Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning large language models with human preferences. While recent research has focused on algorithmic improvements, the importance of prompt-data construction has been overlooked. This paper addresses this gap by exploring data-driven bottlenecks in RLHF performance scaling, particularly reward hacking and decreasing response diversity. We introduce a hybrid reward system combining reasoning task verifiers (RTV) and a generative reward model (GenRM) to mitigate reward hacking. We also propose a novel prompt-selection method, Pre-PPO, to maintain response diversity and enhance learning effectiveness. Additionally, we find that prioritizing mathematical and coding tasks early in RLHF training significantly improves performance. Experiments across two model sizes validate our methods' effectiveness and scalability. Results show that RTV is most resistant to reward hacking, followed by GenRM with ground truth, and then GenRM with SFT Best-of-N responses. Our strategies enable rapid capture of subtle task-specific distinctions, leading to substantial improvements in overall RLHF performance. This work highlights the importance of careful data construction and provides practical methods to overcome performance barriers in RLHF.

Current approaches for training Process Reward Models (PRMs) often involve breaking down responses into multiple reasoning steps using rule-based techniques, such as using predefined placeholder tokens or setting the reasoning step's length into a fixed size. These approaches overlook the fact that specific words do not typically mark true decision points in a text. To address this, we propose AdaptiveStep, a method that divides reasoning steps based on the model's confidence in predicting the next word. This division method provides more decision-making information at each step, enhancing downstream tasks, such as reward model learning. Moreover, our method does not require manual annotation. We demonstrate its effectiveness through experiments with AdaptiveStep-trained PRMs in mathematical reasoning and code generation tasks. Experimental results indicate that the outcome PRM achieves state-of-the-art Best-of-N performance, surpassing greedy search strategy with token-level value-guided decoding, while also reducing construction costs by over 30% compared to existing open-source PRMs. In addition, we provide a thorough analysis and case study on the PRM's performance, transferability, and generalization capabilities.

This paper presents a pioneering exploration of the mechanisms underlying large foundation models' (LFMs) weights, aiming to simplify AI research. Through extensive observation and analysis on prevailing LFMs, we find that regardless of initialization strategies, their weights predominantly follow a Gaussian distribution, with occasional sharp, inverted T-shaped, or linear patterns. We further discover that the weights share the i.i.d. properties of Gaussian noise, and explore their direct relationship. We find that transformation weights can be derived from Gaussian noise, and they primarily serve to increase the standard deviation of pre-trained weights, with their standard deviation growing with layer depth. In other words, transformation weights broaden the acceptable deviation from the optimal weights, facilitating adaptation to downstream tasks. Building upon the above conclusions, we thoroughly discussed the nature of optimal weights, ultimately concluding that they should exhibit zero-mean, symmetry, and sparsity, with the sparse values being a truncated Gaussian distribution and a few outliers. Our experiments in LFM adaptation and editing demonstrate the effectiveness of these insights. We hope these findings can provide a foundational understanding to pave the way for future advancements in the LFM community.

Logical reasoning is a crucial task for Large Language Models (LLMs), enabling them to tackle complex problems. Among reasoning tasks, multi-step reasoning poses a particular challenge. Grounded in the theory of formal logic, we have developed an automated method, Multi-step Deduction (MuseD), for deductive reasoning data. MuseD has allowed us to create training and testing datasets for multi-step reasoning. Our generation method enables control over the complexity of the generated instructions, facilitating training and evaluation of models across different difficulty levels. Through RLHF training, our training data has demonstrated significant improvements in logical capabilities for both in-domain of out-of-domain reasoning tasks. Additionally, we have conducted tests to assess the multi-step reasoning abilities of various models.

While transformers have demonstrated impressive capacities for in-context learning (ICL) in practice, theoretical understanding of the underlying mechanism enabling transformers to perform ICL is still in its infant stage. This work aims to theoretically study the training dynamics of transformers for in-context classification tasks. We demonstrate that, for in-context classification of Gaussian mixtures under certain assumptions, a single-layer transformer trained via gradient descent converges to a globally optimal model at a linear rate. We further quantify the impact of the training and testing prompt lengths on the ICL inference error of the trained transformer. We show that when the lengths of training and testing prompts are sufficiently large, the prediction of the trained transformer approaches the Bayes-optimal classifier. Experimental results corroborate the theoretical findings.

