Papers with direct preference optimization
Direct Preference Optimization of Video Large Multimodal Models from Language Model Reward (2025.naacl-long)
Copied to clipboard
Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander G Hauptmann, Yonatan Bisk, Yiming Yang
| Challenge: | Existing studies have demonstrated that direct preference optimization (DPO) can be effective in generalizing large language models, but its effectiveness in video domain remains limited. |
| Approach: | They propose a framework that utilizes detailed video captions as a proxy of video content to enable language models to incorporate this information as supporting evidence for scoring video Question Answering (QA) predictions. |
| Outcome: | The proposed framework shows that it can be used to align language models with video content and improves performance on open-ended video QA tasks. |
Assessing and Mitigating Medical Knowledge Drift and Conflicts in Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Rapid medical concept drift can lead LLMs to provide incorrect or outdated advice. |
| Approach: | They propose to evaluate how large language models manage knowledge conflicts in clinical guidelines. |
| Outcome: | The proposed benchmark evaluates how LLMs manage varied knowledge conflicts in clinical guidelines. |
Learning to Summarize from LLM-generated Feedback (2025.naacl-long)
Copied to clipboard
| Challenge: | Developing effective text summarizers remains a challenge due to issues like unfaithful statements, key information omissions, and verbosity. |
| Approach: | They propose a large-scale dataset containing multi-dimensional feedback on LLM-generated summaries of varying quality across diverse domains to align them with human preferences for faithfulness, completeness, and conciseness. |
| Outcome: | The proposed model outperforms the 10x larger Llama3-70b-instruct in generating human-preferred summaries. |
Adaptive Helpfulness–Harmlessness Alignment with Preference Vectors (2026.eacl-long)
Copied to clipboard
Ren-Wei Liang, Chin Ting Hsu, Chan-Hung Yu, Saransh Agrawal, Shih-Cheng Huang, Chieh-Yen Lin, Shang-Tse Chen, Kuan-Hao Huang, Shao-Hua Sun
| Challenge: | Existing approaches to balancing helpfulness and harmlessness suffer from performance conflicts, limited controllability, and poor extendability. |
| Approach: | They propose a framework that allows users to control their own preferences and dynamically merge them at test time. |
| Outcome: | The proposed framework improves helpfulness without conservatism and smooth control over preference trade-offs. |
Direct Judgement Preference Optimization (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing judge models are largely trained with supervised finetuning on small data scales to perform limited types of evaluation tasks, limiting generalization. |
| Approach: | They propose to train judge models at large data scales with direct preference optimization . they use four training tasks to form three types of preference pairs targeting different aspects of evaluation . |
| Outcome: | The proposed model outperforms GPT-4o and other similar models on 13 benchmarks. |
Evaluating Psychological Safety of Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | a recent study evaluated the psychological safety of large language models. |
| Approach: | They designed unbiased prompts to evaluate the psychological safety of large language models. |
| Outcome: | The proposed prompts showed that they were fine-tuned with behavioral metrics to reduce toxicity. |
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. |
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. |
LegalDrill: Diagnosis-Driven Synthesis for Legal Reasoning in Small Language Models (2026.acl-industry)
Copied to clipboard
Tianchun Li, Haochen Liu, Vishwa Pardeshi, Xingchen Wang, Tianci Liu, Huijun Zhao, Wei Fan, Jing Gao
| Challenge: | Small language models (SLMs) are promising for real-world deployment but struggle with high-stakes legal reasoning tasks. |
| Approach: | They propose a diagnostic-driven synthesis framework that extracts and refines reasoning trajectories from a capable teacher via fine-grained prompting and a self-reflective verification is employed to adaptively select the most effective data for the SLM student. |
| Outcome: | The proposed framework extracts and refines reasoning trajectories from a capable teacher via fine-grained prompting, then a self-reflective verification is employed to adaptively select the most effective data for the student. |
Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization (2025.emnlp-main)
Copied to clipboard
Shuo Xing, Peiran Li, Yuping Wang, Ruizheng Bai, Yueqi Wang, Chan-Wei Hu, Chengxuan Qian, Huaxiu Yao, Zhengzhong Tu
| Challenge: | emergence of large Vision Language Models (VLMs) has broadened the capabilities of single-modal Large Language Model (LLM) but VLMs are prone to significant hallucinations, especially in the form of cross-modal inconsistencies. |
| Approach: | They propose a new alignment framework that leverages image retrieval to integrate both textual and visual preference signals. |
