Papers with Aligning Large Language Models
Permutative Preference Alignment from Listwise Ranking of Human Judgments (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to align Large Language Models with human preferences are based on the Bradley-Terry model, but when multiple responses are available, the B-T model fails to guarantee an accurate list ranking of the responses. |
| Approach: | They propose an offline listwise approach that incorporates the Normalized Discounted Cumulative Gain (NDCG) as an alternative training objective for LLM alignment. |
| Outcome: | The proposed approach outperforms existing pairwise and listwise methods on evaluation sets and general benchmarks such as AlpacaEval. |
MPPO: Multi Pair-wise Preference Optimization for LLMs with Arbitrary Negative Samples (2025.coling-main)
Copied to clipboard
Shuo Xie, Fangzhi Zhu, Jiahui Wang, Lulu Wen, Wei Dai, Xiaowei Chen, Junxiong Zhu, Kai Zhou, Bo Zheng
| Challenge: | Existing preference optimization methods such as DPO and KTO are inherently derived from PPO, requiring a reference model that adds GPU memory resources and relies heavily on abundant preference data. |
| Approach: | They propose an algorithm that leverages the average likelihood of model responses to fit the reward function and maximizes the utilization of preference data. |
| Outcome: | The proposed algorithm outperforms DPO, ORPO, and SimPO on MT-Bench and Arena-Hard. |
Offline Preference Optimization via Maximum Marginal Likelihood Estimation (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing approaches to align Large Language Models with human preferences are complex and unstable. |
| Approach: | They propose a new approach that maximizes the marginal log-likelihood of a preferred text output by using the preference pair as samples for approximation. |
| Outcome: | The proposed approach maximizes the marginal log-likelihood of a preferred text output, using the preference pair as samples for approximation, and forgoes the need for both an explicit reward model and entropy maximization. |
Disentangling Preference Representation and Text Generation for Efficient Individual Preference Alignment (2025.coling-main)
Copied to clipboard
| Challenge: | Human values are inherently diverse, making it insufficient to align LLMs solely with general preferences. |
| Approach: | They propose a flexible paradigm for individual preference alignment that disentangles preference representation from text generation in LLMs. |
| Outcome: | The proposed method produces aligned quality and better than PEFT-based methods while reducing training time for each new individual preference by 80% to 90%. |
Teaching Language Models to Self-Improve by Learning from Language Feedback (2024.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models (LLMs) generate content that can be untruthful or harmful. |
| Approach: | They propose a method that leverages model feedback for alignment . they use a base language model to generate initial responses, critiqued and refined . |
| Outcome: | The proposed method outperforms strong baselines across diverse tasks and model sizes. |
M2PO: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation (2026.acl-long)
Copied to clipboard
Hao Wang, Linlong Xu, Heng Liu, Yangyang Liu, Xiaohu Zhao, Bo Zeng, Liangying Shao, Yichen Dong, Xinwei Wu, Jiang Zhou, Tianyu Dong, Xiangxiang Zeng, Longyue Wang, Weihua Luo
| Challenge: | prevailing methods for machine translation are often hindered by misleading reward signals. |
| Approach: | They propose a framework that aligns large language models to human preferences . they propose 'M2PO' to correct the bias towards partial errors . |
| Outcome: | The proposed framework outperforms open-source models and achieves parity with proprietary models. |
InvestAlign: Overcoming Data Scarcity in Aligning Large Language Models with Investor Decision-Making Processes Under Herd Behavior (2025.acl-long)
Copied to clipboard
| Challenge: | relying on authentic data for Supervised Fine-Tuning (SFT) is costly and expensive. |
| Approach: | They propose a framework that constructs high-quality SFT datasets by leveraging theoretical solutions to similar and simple optimal investment problems rather than the complex scenarios. |
| Outcome: | The proposed framework achieves faster parameter convergence than using real-user data, suggesting superior learning efficiency. |
Reason-KE++: Aligning the Process, Not Just the Outcome, for Faithful LLM Knowledge Editing (2026.findings-acl)
