More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment

Aligning large language models (LLMs) with human values is an increasingly critical step in post-training. Direct Preference Optimization (DPO) has emerged as a simple, yet effective alternative to reinforcement learning from human feedback (RLHF). Synthetic preference data with its low cost and high quality enable effective alignment through single- or multi-model generated preference data. Our study reveals a striking, safety-specific phenomenon associated with DPO alignment: Although multi-model generated data enhances performance on general tasks (ARC, Hellaswag, MMLU, TruthfulQA, Winogrande) by providing diverse responses, it also tends to facilitate reward hacking during training. This can lead to a high attack success rate (ASR) when models encounter jailbreaking prompts. The issue is particularly pronounced when employing stronger models like GPT-4o or larger models in the same family to generate chosen responses paired with target model self-generated rejected responses, resulting in dramatically poorer safety outcomes. Furthermore, with respect to safety, using solely self-generated responses (single-model generation) for both chosen and rejected pairs significantly outperforms configurations that incorporate responses from stronger models, whether used directly as chosen data or as part of a multi-model response pool. We demonstrate that multi-model preference data exhibits high linear separability between chosen and rejected responses, which allows models to exploit superficial cues rather than internalizing robust safety constraints. Our experiments, conducted on models from the Llama, Mistral, and Qwen families, consistently validate these findings.

Bayesian Computation in Deep Learning

This review paper is intended for the 2nd edition of the Handbook of Markov chain Monte Carlo. We provide an introduction to approximate inference techniques as Bayesian computation methods applied to deep learning models. We organize the chapter by presenting popular computational methods for Bayesian neural networks and deep generative models, explaining their unique challenges in posterior inference as well as the solutions.

Making Reliable and Flexible Decisions in Long-tailed Classification

Long-tailed classification is challenging due to its heavy imbalance in class probabilities. While existing methods often focus on overall accuracy or accuracy for tail classes, they overlook a critical aspect: certain types of errors can carry greater risks than others in real-world long-tailed problems. For example, misclassifying patients (a tail class) as healthy individuals (a head class) entails far more serious consequences than the reverse scenario. To address this critical issue, we introduce Making Reliable and Flexible Decisions in Long-tailed Classification (RF-DLC), a novel framework aimed at reliable predictions in long-tailed problems. Leveraging Bayesian Decision Theory, we introduce an integrated gain to seamlessly combine long-tailed data distributions and the decision-making procedure. We further propose an efficient variational optimization strategy for the decision risk objective. Our method adapts readily to diverse utility matrices, which can be designed for specific tasks, ensuring its flexibility for different problem settings. In empirical evaluation, we design a new metric, False Head Rate, to quantify tail-sensitivity risk, along with comprehensive experiments on multiple real-world tasks, including large-scale image classification and uncertainty quantification, to demonstrate the reliability and flexibility of our method.

ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time

Vision Language Models (VLMs) have become essential backbones for multimodal intelligence, yet significant safety challenges limit their real-world application. While textual inputs are often effectively safeguarded, adversarial visual inputs can easily bypass VLM defense mechanisms. Existing defense methods are either resource-intensive, requiring substantial data and compute, or fail to simultaneously ensure safety and usefulness in responses. To address these limitations, we propose a novel two-phase inference-time alignment framework, Evaluating Then Aligning (ETA): 1) Evaluating input visual contents and output responses to establish a robust safety awareness in multimodal settings, and 2) Aligning unsafe behaviors at both shallow and deep levels by conditioning the VLMs' generative distribution with an interference prefix and performing sentence-level best-of-N to search the most harmless and helpful generation paths. Extensive experiments show that ETA outperforms baseline methods in terms of harmlessness, helpfulness, and efficiency, reducing the unsafe rate by 87.5% in cross-modality attacks and achieving 96.6% win-ties in GPT-4 helpfulness evaluation. The code is publicly available at https://github.com/DripNowhy/ETA.

Cascade Reward Sampling for Efficient Decoding-Time Alignment

Aligning large language models (LLMs) with human preferences is critical for their deployment. Recently, decoding-time alignment has emerged as an effective plug-and-play technique that requires no fine-tuning of model parameters. However, generating text that achieves both high reward and high likelihood remains a significant challenge. Existing methods often fail to generate high-reward text or incur substantial computational costs. In this paper, we propose Cascade Reward Sampling (CARDS) to address both issues, guaranteeing the generation of high-reward and high-likelihood text with significantly low costs. Based on our analysis of reward models (RMs) on incomplete text and our observation that high-reward prefixes induce high-reward complete text, we use rejection sampling to iteratively generate small semantic segments to form such prefixes. The segment length is dynamically determined by the predictive uncertainty of LLMs. This strategy guarantees desirable prefixes for subsequent generations and significantly reduces wasteful token re-generations and the number of reward model scoring. Our experiments demonstrate substantial gains in both generation efficiency and alignment ratings compared to the baselines, achieving five times faster text generation and 99\% win-ties in GPT-4/Claude-3 helpfulness evaluation.

