Ning Yan Gu

Chenxu Yang, Ruipeng Jia, Naibin Gu et al.

DPO is an effective preference optimization algorithm. However, the DPO-tuned models tend to overfit on the dispreferred samples, manifested as overly long generations lacking diversity. While recent regularization approaches have endeavored to alleviate this issue by modifying the objective function, they achieved that at the cost of alignment performance degradation. In this paper, we innovatively incorporate regularization from the perspective of weight updating to curb alignment overfitting. Through the pilot experiment, we discovered that there exists a positive correlation between overfitting and the hyperspherical energy fluctuation. Hence, we introduce orthogonal finetuning for DPO via a weight-Rotated Preference Optimization (RoPO) method, which merely conducts rotational and magnitude-stretching updates on the weight parameters to maintain the hyperspherical energy invariant, thereby preserving the knowledge encoded in the angle between neurons. Extensive experiments demonstrate that our model aligns perfectly with human preferences while retaining the original expressive capacity using only 0.0086% of the trainable parameters, suggesting an effective regularization against overfitting. Specifically, RoPO outperforms DPO by up to 10 points on MT-Bench and by up to 2.8 points on AlpacaEval 2, while enhancing the generation diversity by an average of 6 points.

Shitong Duan, Xiaoyuan Yi, Peng Zhang et al.

Large language models (LLMs) have revolutionized the role of AI, yet pose potential social risks. To steer LLMs towards human preference, alignment technologies have been introduced and gained increasing attention. Nevertheless, existing methods heavily rely on high-quality positive-negative training pairs, suffering from noisy positive responses that are barely distinguishable from negative ones. Given recent LLMs' proficiency in generating helpful responses, this work pivots towards a new research question: can we achieve alignment using solely human-annotated negative samples, preserving helpfulness while reducing harmfulness? For this purpose, we propose Distributional Dispreference Optimization (D$^2$O), which maximizes the discrepancy between dispreferred responses and the generated non-negative ones. In this way, D$^2$O effectively eschews harmful information without incorporating noisy positive samples, while avoiding collapse using self-generated responses as anchors. We demonstrate that D$^2$O can be regarded as learning a distributional preference model reflecting human dispreference against negative responses, which is theoretically an upper bound of the instance-level DPO. Extensive experiments manifest that our method achieves comparable generation quality and surpasses the latest strong baselines in producing less harmful and more informative responses with better training stability and faster convergence.

Yubin Shi, Yixuan Chen, Mingzhi Dong et al.

Despite their prevalence in deep-learning communities, over-parameterized models convey high demands of computational costs for proper training. This work studies the fine-grained, modular-level learning dynamics of over-parameterized models to attain a more efficient and fruitful training strategy. Empirical evidence reveals that when scaling down into network modules, such as heads in self-attention models, we can observe varying learning patterns implicitly associated with each module's trainability. To describe such modular-level learning capabilities, we introduce a novel concept dubbed modular neural tangent kernel (mNTK), and we demonstrate that the quality of a module's learning is tightly associated with its mNTK's principal eigenvalue $λ_{\max}$. A large $λ_{\max}$ indicates that the module learns features with better convergence, while those miniature ones may impact generalization negatively. Inspired by the discovery, we propose a novel training strategy termed Modular Adaptive Training (MAT) to update those modules with their $λ_{\max}$ exceeding a dynamic threshold selectively, concentrating the model on learning common features and ignoring those inconsistent ones. Unlike most existing training schemes with a complete BP cycle across all network modules, MAT can significantly save computations by its partially-updating strategy and can further improve performance. Experiments show that MAT nearly halves the computational cost of model training and outperforms the accuracy of baselines.

Shitong Duan, Xiaoyuan Yi, Peng Zhang et al.

Assessing Large Language Models (LLMs)' underlying value differences enables comprehensive comparison of their misalignment, cultural adaptability, and biases. Nevertheless, current value measurement datasets face the informativeness challenge: with often outdated, contaminated, or generic test questions, they can only capture the shared value orientations among different LLMs, leading to saturated and thus uninformative results. To address this problem, we introduce AdAEM, a novel, self-extensible assessment framework for revealing LLMs' inclinations. Distinct from previous static benchmarks, AdAEM can automatically and adaptively generate and extend its test questions. This is achieved by probing the internal value boundaries of a diverse set of LLMs developed across cultures and time periods in an in-context optimization manner. The optimization process theoretically maximizes an information-theoretic objective to extract the latest or culturally controversial topics, providing more distinguishable and informative insights about models' value differences. In this way, AdAEM is able to co-evolve with the development of LLMs, consistently tracking their value dynamics. Using AdAEM, we generate 12,310 questions grounded in Schwartz Value Theory, conduct an extensive analysis to manifest our method's validity and effectiveness, and benchmark the values of 16 LLMs, laying the groundwork for better value research.