| Outcome: | The proposed framework mitigates hallucinations more effectively than previous methods . it maintains robustness and scalability across a wide range of VLM sizes and architectures . |
Revisiting Self-Play Preference Optimization: On the Role of Prompt Difficulty (2026.findings-acl)
Copied to clipboard
| Challenge: | incorporating difficult prompts into training fails to enhance overall performance, e.g., as prompt difficulty decreases. |
| Approach: | They investigate how prompts of varying difficulty influence self-play preference optimization . they use the reward of sampled responses of a prompt as a proxy for its difficulty . |
| Outcome: | The proposed model improves on difficult prompts and easy prompts, but fails to train on difficult ones and learns from failures. |
Supportiveness-based Knowledge Rewriting for Retrieval-augmented Language Modeling (2025.findings-naacl)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have significantly enhanced their performance in various natural language processing tasks. |
| Approach: | They propose a robust and pluggable knowledge rewriter that is optimized for LLM generation by supporting the model's supportiveness. |
| Outcome: | The proposed model can be used to rewrite knowledge in a supervised manner. |
CALM: Unleashing the Cross-Lingual Self-Aligning Ability of Language Model Question Answering (2025.findings-naacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are pre-trained on extensive multilingual corpora to acquire both language-specific cultural knowledge and general knowledge. |
| Approach: | They propose to use the **C**ross-Lingual Self-**Aligning ability of **L**anguage **M**odels to align knowledge across languages. |
| Outcome: | The proposed model performs well in both zero-shot and retrieval-augmented settings. |
Word Alignment as Preference for Machine Translation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Hallucination and omission are a problem in machine translation because of an LLM's size and low-resource languages. |
| Approach: | They propose to use word alignment as preference to optimize an LLM-based MT model to mitigate hallucination and omission problems. |
| Outcome: | The proposed model is able to mitigate hallucination and omission by using word alignment as preference. |
Dually Self-Improved Counterfactual Data Augmentation Using Large Language Model (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to generate counterfactual data augmentation are limited due to imbalance and biases in real-world training data. |
| Approach: | They propose a self-improved method for generating high-quality counterfacts using large language models. |
| Outcome: | The proposed method generates high-quality counterfacts on the natural language inference task using lightweight and task-specific LLMs. |
VLFeedback: A Large-Scale AI Feedback Dataset for Large Vision-Language Models Alignment (2024.emnlp-main)
Copied to clipboard
Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, Qi Liu
| Challenge: | Large vision-language models (LVLMs) are evolving rapidly and require data with human supervision to achieve better alignment. |
| Approach: | They introduce VLFeedback, the first large-scale vision-language feedback dataset . they train Silkie, an LVLM fine-tuned via direct preference optimization . |
| Outcome: | The proposed model outperforms its base model in helpfulness, visual faithfulness, and safety metrics and exhibits enhanced resilience against red-teaming attacks. |
ADELIE: Aligning Large Language Models on Information Extraction (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) struggle to follow complex instructions of IE tasks due to not being aligned with humans. |
| Approach: | They propose an aligned large language moDEL that effectively solves various IE tasks including closed IE, open IE and on-demand IE. |
| Outcome: | The proposed model achieves state-of-the-art (SoTA) performance among open-source models. |
User-Assistant Bias in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Modern large language models are typically trained using structured role tags . asymmetries in training data associated with different role tags can potentially introduce inductive biases. |
| Approach: | They propose a task-agnostic benchmark to test user–assistant bias in large language models . they find human-preference alignment amplifies user bias, reasoning fine-tuning reduces it . |
| Outcome: | The proposed benchmark tests show that most instruction-tuned models exhibit strong user bias . human-preference alignment amplifies user bias, while reasoning fine-tuning reduces it. |
Safety Is Not Universal: The Selective Safety Trap in LLM Alignment (2026.findings-acl)
Copied to clipboard
Iago Alves Brito, Walcy Rios, Julia Soares Dollis, Diogo Fernandes Costa Silva, Arlindo Rodrigues Galvão Filho
| Challenge: | Existing safety evaluations of large language models aggregate harms under generic categories such as "Identity Hate" a bilingual benchmark identifies a selective safety trap, where defense rates vary by up to 42% within the same model solely based on the target group. |