Copied to clipboard
| Challenge: | Current methods for modifying parameters to integrate new knowledge are not accurate enough. |
| Approach: | They propose an SFT+RL framework that instills process-level faithfulness by a stage-aware Reward mechanism and a Stage-assisted Reward Mechanism. |
| Outcome: | The proposed framework instills process-level faithfulness while boosting final accuracy. |
Dynamic Rewarding with Prompt Optimization Enables Tuning-free Self-Alignment of Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Empirical evaluations on eight recent LLMs reveal that DRPO significantly enhances alignment performance, enabling base models to outperform their SFT/RLHF-tuned counterparts. |
| Approach: | They propose a tuning-free approach to self-alignment called Dynamic Rewarding with Prompt Optimization (DRPO) it leverages a dynamic rewarding mechanism to identify and rectify alignment weaknesses . |
| Outcome: | The proposed approach outperforms existing methods and is highly adaptable to various alignment challenges. |
Internal Value Alignment in Large Language Models through Controlled Value Vector Activation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing LLMs do not possess consistent values, but many have been developed to align them at the behavioral level, including supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). |
| Approach: | They propose a Controlled Value Vector Activation method that directly aligns the internal values of Large Language Models by interpreting how a value is encoded in their latent representations. |
| Outcome: | The proposed method achieves highest success rate across 10 basic values without hurting model performance and fluency, and ensures target values even with opposite and potentially malicious input prompts. |
Activation Reward Models for Few-Shot Model Alignment (2026.findings-acl)
Copied to clipboard
Tianning Chai, Chancharik Mitra, Brandon Huang, Gautam Rajendrakumar Gare, Zhiqiu Lin, Assaf Arbelle, Leonid Karlinsky, Rogerio Feris, Trevor Darrell, Deva Ramanan, Roei Herzig
| Challenge: | A common approach is to use reward models that enable reinforcement-learning post-training. |
| Approach: | They propose a method that steers LLM activations to align with few-shot preference data without finetuning. |
| Outcome: | The proposed method surpasses zero-shot, few-shot and voting-based benchmarks on reward hacking and noise signals. |
ProMedical: Hierarchical Fine-Grained Criteria Modeling for Medical LLM Alignment via Explicit Injection (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are difficult to align with high-stakes medical standards due to dissonance between coarse-grained preference signals and complex protocols. |
| Approach: | They propose a framework that aligns Large Language Models with medical standards . they use a dataset generated via a human-in-the-loop pipeline to augment medical instructions . |
| Outcome: | The proposed framework disentangles safety constraints from general proficiency, enabling precise guidance during reinforcement learning. |
MEAV: Model Editing with Alignment Vectors for inference time LLM alignment in single and multidomain preference spectrum (2026.findings-acl)
Copied to clipboard
Sadat Shahriar, Zheng Qi, Nikolaos Pappas, Srikanth Doss, Kishaloy Halder, Monica Sunkara, Manuel Mager, Yassine Benajiba
| Challenge: | Existing training-time alignment methods require full retraining when a change is needed. |
| Approach: | They propose an inference-time model-editing-based alignment method that learns encoded representations of preference dimensions and allows dynamic adjusting of the model behavior. |
| Outcome: | The proposed method can be used to align large language models to human preferences . it reduces the cost of inference by half compared to the prompt engineering approach . |
Self-Guided Alignment: Adaptive Preference Sensing for Multi-Objective Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to align LLMs with diverse human values rely on ground-truth scores . existing approaches implicitly approximate an average-user preference, thereby failing to capture heterogeneity of human values or accommodate conflicting user needs. |
| Approach: | They propose a framework that transforms passive reward dependency into an intrinsic adaptive sensing capability. |
| Outcome: | The proposed framework outperforms state-of-the-art models in multiple model scales and improves preference alignment. |