Reinforcement learning from human feedback (RLHF) is widely used to align large language models (LLMs) with human preferences. However, RLHF-trained reward models often exhibit length bias -- a systematic tendency to favor longer responses by conflating verbosity with quality. We propose a causal framework for analyzing and mitigating length bias in RLHF reward modeling. Central to our approach is a counterfactual data augmentation method that generates response pairs designed to isolate content quality from verbosity. These counterfactual examples are then used to train the reward model, enabling it to assess responses based on content quality independently of verbosity. Specifically, we construct (1) length-divergent pairs with similar content and (2) content-divergent pairs of similar length. Empirical evaluations show that our method reduces length bias in reward assignment and leads to more concise, content-focused outputs from the policy model. These findings demonstrate that the proposed approach effectively reduces length bias and improves the robustness and content sensitivity of reward modeling in RLHF pipelines.

Recent research has attempted to associate preference optimization (PO) performance with the underlying preference datasets. In this work, our observation is that the differences between the preferred response $y^+$ and dispreferred response $y^-$ influence what LLMs can learn, which may not match the desirable differences to learn. Therefore, we use distance and reward margin to quantify these differences, and combine them to get Distance Calibrated Reward Margin (DCRM), a metric that measures the quality of a response pair for PO. Intuitively, DCRM encourages minimal noisy differences and maximal desired differences. With this, we study 3 types of commonly used preference datasets, classified along two axes: the source of the responses and the preference labeling function. We establish a general correlation between higher DCRM of the training set and better learning outcome. Inspired by this, we propose a best-of-$N^2$ pairing method that selects response pairs with the highest DCRM. Empirically, in various settings, our method produces training datasets that can further improve models' performance on AlpacaEval, MT-Bench, and Arena-Hard over the existing training sets.

Large vision-language models (L VLMs) remain vulnerable to hallucination, often generating content misaligned with visual inputs. Although recent training-based approaches aim to mitigate hallucination, they typically rely on predefined or randomly edited negative samples that do not reflect actual model errors, thus limiting training efficacy. In this work, we propose an Online Vision-language Preference Learning (OViP) framework that dynamically constructs contrastive training data based on the model's own hallucinated outputs. By identifying semantic differences between sampled response pairs and synthesizing negative images using a diffusion model, OViP generates more relevant supervision signals in real time. This failure-driven training enables adaptive alignment of both textual and visual preferences. Moreover, we refine existing evaluation protocols to better capture the trade-off between hallucination suppression and expressiveness. Experiments on hallucination and general benchmarks demonstrate that OViP not only reduces hallucinations while preserving core multi-modal capabilities, but also substantially improves training efficiency. However, L VLMs continue to struggle with persistent hallucination issues (Li et al., 2023b; Bai et al., 2024), often exhibiting incorrect references to visual content (Liu et al., 2024a; Zhou et al., 2023; Bai et al., 2024). These errors manifest as misattributing object properties, describing nonexistent entities, or fabricating spatial relationships that do not align with the image. Such inconsistencies undermine the model's faithfulness to the input and hinder further reasoning capabilities, significantly limiting the reliability of L VLMs in real-world applications. Recent success of Direct Preference Optimization (DPO) (Rafailov et al., 2023) in LLMs alignment has inspired the exploration of multi-modal DPO to mitigate hallucination in L VLMs (Y u et al., 2024a;b; Xie et al., 2024; Sarkar et al., 2024). However, early efforts directly extend the original DPO designs from LLMs to L VLMs by constructing preference pairs solely on textual responses given the same image input, primarily focusing on response-side preference optimization and showing limited effectiveness.

Large language models (LLMs) are being widely applied across various fields, but as tasks become more complex, evaluating their responses is increasingly challenging. Compared to human evaluators, the use of LLMs to support performance evaluation offers a more efficient alternative. However, most studies focus mainly on aligning LLMs' judgments with human preferences, overlooking the existence of biases and mistakes in human judgment. Furthermore, how to select suitable LLM judgments given multiple potential LLM responses remains underexplored. To address these two aforementioned issues, we propose a three-stage meta-judge selection pipeline: 1) developing a comprehensive rubric with GPT-4 and human experts, 2) using three advanced LLM agents to score judgments, and 3) applying a threshold to filter out low-scoring judgments. Compared to methods using a single LLM as both judge and meta-judge, our pipeline introduces multi-agent collaboration and a more comprehensive rubric. Experimental results on the JudgeBench dataset show about 15.55\% improvement compared to raw judgments and about 8.37\% improvement over the single-agent baseline. Our work demonstrates the potential of LLMs as meta-judges and lays the foundation for future research on constructing preference datasets for LLM-as-a-judge reinforcement learning.