| Approach: | They propose a bilingual adversarial benchmark to audit selective safety in large language models . defense rates vary by up to 42% within the same model solely based on target group . |
| Outcome: | The proposed benchmark identifies a selective safety trap in large language models . defense rates vary by up to 42% within the same model solely based on the target group. |
PURE: Aligning LLM via Pluggable Query Reformulation for Enhanced Helpfulness (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) depend on vast amounts of text data sourced from the Internet for their training. |
| Approach: | They propose a new alignment paradigm that reformulates risky queries into highly relevant yet harmless ones before feeding them into LLMs. |
| Outcome: | The proposed approach eliminates the high costs of training base LLMs and achieves a promising balance of harmlessness and helpfulness. |
Context-DPO: Aligning Language Models for Context-Faithfulness (2025.findings-acl)
Copied to clipboard
Baolong Bi, Shaohan Huang, Yiwei Wang, Tianchi Yang, Zihan Zhang, Haizhen Huang, Lingrui Mei, Junfeng Fang, Zehao Li, Furu Wei, Weiwei Deng, Feng Sun, Qi Zhang, Shenghua Liu
| Challenge: | Context-DPO is the first alignment method specifically designed to enhance contextfaithfulness for large language models. |
| Approach: | They propose a benchmark that simulates Retrieval-Augmented Generation scenarios with knowledge conflicts to evaluate context-faithfulness. |
| Outcome: | The proposed method improves LLMs' context-faithfulness by 35% to 280% over open-source models. |
Low-Resource Language Expansion and Translation Capacity Enhancement for LLM: A Study on the Uyghur (2025.coling-main)
Copied to clipboard
Kaiwen Lu, Yating Yang, Fengyi Yang, Rui Dong, Bo Ma, Aihetamujiang Aihemaiti, Abibilla Atawulla, Lei Wang, Xi Zhou
| Challenge: | Extensive experiments have shown that our strategy effectively expands the low-resource languages supported by large language models and significantly enhances the model’s translation ability in Uyghur with less parallel data. |
| Approach: | They propose a direct preference optimization based on translation self-evolution to expand low-resource languages into large language models by using Uyghur as an example. |
| Outcome: | The proposed strategy expands low-resource languages supported by large language models and significantly enhances the model’s translation ability in Uyghur with less parallel data. |
ChatGLM-Math: Improving Math Problem-Solving in Large Language Models with a Self-Critique Pipeline (2024.findings-emnlp)
Copied to clipboard
Yifan Xu, Xiao Liu, Xinghan Liu, Zhenyu Hou, Yueyan Li, Xiaohan Zhang, Zihan Wang, Aohan Zeng, Zhengxiao Du, Zhao Wenyi, Jie Tang, Yuxiao Dong
| Challenge: | Large language models (LLMs) have shown excellent mastering of human language but struggle in real-world applications that require mathematical problem-solving. |
| Approach: | They propose a pipeline to train a general Math-Critique model from the LLM itself to provide feedback signals and employ rejective fine-tuning and direct preference optimization over the Llm's own generations for data collection. |
| Outcome: | The proposed pipeline outperforms existing LLMs that could be two times larger. |
InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance (2024.emnlp-main)
Copied to clipboard
Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, Xipeng Qiu
| Challenge: | Existing methods for enhancing harmlessness and helpfulness of large language models (LLMs) involve complex and resource-intensive training processes. |
| Approach: | They propose a method that decouples harmlessness from helpfulness during inference phase. |
| Outcome: | The proposed method significantly reduces the attack success rate (ASR) of harmful instructions and jailbreak instructions while maintaining almost unchanged performance in downstream tasks. |
InfoPO: On Mutual Information Maximization for Large Language Model Alignment (2025.naacl-long)
Copied to clipboard
Teng Xiao, Zhen Ge, Sujay Sanghavi, Tian Wang, Julian Katz-Samuels, Marc Versage, Qingjun Cui, Trishul Chilimbi
| Challenge: | Recent studies have shown that direct preference optimization and its variants can be useful for fine-tuning large language models with human preferences data. |
| Approach: | They propose a preference fine-tuning algorithm that effectively and efficiently aligns large language models using preference data. |
| Outcome: | Extensive experiments show that the proposed algorithm outperforms established baselines on reasoning tasks. |
JAWAHER: A Multidialectal Dataset of Arabic Proverbs for LLM Benchmarking (2025.naacl-long)
Copied to clipboard
Samar Mohamed Magdy, Sang Yun Kwon, Fakhraddin Alwajih, Safaa Taher Abdelfadil, Shady Shehata, Muhammad Abdul-Mageed
| Challenge: | Recent advances in instruction fine-tuning and alignment methods have enhanced the adaptability of large language models to user preferences. |
| Approach: | They propose a benchmark to assess LLMs’ capacity to comprehend and interpret Arabic proverbs. |
| Outcome: | The proposed model can generate accurate translations, but struggle to produce culturally nuanced and contextually relevant explanations. |
