In this paper, we study format biases in reinforcement learning from human feedback (RLHF). We observe that many widely-used preference models—including human evaluators, GPT-4, and top-ranking models on the RewardBench benchmark—exhibit strong biases towards specific format patterns, such as lists, links, bold text, and emojis. Furthermore, large language models (LLMs) can exploit these biases to achieve higher rankings on popular benchmarks like AlpacaEval and LMSYS Chatbot Arena. One notable example is verbosity bias, where current preference models favor longer responses that appear more comprehensive, even when their quality is equal to or lower than shorter responses. However, format biases beyond verbosity remain largely underexplored. In this work, we extend the study of biases in preference learning beyond the commonly recognized length bias, offering a comprehensive analysis of a wider range of format biases. Additionally, we show that with a small amount of biased data (less than 1%), we can inject significant bias into the reward model. Moreover, these format biases can also be easily exploited by downstream alignment algorithms, such as *best-of-n sampling* and online iterative *DPO*, as it is usually easier to manipulate the format than to improve the quality of responses. Our findings emphasize the need to disentangle format and content both for designing alignment algorithms and evaluating models.

The advancement of Large Language Models (LLMs) has made ensuring their trustworthiness increasingly critical, especially in terms of fairness across diverse human groups. While modern LLMs are aligned with user preferences through Reinforcement Learning from Human Feedback (RLHF), the reward models used for alignment are trained on preference data that may both reflect societal biases and suffer from demographic skewness, as labeler populations are often uneven due to systemic accessibility or participation gaps. In this work, we reveal that reward models can exhibit significant discrepancies across different demographic groups, posing a fundamental challenge to fair and robust alignment. Using real-world datasets, we conduct the most comprehensive study to date, auditing various state-of-the-art reward models across nine sensitive attributes, including age, gender, ethnicity, etc. Our evaluation spans both (1) the agreement level between reward models and specific user groups, and (2) the reward model’s preference toward responses associated with different groups. Based on these findings, we propose the first method to mitigate group disparities in reward modeling. Code is available at https://github.com/Violet24K/FaRM.

Designing polymers for targeted applications and accurately predicting their properties is a key challenge in materials science owing to the vast and complex polymer chemical space. While molecular language models have proven effective in solving analogous problems for molecular discovery, similar advancements for polymers are limited. To address this gap, we propose polyBART, a language model-driven polymer discovery capability that enables rapid and accurate exploration of the polymer design space. Central to our approach is Pseudo-polymer SELFIES (PSELFIES), a novel representation that allows for the transfer of molecular language models to the polymer space. polyBART is, to the best of our knowledge, the first language model capable of bidirectional translation between polymer structures and properties, achieving state-of-the-art results in property prediction and design of novel polymers for electrostatic energy storage. Further, polyBART is validated through a combination of both computational and laboratory experiments. We report what we believe is the first successful synthesis and validation of a polymer designed by a language model, predicted to exhibit high thermal degradation temperature and confirmed by our laboratory measurements. Our work presents a generalizable strategy for adapting molecular language models to the polymer space and introduces a polymer foundation model, advancing generative polymer design that may be adapted for a variety of applications.

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA’s performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here.

Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, it is desirable for them to be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.

Foundation models have demonstrated a great ability to achieve general human-level intelligence far beyond traditional approaches. As the technique keeps attracting attention from the AI community, more and more foundation models have become publicly available.However, most of those models exhibit a major deficiency in specialized-domain and specialized-task applications, where the step of domain- and task-aware finetuning is still required to obtain scientific language models. As the number of available foundation models and specialized tasks keeps growing, the job of training scientific language models becomes highly nontrivial. In this paper, we take the first step to address this issue. We introduce an extensible and lightweight toolkit, LMFlow, which aims to simplify the domain- and task-aware finetuning of general foundation models.LMFlow offers a complete finetuning workflow for a foundation model to support specialized training with limited computing resources.Furthermore, it supports continuous pretraining, instruction tuning, parameter-efficient finetuning, alignment tuning, inference acceleration, long context generalization, model customization, and even multimodal finetuning, along with carefully designed and extensible APIs. This toolkit has been thoroughly tested and is available at https://github.com/OptimalScale/LMFlow.

This paper presents our contribution to the SemEval-2022 Task 2: Multilingual Idiomaticity Detection and Sentence Embedding.We explore the impact of three different pre-trained multilingual language models in the SubTaskA.By enhancing the model generalization and robustness, we use the exponential moving average (EMA) method and the adversarial attack strategy. In SubTaskB, we add an effective cross-attention module for modeling the relationships of two sentences. We jointly train the model with a contrastive learning objective and employ a momentum contrast to enlarge the number of negative pairs. Additionally, we use the alignment and uniformity properties to measure the quality of sentence embeddings.Our approach obtained competitive results in both subtasks.