Adversarial DPO: Harnessing Harmful Data for Reducing Toxicity with Minimal Impact on Coherence and Evasiveness in Dialogue Agents (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing toxicity within large language models can negatively impact the user experience, causing performance degradation. |
| Approach: | They propose an adversarial DPO algorithm that improves direct preference optimization (DPO) by incorporating harmful data into the generative model. |
| Outcome: | The proposed training algorithm improves the model’s resilience against harmful conversations while minimizing performance degradation. |
Similar Papers
Adversary-Aware DPO: Enhancing Safety Alignment in Vision Language Models via Adversarial Training (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Adversary-aware DPO (ADPO) is a training framework that explicitly considers adversary. |
| Approach: | a new framework integrates adversarial training into a pre-trained large language model to enhance safety alignment . adversary-aware DPO provides a framework that explicitly considers adversary . |
| Outcome: | a new training framework outperforms baselines in safety alignment and general utility of large language models. |
How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis (2025.emnlp-main)
Copied to clipboard
| Challenge: | Direct Preference Optimization (DPO) is a popular choice of safety fine-tuning algorithms, but prior explanations of its effects only account for dampened toxic neurons in the MLP layers. |
| Approach: | They analysed four language models and found that toxic neurons only account for 2.5% to 24% of DPO’s effects across models. |
| Outcome: | The proposed method outperforms DPO in reducing toxicity while preserving perplexity, without requiring any weight updates. |
Adversarial Preference Learning for Robust LLM Alignment (2025.findings-acl)
Copied to clipboard
Yuanfu Wang, Pengyu Wang, Chenyang Xi, Bo Tang, Junyi Zhu, Wenqiang Wei, Chen Chen, Chao Yang, Jingfeng Zhang, Chaochao Lu, Yijun Niu, Keming Mao, Zhiyu Li, Feiyu Xiong, Jie Hu, Mingchuan Yang
| Challenge: | Modern language models rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors, but they remain vulnerable to adversarial attacks due to three key limitations: (1) the inefficiency and high cost of human annotation; (2) the vast diversity of potential adversarials; and (3) the risk of feedback bias and reward hacking. |
| Approach: | They propose an iterative adversarial training method that incorporates three key innovations to address these challenges. |
| Outcome: | Experiments on Mistral-7B-Instruct-v0.3 show that the proposed method significantly enhances robustness and reduces harmful outputs from 5.88% to 0.43%. |
Towards Building a Robust Toxicity Predictor (2023.acl-industry)
Copied to clipboard
| Challenge: | Recent studies have focused on robustness of toxicity language predictors, but this is problematic for real-world toxicity detection. |
| Approach: | They propose a novel adversarial attack that exploits greedy search strategies to fool toxic text classifiers. |
| Outcome: | The proposed attack can detect weaker toxicity language detectors even against unseen attacks. |
Bot-Adversarial Dialogue for Safe Conversational Agents (2021.naacl-main)
Copied to clipboard
| Challenge: | a new method for evaluating chatbot safety is proposed to mimic human-generated data . a bot-adversarial dialogue model learns undesirable features from this data, a study finds . |
| Approach: | They propose a human-and-model-in-the-loop framework for evaluating toxicity of chatbots . they propose two methods for safe conversational agents by either training on data or ”baking-in” safety to the generative model itself. |
| Outcome: | The proposed methods are safer than existing models while maintaining usability metrics, the authors say . they show that the proposed methods can be used to make safer models with human-model interactions . |
RS-DPO: A Hybrid Rejection Sampling and Direct Preference Optimization Method for Alignment of Large Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Reinforcement learning with human feedback (RLHF) is widely employed to align large language models with user intent. |
| Approach: | They propose to combine rejection sampling and direct preference optimization to improve alignment with user intent by identifying pairs of contrastive samples from human annotator and alternative LLMs. |
| Outcome: | The proposed method outperforms existing methods including RS, PPO, and DPO in a limited resource environment. |
Constructing Highly Inductive Contexts for Dialogue Safety through Controllable Reverse Generation (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to detect toxic generation of pretrained language models rely on templates, data extraction, crowdsourcing workers or automatic generation. |
| Approach: | They propose a method to construct adversarial contexts conditioned on a given response . they augment existing dataset BAD+ and construct a new dataset B AD+ . |
| Outcome: | The proposed method can detect toxic or biased content in large pretrained language models. |
Full-Step-DPO: Self-Supervised Preference Optimization with Step-wise Rewards for Mathematical Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to improve long-chain mathematical reasoning focus on the first erroneous step, but ignore all other steps and rely heavily on external signals. |
| Approach: | They propose a DPO framework that leverages step-wise rewards from the entire reasoning chain instead of optimizing only the first erroneous step. |
| Outcome: | The proposed framework improves on in-domain and out-of-domain mathematical reasoning benchmarks. |
BiasDPO: Mitigating Bias in Language Models through Direct Preference Optimization (2024.acl-srw)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have been shown to be effective in complex language tasks, but their potential to perpetuate biases poses significant concerns. |
| Approach: | They propose a new framework employing Direct Preference Optimization to mitigate biases in LLMs. |
| Outcome: | The proposed model outperforms the baseline model on almost all bias benchmarks and achieves better performance than open-source models. |
Causal Direct Preference Optimization for Language Model Alignment (2026.findings-eacl)
Copied to clipboard
| Challenge: | Empirical evaluations show that CDPO surpasses DPO-based baselines by achieving unbiased fine-tuning through causal reasoning. |
| Approach: | They propose a framework that incorporates causal inference principles to mitigate the influence of confounders and sharpen the signal of genuine human preferences. |
| Outcome: | The proposed framework preserves the tractability of direct optimization while enhancing robustness to spurious correlations and annotation biases. |