HFT: Half Fine-Tuning for Large Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) with one or more fine-tuning phases can unlock various capabilities, but can be catastrophic forgetting during sequential training. |
| Approach: | They propose a method to regularly reset partial parameters to mitigate forgetting issues by using half fine-tuning instead of full fine-uning. |
| Outcome: | The proposed approach reduces the risk of catastrophic forgetting during training and the parametric knowledge lost during training may be overwhelmed by incoming training data. |
Self-Training with Direct Preference Optimization Improves Chain-of-Thought Reasoning (2024.acl-long)
Copied to clipboard
| Challenge: | Recent studies focus on enhancing large-scale language models' reasoning abilities, but the research question of how to GSM8K Performance vs. computational cost remains. |
| Approach: | They propose to train small-scale language models with their own outputs to avoid relying on large models' outputs. |
| Outcome: | The proposed approach outperforms baseline models with comparable sizes while minimizing the required compute. |
Hybrid Alignment Training for Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to align large language models with instructions and preferences are conflicting . et al., 2023b) show that hybrid alignment training can outperform baselines . |
| Approach: | They propose a hybrid alignment training approach based on alternating alignment and modified elastic weight consolidation methods to achieve better collaboration between different alignment tasks. |
| Outcome: | The proposed approach outperforms baseline alignment training methods on summarization and dialogue tasks. |
Investigating and Enhancing the Robustness of Large Multimodal Models Against Temporal Inconsistency (2025.acl-long)
Copied to clipboard
Jiafeng Liang, Shixin Jiang, Xuan Dong, Ning Wang, Zheng Chu, Hui Su, Jinlan Fu, Ming Liu, See-Kiong Ng, Bing Qin
| Challenge: | Large Multimodal Models (LMMs) have demonstrated impressive performance on general video comprehension benchmarks, but their robustness needs to be thoroughly investigated for broader applications. |
| Approach: | They propose a temporal robustness benchmark which introduces temporal inconsistency perturbations separately at the visual and textual modalities to assess the robustness of models. |
| Outcome: | The proposed method improves the model’s robustness and reliability in temporal analysis. |
Textual Aesthetics in Large Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on image aesthetics have focused on content correctness and helpfulness of responses. |
| Approach: | They propose a textual aesthetics-powered fine-tuning method that leverages textual visual aesthetics without compromising content correctness. |
| Outcome: | The proposed method improves aesthetic scores and performs well on general evaluation datasets. |
A Necessary Step toward Faithfulness: Measuring and Improving Consistency in Free-Text Explanations (2025.emnlp-main)
Copied to clipboard
| Challenge: | a measure of faithful free-text explanations is difficult to generate by language models and assess by humans. |
| Approach: | They propose a measure of Prediction-EXplanation consistency by extending the concept of weight of evidence. |
| Outcome: | The proposed measure improves explanation faithfulness by up to 9.7%, the authors show . they show that applying preference optimization improves the consistency of generated explanations across three model families. |
LoGU: Long-form Generation with Uncertainty Expressions (2025.acl-long)
Copied to clipboard
Ruihan Yang, Caiqi Zhang, Zhisong Zhang, Xinting Huang, Sen Yang, Nigel Collier, Dong Yu, Deqing Yang
| Challenge: | Large Language Models (LLMs) generate factually incorrect content, i.e., hallucinations, despite impressive performance. |
| Approach: | They propose a framework to enable models to express uncertainty when unsure . they propose atomic claims to refine uncertainty and refine it using supervised fine-tuning and direct preference optimization to enhance uncertainty expression. |
| Outcome: | The proposed framework significantly improves accuracy, reduces hallucinations, and maintains comprehensiveness of responses. |
A Modular Approach for Clinical SLMs Driven by Synthetic Data with Pre-Instruction Tuning, Model Merging, and Clinical-Tasks Alignment (2025.acl-long)
Copied to clipboard
Jean-Philippe Corbeil, Amin Dada, Jean-Michel Attendu, Asma Ben Abacha, Alessandro Sordoni, Lucas Caccia, Francois Beaulieu, Thomas Lin, Jens Kleesiek, Paul Vozila
| Challenge: | Large language models such as GPT-4 have limited their deployment in clinical settings . a novel framework for adapting SLMs into high-performing clinical models is needed . |
| Approach: | They propose a framework for adapting large language models into high-performing clinical models . they pre-instruct experts on relevant medical and clinical corpora and model merging . |
| Outcome: | The proposed framework outperforms the existing model on the CLUE+ benchmark on medical entities and radiology reports. |
