Wenbin Lai

Rui Zheng, Shihan Dou, Songyang Gao et al.

Large language models (LLMs) have formulated a blueprint for the advancement of artificial general intelligence. Its primary objective is to function as a human-centric (helpful, honest, and harmless) assistant. Alignment with humans assumes paramount significance, and reinforcement learning with human feedback (RLHF) emerges as the pivotal technological paradigm underpinning this pursuit. Current technical routes usually include \textbf{reward models} to measure human preferences, \textbf{Proximal Policy Optimization} (PPO) to optimize policy model outputs, and \textbf{process supervision} to improve step-by-step reasoning capabilities. However, due to the challenges of reward design, environment interaction, and agent training, coupled with huge trial and error cost of large language models, there is a significant barrier for AI researchers to motivate the development of technical alignment and safe landing of LLMs. The stable training of RLHF has still been a puzzle. In the first report, we dissect the framework of RLHF, re-evaluate the inner workings of PPO, and explore how the parts comprising PPO algorithms impact policy agent training. We identify policy constraints being the key factor for the effective implementation of the PPO algorithm. Therefore, we explore the PPO-max, an advanced version of PPO algorithm, to efficiently improve the training stability of the policy model. Based on our main results, we perform a comprehensive analysis of RLHF abilities compared with SFT models and ChatGPT. The absence of open-source implementations has posed significant challenges to the investigation of LLMs alignment. Therefore, we are eager to release technical reports, reward models and PPO codes, aiming to make modest contributions to the advancement of LLMs.

Rui Zheng, Wei Shen, Yuan Hua et al.

The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences. As universal AI assistants, there's a growing expectation for them to perform consistently across various domains. However, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples. This focus on quick reward gains undermines both the stability in training and the model's ability to generalize to new, unseen data. In this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. Given the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance. Then, we optimize the policy to perform well on challenging groups. Lastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.

You Huang, Wenbin Lai, Jiayi Ji et al.

The Segment Anything Model (SAM) has advanced interactive segmentation but is limited by the high computational cost on high-resolution images. This requires downsampling to meet GPU constraints, sacrificing the fine-grained details needed for high-precision interactive segmentation. To address SAM's limitations, we focus on visual length extrapolation and propose a lightweight model named HRSAM. The extrapolation enables HRSAM trained on low resolutions to generalize to high resolutions. We begin by finding the link between the extrapolation and attention scores, which leads us to base HRSAM on Swin attention. We then introduce the Flexible Local Attention (FLA) framework, using CUDA-optimized Efficient Memory Attention to accelerate HRSAM. Within FLA, we implement Flash Swin attention, achieving over a 35% speedup compared to traditional Swin attention, and propose a KV-only padding mechanism to enhance extrapolation. We also develop the Cycle-scan module that uses State Space models to efficiently expand HRSAM's receptive field. We further develop the HRSAM++ within FLA by adding an anchor map, providing multi-scale data augmentation for the extrapolation and a larger receptive field at slight computational cost. Experiments show that, under standard training, HRSAMs surpass the previous SOTA with only 38% of the latency. With SAM-distillation, the extrapolation enables HRSAMs to outperform the teacher model at lower latency. Further finetuning achieves performance significantly exceeding the previous SOTA.