Recent self-rewarding large language models (LLM) have successfully applied LLM-as-a-Judge to iteratively improve the alignment performance without the need of human annotations for preference data. These methods commonly utilize the same LLM to act as both the policy model (which generates responses) and the reward model (which scores and ranks those responses). The ranked responses are then used as preference pairs to train the LLM via direct alignment technologies (e.g. DPO). However, it is noteworthy that throughout this process, there is no guarantee of accuracy in the rewarding and ranking, which is critical for ensuring accurate rewards and high-quality preference data. Empirical results from relatively small LLMs (e.g., 7B parameters) also indicate that improvements from self-rewarding may diminish after several iterations in certain situations, which we hypothesize is due to accumulated bias in the reward system. This bias can lead to unreliable preference data for training the LLM. To address this issue, we first formulate and analyze the generalized iterative preference fine-tuning framework for self-rewarding language model. We then introduce the regularization to this generalized framework to mitigate the overconfident preference labeling in the self-rewarding process. Based on this theoretical insight, we propose a Consistency Regularized sElf-rewarding lAnguage Model (CREAM) that leverages the consistency of rewards across different iterations to regularize the self-rewarding training, helping the model to learn from more reliable preference data. With this explicit regularization, our empirical results demonstrate the superiority of CREAM in improving both reward consistency and alignment performance. The code is publicly available at https://github.com/Raibows/CREAM.

Inserting foreground objects into specific background scenes and eliminating the illumination inconsistency (eg., color, brightness) between them is an important and challenging task. It typically involves multiple processing tasks, such as image harmonization and shadow generation. In these two domains, there are already many mature solutions, but they often only focus on one of the tasks. Recently, some image composition methods have utilized diffusion models to address both of these issues simultaneously, but they cannot guarantee complete reconstruction of the foreground content. In this work, we propose CFDiffusion, which can simultaneously handle image harmonization and shadow generation. We first employ a shadow mask predictor to estimate the shadow mask of the foreground object. Next, we design a harmonization-shadow generator based on a diffusion model to harmonize the foreground and generate shadows concurrently. Additionally, we propose a foreground content enhancement module to ensure the complete preservation of foreground content at the insertion location, and we also develop an adaptive encoder to guide the harmonization process in the foreground area. The experimental results on the iHarmony4 dataset and the IH-SG dataset demonstrate the superiority of our CFDiffusion approach.