess the capability of handling both the source and target languages in machine translation. When it is challenging to find large models that support the desired languages, resorting to continuous learning methods becomes a costly endeavor. To mitigate these expenses, we propose an innovative approach called RD (Relay Decoding), which entails concatenating two distinct large models that individually support the source and target languages. By incorporating a simple mapping layer to facilitate the connection between these two models and utilizing a limited amount of parallel data for training, we successfully achieve superior results in the machine translation task. Experimental results conducted on the Multi30k and WikiMatrix datasets validate the effectiveness of our proposed method.