Reward models (RMs) guide the alignment of large language models (LLMs), steering them toward behaviors preferred by humans. Evaluating RMs is the key to better aligning LLMs. However, the current evaluation of RMs may not directly correspond to their alignment performance due to the limited distribution of evaluation data and evaluation methods that are not closely related to alignment objectives. To address these limitations, we propose RMB, a comprehensive RM benchmark that covers over 49 real-world scenarios and includes both pairwise and Best-of-N (BoN) evaluations to better reflect the effectiveness of RMs in guiding alignment optimization. We demonstrate a positive correlation between our benchmark and the downstream alignment task performance. Based on our benchmark, we conduct extensive analysis on the state-of-the-art RMs, revealing their generalization defects that were not discovered by previous benchmarks, and highlighting the potential of generative RMs. Furthermore, we delve into open questions in reward models, specifically examining the effectiveness of majority voting for the evaluation of reward models and analyzing the impact factors of generative RMs, including the influence of evaluation criteria and instructing methods. Our evaluation code and datasets are available at https://github.com/Zhou-Zoey/RMB-Reward-Model-Benchmark.

Learning from preference feedback is a common practice for aligning large language models~(LLMs) with human value. Conventionally, preference data is learned and encoded into a scalar reward model that connects a value head with an LLM to produce a scalar score as preference or reward. However, scalar models lack interpretability and are known to be susceptible to biases in datasets. This paper investigates leveraging the generation capability of LLMs to address both limitations in one shot. Specifically, we prompt the pre-trained LLM to generate positive and negative judgments, both supported with rationales in natural language form. The self-generated contrastive judgment pairs are used to train the generative judge with Direct Preference Optimization (DPO). This proposal of training the generative Judge using self-generated Contrastive judgments (Con-J) ensures natural interpretability due to the generated rationales together with the judgments, as well as high robustness against bias without the need for an additional reward head. Experimental results show that the performance of Con-J is comparable to the scalar reward model trained on the same collection of preference data, and demonstrate its superior interpretability and robustness in encoding human preferences.

Reward models (RM) play a critical role in aligning generations of large language models (LLM) to human expectations. However, prevailing RMs fail to capture the stochasticity within human preferences and cannot effectively evaluate the reliability of reward predictions. To address these issues, we propose Uncertain-aware RM (URM) and Uncertain-aware RM Ensemble (URME) to incorporate and manage uncertainty in reward modeling. URM can model the distribution of disentangled attributes within human preferences, while URME quantifies uncertainty through discrepancies in the ensemble, thereby identifying potential lack of knowledge during reward evaluation. Experiment results indicate that the proposed URM achieves state-of-the-art performance compared to models with the same size, demonstrating the effectiveness of modeling uncertainty within human preferences. Furthermore, empirical results show that through uncertainty quantification, URM and URME can identify unreliable predictions to improve the quality of reward evaluations.

We introduce Adversarial Policy Optimization (AdvPO), a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback (RLHF) for Large Language Models (LLMs). Over-optimization occurs when a reward model serves as an imperfect proxy for human preference, and RL-driven policy optimization erroneously exploits reward inaccuracies. In this paper, we begin by introducing a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model, without the need for computationally expensive reward ensembles. AdvPO then addresses a distributionally robust optimization problem centred around the confidence interval of the reward model's predictions for policy improvement. Through comprehensive experiments on the Anthropic HH and TL;DR summarization datasets, we illustrate the efficacy of AdvPO in mitigating the overoptimization issue, consequently resulting in enhanced performance as evaluated through human-assisted evaluation.

The rapid expansion of large foundation models within the pre-training and fine-tuning framework has underscored that larger models often yield better results. However, the scaling up of large foundation models has led to soaring costs in fine-tuning and parameter storage, rendering extensive adaptations impractical. This challenge has sparked the development of parameter-efficient fine-tuning (PEFT), which focuses on optimizing a select subset of parameters while keeping the rest fixed, significantly lowering computational and storage overheads. While recent years have witnessed a significant success in PEFT, a deep understanding of the fundamental principles behind these methods remains unexplored. To this end, here we take the first step to unify all approaches by dissecting them from a decomposition perspective. We initiate a comprehensive mathematical analysis of these methods, allowing us to delve deeply into their underlying mechanisms, and we explore the reasons behind the variations in performance among different techniques. Furthermore, inspired by our theoretical analysis, we introduce two novel PEFT methods alongside a simple yet effective framework designed to enhance the performance of PEFT techniques across various applications. Our empirical validations, conducted across multiple datasets, demonstrate the efficacy of these methods, showcasing both theoretical validity and practical performance improvements under the guidance of our analytical findings. We believe our work will deepen researchers' understanding of PEFT and other techniques, prompting further contemplation and advancing the research across the whole community.