Recent advancements in large language models (LLMs) have shifted the post-training paradigm from traditional instruction tuning and human preference alignment toward reinforcement learning (RL) focused on reasoning capabilities. However, numerous technical reports indicate that purely rule-based reward RL frequently results in poor-quality reasoning chains or inconsistencies between reasoning processes and final answers, particularly when the base model is of smaller scale. During the RL exploration process, models might employ low-quality reasoning chains due to the lack of knowledge, occasionally producing correct answers randomly and receiving rewards based on established rule-based judges. This constrains the potential for resource-limited organizations to conduct direct reinforcement learning training on smaller-scale models. We propose a novel confidence-based reward model tailored for enhancing STEM reasoning capabilities. Unlike conventional approaches, our model penalizes not only incorrect answers but also low-confidence correct responses, thereby promoting more robust and logically consistent reasoning. We validate the effectiveness of our approach through static evaluations, Best-of-N inference tests, and PPO-based RL training. Our method outperforms several state-of-the-art open-source reward models across diverse STEM benchmarks. We release our codes and model in https://github.com/qianxiHe147/C2RM.

User information needs are often highly diverse and varied. A key challenge in current research is how to achieve controllable multi-objective generation while enabling rapid adaptation to accommodate diverse user demands during test time. Existing solutions, such as Rewarded Soup, focus on merging language models individually tuned on single objectives. While easy to implement and widely used, these approaches face limitations in achieving optimal performance due to their disregard for the impacts of competing objectives on model tuning. To address this issue, we propose Bone Soup, a novel model merging approach that first seeks a series of backbone models by considering the impacts of multiple objectives and then makes the soup (i.e., merge the backbone models). Specifically, Bone Soup begins by training multiple backbone models for different objectives using multi-objective reinforcement learning. Each backbone model is guided by a combination of backbone reward signals. To ensure that these models are optimal for the Pareto front, the backbone rewards are crafted by combining standard reward functions into basis vectors, which can then be modified through a rule-based construction method. Bone Soup leverages a symmetric circulant matrix mapping to generate the merging coefficients, which are used to merge the backbone models according to user preferences. Extensive experimental results demonstrate that Bone Soup exhibits strong controllability and Pareto optimality in controllable multi-objective generation, providing a more effective and efficient approach to addressing diverse user needs at test time.

To align language models (LMs, e.g., OpenAI, 2023; Reid et al., 2024) with human values, reinforcement learning (RL, Sutton and Barto, 2018) methods have been widely adopted to optimize the non-differentiable human preference, leading to the paradigm of reinforcement learning from human feedback (RLHF, Ouyang et al., 2022; Bai et al., 2022b). A prevailing approach in RLHF is to optimize the LMs by proximal policy optimization (PPO, Schulman et al., 2017) against a bandit reward model learned from human preference data, with KL regularization towards a pre-specified target distribution to avoid over-optimization on the reward model (Ziegler et al., 2019; Stiennon et al., 2020; Castricato et al., 2022). While this bandit approach is easier for reward modeling and has achieved remarkable success, language generation is intrinsically sequential, rather than simultaneous. Thus, from the view of optimizing human preference, assigning a bandit reward to entire text sequence induces the sparse reward (delayed feedback) issue (Andrychowicz et al., 2017; Marbach and Tsitsiklis, 2003), that often hurts RL-based LM training by increasing gradient variance and lowering sample efficiency (Takanobu et al., 2019; Wang et al., 2020; Guo et al., 2022; Snell et al., 2022).

Learning from preference feedback has emerged as an essential step for improving the generation quality and performance of modern language models (LMs). Despite its widespread use, the way preference-based learning is applied varies wildly, with differing data, learning algorithms, and evaluations used, making disentangling the impact of each aspect difficult. In this work, we identify four core aspects of preference-based learning: preference data, learning algorithm, reward model, and policy training prompts, systematically investigate the impact of these components on downstream model performance, and suggest a recipe for strong learning for preference feedback. Our findings indicate that all aspects are important for performance, with better preference data leading to the largest improvements, followed by the choice of learning algorithm, the use of improved reward models, and finally the use of additional unlabeled prompts for policy training. Notably, PPO outperforms DPO by up to 2.5% in math and 1.2% in general domains. High-quality preference data leads to improvements of up to 8% in instruction following and truthfulness. Despite significant gains of up to 5% in mathematical evaluation when scaling up reward models, we surprisingly observe marginal improvements in other categories.