AlignDistil: Token-Level Language Model Alignment as Adaptive Policy Distillation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for LLM alignment optimize tokens using a sparse, response-level reward or preference annotation. |
| Approach: | They propose an RLHF-equivalent distillation method for token-level reward optimization that incorporates the reward learned by DPO into the RLHG objective and builds a token-based teacher distribution. |
| Outcome: | The proposed method bridges the accuracy gap between the reward from the DPO model and the pure reward model by building a contrastive DPO reward with a normal and a reverse DPO. |
Retrieval-Augmented Fine-Tuning With Preference Optimization For Visual Program Generation (2025.acl-long)
Copied to clipboard
| Challenge: | Visual programming languages (VPLs) allow users to create programs through graphical interfaces, which results in easier accessibility and widespread usage in various domains. |
| Approach: | They propose to train VPLs from user instructions using large language models . they propose to use retrieval-augmented fine-tuning to leverage repetitive use of subroutines . |
| Outcome: | The proposed method outperforms prompting-based methods for LD generation accuracy even with smaller backbone models. |
PEToolLLM: Towards Personalized Tool Learning in Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing tool learning studies focus on general-purpose tool-use capability, but ignore the importance of personalized tool-user preferences. |
| Approach: | They propose a framework to adapt Large Language Models to personalized tool learning task, which is trained through supervised fine-tuning and direct preference optimization. |
| Outcome: | Extensive experiments on PEToolBench show that the proposed framework outperforms existing LLMs in the personalized tool learning task. |
Filtered Direct Preference Optimization (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on the impact of RLHF on text quality have focused on reward-model-free RL. |
| Approach: | They propose an extension of direct preference optimization to improve model performance by analyzing the quality of the preference dataset. |
| Outcome: | The proposed method improves the performance of models optimized with DPO over those optimized with reward-model-based RLHF. |
SMART: Simulated Students Aligned with Item Response Theory for Question Difficulty Prediction (2025.emnlp-main)
Copied to clipboard
| Challenge: | Traditionally, estimating item difficulties requires real students to respond to items . a cold-start approach cannot be applied to previously unseen items either . |
| Approach: | They propose a method for aligning simulated students with instructed ability to predict difficulty of open-ended items. |
| Outcome: | The proposed method outperforms existing methods on two real-world student responses. |
Towards Universal Debiasing for Language Models-based Tabular Data Generation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing large language models have exacerbated fairness issues in tabular data generation . inherent historical biases in tabulated data cause LLMs to exacerbate fairness problems . |
| Approach: | They propose a universal debiasing framework that minimizes group-level dependencies . it leverages the autoregressive structure and analytic sampling distributions of LLM-based tabular data generators . |
| Outcome: | The proposed framework minimizes group-level dependencies while reducing mutual information between advantaged and protected attributes. |
MASH: Evading Black-Box AI-Generated Text Detectors via Style Humanization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing detection methods rely on white-box assumptions or require prohibitively high computational and interaction costs, rendering them ineffective under practical black-box scenarios. |
| Approach: | They propose a framework that evades black-box detection methods based on style transfer by using style-injection supervised fine-tuning and direct preference optimization to shape distributions of AI-generated texts to resemble those of human-written texts. |
| Outcome: | The proposed framework achieves an average Attack Success Rate (ASR) of 92%, surpassing the strongest baselines by an average of 24% while maintaining superior linguistic quality. |
How Much Would a Clinician Edit This Draft? Evaluating LLM Alignment for Patient Message Response Drafting (2026.acl-long)
Copied to clipboard
Parker Seegmiller, Joseph Gatto, Sarah E. Greer, Ganza Belise Isingizwe, Rohan Ray, Timothy E. Burdick, Sarah Masud Preum
| Challenge: | Large language models (LLMs) have been shown to be effective in drafting patient portal responses, yet their integration into clinical workflows raises various concerns. |
| Approach: | They propose a taxonomy of thematic elements in clinician responses and a framework for assessing clinician editing load of LLM-drafted responses at both content and theme levels. |
| Outcome: | The proposed framework assesses the editing load of LLM-drafted responses at both content and theme levels. |