Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via .Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant --, a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to correct even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint.Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.

Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited.

Offline preference optimization offers a simpler and more stable alternative to RLHF for aligning language models. However, their effectiveness is critically dependent on ranking accuracy, a metric where further gains are highly im-pactful. This limitation arises from a fundamental problem that we identify and formalize as the Overfitting-Underfitting Dilemma: current margin designs cause models to apply excessive, wasteful gradients to correctly ranked samples (over-fitting) while providing insufficient corrective signals for misranked ones (underfitting). To resolve this dilemma, we propose Adaptive Margin-attached Preference Optimization (AMaPO), a simple yet principled algorithm. AMaPO employs an instance-wise adaptive margin, refined by Z-normalization and exponential scaling, which dynamically reallocates learning effort by amplifying gradients for mis-ranked samples and suppressing them for correct ones. Extensive experiments on widely used benchmarks demonstrate that AMaPO not only achieves better ranking accuracy and superior downstream alignment performance, but targeted analysis also confirms that it successfully mitigates the core overfitting and underfitting issues.

Project-Based Learning (PBL) involves a variety of highly correlated multimodal data, making it a vital educational approach within STEM disciplines. With the rapid development of multimodal large language models (MLLMs), researchers have begun exploring their potential to enhance tasks such as information retrieval, knowledge comprehension, and data generation in educational settings. However, existing benchmarks fall short in providing both a free-form output structure and a rigorous human expert validation process, limiting their effectiveness in evaluating real-world educational tasks. Additionally, few methods have developed automated pipelines to assist with the complex responsibilities of teachers leveraging MLLMs, largely due to model hallucination and instability, which lead to unreliable implementation. To address this gap, we introduce PBLBench, a novel benchmark designed to evaluate complex reasoning grounded in domain-specific knowledge and long-context understanding, thereby challenging models with tasks that closely resemble those handled by human experts. To establish reliable ground truth, we adopt the Analytic Hierarchy Process (AHP), utilizing expert-driven pairwise comparisons to derive structured and weighted evaluation criteria. We assess the performance of 15 leading MLLMs/LLMs using PBLBench and demonstrate that even the most advanced models achieve only 59% rank accuracy, underscoring the significant challenges presented by this benchmark. We believe PBLBench will serve as a catalyst for the development of more capable AI agents, ultimately aiming to alleviate teacher workload and enhance educational productivity.

Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via ranking accuracy.Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the idealized ranking accuracy that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant alignment gap -- i.e., a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to correct even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint.Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.

Bipartite ranking aims to learn a real-valued ranking function that orders positive instances before negative instances. Recent efforts of bipartite ranking are focused on optimizing ranking accuracy at the top of the ranked list. Most existing approaches are either to optimize task specific metrics or to extend the rank loss by emphasizing more on the error associated with the top ranked instances, leading to a high computational cost that is super-linear in the number of training instances. We propose a highly efficient approach, titled TopPush, for optimizing accuracy at the top that has computational complexity linear in the number of training instances. We present a novel analysis that bounds the generalization error for the top ranked instances for the proposed approach. Empirical study shows that the proposed approach is highly competitive to the state-of-the-art approaches and is 10-100 times faster.

Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via $\textit{ranking accuracy}$. Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the $\textit{idealized ranking accuracy}$ that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant $\textit{alignment gap}$ -- $\textit{i.e.}$, a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to fix even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint. Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.

Bipartite ranking aims to learn a real-valued ranking function that orders positive instances before negative instances. Recent efforts of bipartite ranking are focused on optimizing ranking accuracy at the top of the ranked list. Most existing approaches are either to optimize task specific metrics or to extend the rank loss by emphasizing more on the error associated with the top ranked instances, leading to a high computational cost that is super-linear in the number of training instances. We propose a highly efficient approach, titled TopPush, for optimizing accuracy at the top that has computational complexity linear in the number of training instances. We present a novel analysis that bounds the generalization error for the top ranked instances for the proposed approach. Empirical study shows that the proposed approach is highly competitive to the state-of-the-art approaches and is 10-100 times faster.

We introduce a new proxy score for evaluating bitext mining based on similarity in a multilingual embedding space: xSIM++. In comparison to xSIM, this improved proxy leverages rule-based approaches to extend English sentences in any evaluation set with synthetic, hard-to-distinguish examples which more closely mirror the scenarios we encounter during large-scale mining. We validate this proxy by running a significant number of bitext mining experiments for a set of low-resource languages, and subsequently train NMT systems on the mined data. In comparison to xSIM, we show that xSIM++ is better correlated with the downstream BLEU scores of translation systems trained on mined bitexts, providing a reliable proxy of bitext mining performance without needing to run expensive bitext mining pipelines. xSIM++ also reports performance for different error types, offering more fine-grained feedback for model development.