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions.

Learning from human preference has been considered key to aligning Large Language Models (LLMs) with human values. However, contrary to popular belief, our preliminary study reveals that reward models trained on human preference datasets tend to give higher scores to long off-topic responses than short on-topic ones. Motivated by this observation, we explore a preference-free approach utilizing `relevance' as a key objective for alignment. On our first attempt, we find that the relevance score obtained by a retriever alone is vulnerable to reward hacking, i.e., overoptimizing to undesired shortcuts, when we utilize the score as a reward for reinforcement learning. To mitigate it, we integrate effective inductive biases into the vanilla relevance to regularize each other, resulting in a mixture of reward functions: Regularized Relevance Reward ($R^3$). $R^3$ significantly improves performance on preference benchmarks by providing a robust reward signal. Notably, $R^3$ does not require any human preference datasets (i.e., preference-free), outperforming open-source reward models in improving human preference. Our analysis demonstrates that $R^3$ has advantages in elevating human preference while minimizing its side effects. Finally, we show the generalizability of $R^3$, consistently improving instruction-tuned models in various backbones and sizes without additional dataset cost. Our code is available at https://github.com/naver-ai/RRR.

Large Language Models (LLMs) have revolutionized natural language processing, yet aligning these models with human values and preferences using RLHF remains a significant challenge. This challenge is characterized by various instabilities, such as reward hacking and catastrophic forgetting. In this technical report, we propose two innovations to stabilize RLHF training: (i) Advantage Model, which directly models advantage score i.e., extra reward compared to the expected rewards and regulates score distributions across tasks to prevent reward hacking. Large language models (LLMs) have become a fundamental element in advancing natural language processing (NLP) and artificial intelligence (AI), showcasing an impressive ability to generate text that is both semantically and contextually relevant (OpenAI, 2023; Köpf et al., 2023; Touvron et al., 2023). Despite these advancements, LLMs have the risk of engaging in undesirable behaviors, such as fabricating information or producing biased, toxic, or even dangerous content, since LLMs are trained on a wide array of data, which can include low-quality sources. This has highlighted the necessities of LLM Alignments with human values, intentions, and preferences (Brown et al., 2020; Ouyang et al., 2022; Bai et al., 2022a; Glaese et al., 2022). Many approaches have been put forth to address the challenge LLM Alignments (Bai et al., 2022a; OpenAI, 2023; Askell et al., 2021). Among these approaches, Reinforcement Learning from Human Feedback (RLHF) has demonstrated its efficacy in aligning language models with human preferences.

During the last stage of RLHF, a large language model is aligned to human intents via PPO training, a process that generally requires large-scale computational resources. In this technical report, we empirically investigate an efficient implementation of RLHF using low-rank adaptation (LoRA), which allows us to align the LLaMA 7B checkpoint on the Alpaca dataset (Taori et al., 2023) using only two A100 GPUs instead of the eight required for full model fine-tuning. Despite tuning only 0.2% of LLaMA 7B's parameters, our implementation achieves better performance than the publicly-released AlpacaFarm checkpoint (Dubois et al., 2023) with full model fine-tuning. Next, we analyze several configurations of our LoRA-based PPO implementation, varying the form of the KL regularization term in the training objective. We find that (1) removing this penalty term does not harm performance on the AlpacaFarm evaluation set under our LoRA setup; (2) other regularizers, such as Jensen-Shannon divergence, lead to improved performance; and (3) while PPO training negatively impacts the factuality of model-generated responses, training with LoRA largely mitigates this effect. We release our code and pretrained checkpoints to facilitate future research on more efficient RLHF.