Entropy-MCMC: Sampling from Flat Basins with Ease

Bayesian deep learning counts on the quality of posterior distribution estimation. However, the posterior of deep neural networks is highly multi-modal in nature, with local modes exhibiting varying generalization performance. Given a practical budget, sampling from the original posterior can lead to suboptimal performance, as some samples may become trapped in "bad" modes and suffer from overfitting. Leveraging the observation that "good" modes with low generalization error often reside in flat basins of the energy landscape, we propose to bias sampling on the posterior toward these flat regions. Specifically, we introduce an auxiliary guiding variable, the stationary distribution of which resembles a smoothed posterior free from sharp modes, to lead the MCMC sampler to flat basins. By integrating this guiding variable with the model parameter, we create a simple joint distribution that enables efficient sampling with minimal computational overhead. We prove the convergence of our method and further show that it converges faster than several existing flatness-aware methods in the strongly convex setting. Empirical results demonstrate that our method can successfully sample from flat basins of the posterior, and outperforms all compared baselines on multiple benchmarks including classification, calibration, and out-of-distribution detection.

Long-tailed Classification from a Bayesian-decision-theory Perspective

Long-tailed classification poses a challenge due to its heavy imbalance in class probabilities and tail-sensitivity risks with asymmetric misprediction costs. Recent attempts have used re-balancing loss and ensemble methods, but they are largely heuristic and depend heavily on empirical results, lacking theoretical explanation. Furthermore, existing methods overlook the decision loss, which characterizes different costs associated with tailed classes. This paper presents a general and principled framework from a Bayesian-decision-theory perspective, which unifies existing techniques including re-balancing and ensemble methods, and provides theoretical justifications for their effectiveness. From this perspective, we derive a novel objective based on the integrated risk and a Bayesian deep-ensemble approach to improve the accuracy of all classes, especially the "tail". Besides, our framework allows for task-adaptive decision loss which provides provably optimal decisions in varying task scenarios, along with the capability to quantify uncertainty. Finally, We conduct comprehensive experiments, including standard classification, tail-sensitive classification with a new False Head Rate metric, calibration, and ablation studies. Our framework significantly improves the current SOTA even on large-scale real-world datasets like ImageNet.

Trustworthy Long-Tailed Classification

Classification on long-tailed distributed data is a challenging problem, which suffers from serious class-imbalance and accordingly unpromising performance especially on tail classes. Recently, the ensembling based methods achieve the state-of-the-art performance and show great potential. However, there are two limitations for current methods. First, their predictions are not trustworthy for failure-sensitive applications. This is especially harmful for the tail classes where the wrong predictions is basically frequent. Second, they assign unified numbers of experts to all samples, which is redundant for easy samples with excessive computational cost. To address these issues, we propose a Trustworthy Long-tailed Classification (TLC) method to jointly conduct classification and uncertainty estimation to identify hard samples in a multi-expert framework. Our TLC obtains the evidence-based uncertainty (EvU) and evidence for each expert, and then combines these uncertainties and evidences under the Dempster-Shafer Evidence Theory (DST). Moreover, we propose a dynamic expert engagement to reduce the number of engaged experts for easy samples and achieve efficiency while maintaining promising performances. Finally, we conduct comprehensive experiments on the tasks of classification, tail detection, OOD detection and failure prediction. The experimental results show that the proposed TLC outperforms the state-of-the-art methods and is trustworthy with reliable uncertainty.

Graph Communal Contrastive Learning

Graph representation learning is crucial for many real-world applications (e.g. social relation analysis). A fundamental problem for graph representation learning is how to effectively learn representations without human labeling, which is usually costly and time-consuming. Graph contrastive learning (GCL) addresses this problem by pulling the positive node pairs (or similar nodes) closer while pushing the negative node pairs (or dissimilar nodes) apart in the representation space. Despite the success of the existing GCL methods, they primarily sample node pairs based on the node-level proximity yet the community structures have rarely been taken into consideration. As a result, two nodes from the same community might be sampled as a negative pair. We argue that the community information should be considered to identify node pairs in the same communities, where the nodes insides are semantically similar. To address this issue, we propose a novel Graph Communal Contrastive Learning (gCooL) framework to jointly learn the community partition and learn node representations in an end-to-end fashion. Specifically, the proposed gCooL consists of two components: a Dense Community Aggregation (DeCA) algorithm for community detection and a Reweighted Self-supervised Cross-contrastive (ReSC) training scheme to utilize the community information. Additionally, the real-world graphs are complex and often consist of multiple views. In this paper, we demonstrate that the proposed gCooL can also be naturally adapted to multiplex graphs. Finally, we comprehensively evaluate the proposed gCooL on a variety of real-world graphs. The experimental results show that the gCooL outperforms the state-of-the-art methods.

Identifying Incorrect Classifications with Balanced Uncertainty

Uncertainty estimation is critical for cost-sensitive deep-learning applications (i.e. disease diagnosis). It is very challenging partly due to the inaccessibility of uncertainty groundtruth in most datasets. Previous works proposed to estimate the uncertainty from softmax calibration, Monte Carlo sampling, subjective logic and so on. However, these existing methods tend to be over-confident about their predictions with unreasonably low overall uncertainty, which originates from the imbalance between positive (correct classifications) and negative (incorrect classifications) samples. For this issue, we firstly propose the distributional imbalance to model the imbalance in uncertainty estimation as two kinds of distribution biases, and secondly propose Balanced True Class Probability (BTCP) framework, which learns an uncertainty estimator with a novel Distributional Focal Loss (DFL) objective. Finally, we evaluate the BTCP in terms of failure prediction and out-of-distribution (OOD) detection on multiple datasets. The experimental results show that BTCP outperforms other uncertainty estimation methods especially in identifying incorrect classifications.