Aligning large language models (LLMs) with human preferences has achieved remarkable success. However, existing Chinese preference datasets are limited by small scale, narrow domain coverage, and lack of rigorous data validation. Additionally, the reliance on human annotators for instruction and response labeling significantly constrains the scalability of human preference datasets. To address these challenges, we design an LLM-based Chinese preference dataset annotation pipeline with no human intervention. Specifically, we crawled and carefully filtered 92k high-quality Chinese queries and employed 15 mainstream LLMs to generate and score chosen-rejected response pairs. Based on it, we introduce COIG-P (Chinese Open Instruction Generalist - Preference), a high-quality, large-scale Chinese preference dataset, comprises 1,009k Chinese preference pairs spanning 6 diverse domains: Chat, Code, Math, Logic, Novel, and Role. Building upon COIG-P, to reduce the overhead of using LLMs for scoring, we trained a 8B-sized Chinese Reward Model (CRM) and meticulously constructed a Chinese Reward Benchmark (CRBench). Evaluation results based on AlignBench \citep{liu2024alignbenchbenchmarkingchinesealignment} show that that COIG-P significantly outperforms other Chinese preference datasets, and it brings significant performance improvements ranging from 2% to 12% for the Qwen2/2.5 and Infinity-Instruct-3M-0625 model series, respectively. The results on CRBench demonstrate that our CRM has a strong and robust scoring ability. We apply it to filter chosen-rejected response pairs in a test split of COIG-P, and our experiments show that it is comparable to GPT-4o in identifying low-quality samples while maintaining efficiency and cost-effectiveness. Our codes and data are released in https://github.com/multimodal-art-projection/COIG-P.

Modern large language models (LLMs) are optimized for human-aligned responses using Reinforcement Learning from Human Feedback (RLHF). However, existing RLHF approaches assume a universal preference model and fail to account for individual user preferences, limiting their effectiveness in personalized applications. We introduce a framework that extends RLHF to enable user personalization by leveraging the assumption that user preferences lie in a low-dimensional space. Instead of training a separate model per user, we represent user-specific rewards as a linear combination of base reward functions. Using only 10 user responses, our method can infer user-specific rewards and align LLM outputs accordingly. We validate our approach through experiments with both synthetic and real users, demonstrating significant personalization achieved by our method. In human evaluations, our method achieves a 67% win rate over default GPT-4o responses. Code and demo are available at https://idanshen.github.io/PReF/.

Training data quality is one of the most important drivers of final model quality. In this work, we introduce a method for evaluating data integrity based on the assumption that low-quality input prompts result in high variance and low quality responses. This is achieved by measuring the rejected response quality and the reward gap between the chosen and rejected preference pair. Our method, Rejecting Instruction Preferences (RIP) can be used to filter prompts from existing training sets, or to make high quality synthetic datasets, yielding large performance gains across various benchmarks compared to unfiltered data. Using Llama 3.1-8B-Instruct, RIP improves AlpacaEval2 LC Win Rate by 9.4%, Arena-Hard by 8.7%, and WildBench by 9.9%. Using Llama 3.3-70B-Instruct, RIP improves Arena-Hard from 67.5 to 82.9, which is from 18th place to 6th overall in the leaderboard.

Aligning large language models (LLMs) to human preferences is a crucial step in building helpful and safe AI tools, which usually involve training on supervised datasets. Popular algorithms such as Direct Preference Optimization rely on pairs of AI-generated responses ranked according to human feedback. The response pair annotation process is the most labor-intensive and costly part of the alignment pipeline, and improving its efficiency and annotation quality would have a meaningful impact on AI development. We propose REAL: Response Embedding-based Alignment for LLMs, a strategy for constructing a high-quality training dataset that focuses on acquiring the most informative response pairs for labeling out of a set of response candidates. Our selection process is based on embedding responses independently of prompts. Experimental results on real-world dataset SHP2 and synthetic HH-RLHF benchmarks indicate that choosing dissimilar response pairs enhances the direct alignment of LLMs while reducing inherited labeling errors. The model aligned on dissimilar response pairs obtained a better margin and win rate on the dialogue task. Our findings suggest that focusing on distinct pairs can reduce the label error to improve the efficiency of LLM alignment, saving up to 65% of annotators' work.

While Large Language Models (LLMs) have become central tools in various fields, they often provide inaccurate or false information. This study examines user preferences regarding falsehood responses from LLMs. Specifically, we evaluate preferences for LLM responses where false statements are explicitly marked versus unmarked responses and preferences for confident falsehoods compared to LLM disclaimers acknowledging a lack of knowledge. Additionally, we investigate how requiring users to assess the truthfulness of statements influences these preferences. Surprisingly, 61\% of users prefer unmarked falsehood responses over marked ones, and 69\% prefer confident falsehoods over LLMs admitting lack of knowledge. In all our experiments, a total of 300 users participated, contributing valuable data to our analysis and conclusions. When users are required to evaluate the truthfulness of statements, preferences for unmarked and falsehood responses decrease slightly but remain high. These findings suggest that user preferences, which influence LLM training via feedback mechanisms, may inadvertently encourage the generation of falsehoods. Future research should address the ethical and practical implications of aligning LLM behavior with such preferences.

Reinforcement Learning with Human Feedback (RLHF) is the key to the success of large language models (LLMs) in recent years. In this work, we first introduce the concepts of knowledge breadth and knowledge depth, which measure the comprehensiveness and depth of an LLM or knowledge source respectively. We reveal that the imbalance in the number of prompts and responses can lead to a potential disparity in breadth and depth learning within alignment tuning datasets by showing that even a simple uniform method for balancing the number of instructions and responses can lead to significant improvements. Building on this, we further propose Balanced Preference Optimization (BPO), designed to dynamically augment the knowledge depth of each sample. BPO is motivated by the observation that the usefulness of knowledge varies across samples, necessitating tailored learning of knowledge depth. To achieve this, we introduce gradient-based clustering, estimating the knowledge informativeness and usefulness of each augmented sample based on the model's optimization direction. Our experimental results across various benchmarks demonstrate that BPO outperforms other baseline methods in alignment tuning while maintaining training efficiency. Furthermore, we conduct a detailed analysis of each component of BPO, providing guidelines for future research in preference data optimization.