Papers with alignment techniques
Explicit vs. Implicit: Investigating Social Bias in Large Language Models through Self-Reflection (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to quantify and quantify social biases in Large Language Models (LLMs) focus on explicit bias, with little attention to implicit bias. |
| Approach: | They propose a self-reflection-based evaluation framework that measures implicit bias and evaluates explicit bias by prompting LLMs to analyze their own generated content. |
| Outcome: | The proposed framework compares explicit and implicit biases in large language models . it demonstrates that explicit bias manifests as mild stereotypes, while implicit bias exhibits strong stereotypes. |
Reformatted Alignment (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Current methods to improve data quality are labor-intensive or prone to factual errors caused by LLM hallucinations. |
| Approach: | They propose a method which reformats the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence. |
| Outcome: | The proposed approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques. |
Aligning Black-Box LLMs for Aspect Sentiment Quad Prediction (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models lack task-specific alignment with ASQP . supervised small language models (SLMs) lack the extensive knowledge of LLMs. |
| Approach: | They propose a framework that combines large language models and small language models to align LLM outputs with human preferences. |
| Outcome: | The proposed framework improves Aspect Sentiment Quad Prediction performance by combining SLMs and LLMs. |
Ensuring Safe and High-Quality Outputs: A Guideline Library Approach for Language Models (2024.naacl-long)
Copied to clipboard
Yi Luo, Zhenghao Lin, YuHao Zhang, Jiashuo Sun, Chen Lin, Chengjin Xu, Xiangdong Su, Yelong Shen, Jian Guo, Yeyun Gong
| Challenge: | Guide-Align is a guideline-oriented approach to augment the safety and quality of Large Language Models. |
| Approach: | They propose a guideline-oriented method to augment the safety and quality of large language models. |
| Outcome: | The proposed method outperforms existing methods on three benchmarks and shows significant improvements in security and quality. |
Wait, that’s not an option: LLMs Robustness with Incorrect Multiple-Choice Options (2025.acl-long)
Copied to clipboard
| Challenge: | Using a framework that combines instruction-following with critical reasoning, we show that the ability of LLMs to override defaults when faced with invalid options is impaired by alignment techniques. |
| Approach: | They propose a framework for evaluating LLMs’ capacity to balance instruction-following with critical reasoning when presented with multiple-choice questions containing no valid answers. |
| Outcome: | The proposed framework improves models' ability to override defaults when faced with invalid options while minimizing the impact of model size and training techniques on the model. |
Learning to Explore and Select for Coverage-Conditioned Retrieval-Augmented Generation (2025.findings-naacl)
Copied to clipboard
| Challenge: | Recent advances of large language models (LLMs) have enabled them to provide long and detailed responses by leveraging their parametric knowledge. |
| Approach: | They introduce a dataset of 10K hierarchical sets of information-seeking subqueries that define structured boundaries for outline creation and evaluation in C2 scenarios. |
| Outcome: | The proposed model generates higher-quality outlines that better fulfill diverse user needs when trained with alignment techniques like DPO. |
Controllable Preference Optimization: Toward Controllable Multi-Objective Alignment (2024.emnlp-main)
Copied to clipboard
Yiju Guo, Ganqu Cui, Lifan Yuan, Ning Ding, Zexu Sun, Bowen Sun, Huimin Chen, Ruobing Xie, Jie Zhou, Yankai Lin, Zhiyuan Liu, Maosong Sun
| Challenge: | Existing algorithms for achieving optimal alignment are mostly unidirectional . a recent study suggests that large language models can be ground with evident preferences . |
| Approach: | They propose to ground large language models with evident preferences . they propose to use controllable preference optimization to specify different objectives . |
| Outcome: | The proposed models can provide responses that match various preferences among the ”3H” desiderata. |
Towards Proactive Personalization through Profile Customization for Individual Users in Dialogues (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing alignment methods focus on universal human values or static, single-turn preferences, thereby failing to address the critical needs of long-term personalization and the initial user cold-start problem. |
| Approach: | They propose a user-centric lifelong agent that continuously infers and adapts to user preferences. |
| Outcome: | The proposed agent achieves superior performance over strong prompt-based and policy optimization baselines, not only in idealized but also in noisy conversational contexts. |
Self-Evolution Fine-Tuning for Policy Optimization (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent years have showcased the remarkable capabilities and performance of large language models (LLMs) across a broad range of tasks. |
| Approach: | They propose supervised fine-tuning (SEFT) for LLM alignment to eliminate the need for annotated samples while retaining the stability and efficiency of SFT. |
| Outcome: | The proposed method eliminates the need for annotated samples while maintaining the stability and efficiency of SFT. |
A Survey on Personalized Alignment—The Missing Piece for Large Language Models in Real-World Applications (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their transition to real-world applications reveals a critical limitation: the inability to adapt to individual preferences while maintaining alignment with universal human values. |
| Approach: | They propose a framework that enables LLMs to adapt their behavior within ethical boundaries based on individual preferences. |
| Outcome: | The proposed framework analyzes implementation approaches and evaluates their effectiveness across various scenarios. |
Translate & Fill: Improving Zero-Shot Multilingual Semantic Parsing with Synthetic Data (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to improve semantic parsing performance on target languages are limited. |
| Approach: | They propose a Translate-and-Fill method that produces silver training data for a multilingual semantic parser. |
| Outcome: | The proposed method produces silver training data for a multilingual parser. |
Exploiting the Index Gradients for Optimization-Based Jailbreaking on Large Language Models (2025.coling-main)
Copied to clipboard
| Challenge: | Despite advances in training Large Language Models, they remain vulnerable to jailbreak, an adversarial attack method. |
| Approach: | They propose an adversarial jailbreak algorithm that exploits the gradient information of the suffix tokens to accelerate the optimization process. |
| Outcome: | The proposed model achieves 1.5x speedup while maintaining high attack success rates. |
MetaAlign: Align Large Language Models with Diverse Preferences during Inference Time (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to align large language models with human preferences often result in a static alignment that cannot account for the diversity of human preferences in practical applications. |
| Approach: | They propose a method to help large language models dynamically align with various explicit or implicit preferences specified at inference time. |
| Outcome: | The proposed method can help LLMs dynamically align with various explicit or implicit preferences specified at the inference stage, validating the feasibility of MetaAlign. |
From Distributional to Overton Pluralism: Investigating Large Language Model Alignment (2025.naacl-long)
Copied to clipboard
| Challenge: | a large language model's (LLM) output distribution is changed by an alignment process . a recent study shows that aligned models surface information that cannot be recovered from base models without fine-tuning. |
| Approach: | They analyze two aspects of the alignment process that change output distributions . they find alignment suppresses irrelevant and unhelpful content . |
| Outcome: | The proposed model can be imitated without fine-tuning by using in-context examples and lower-resolution semantic hints about response content. |
MT-Bench-101: A Fine-Grained Benchmark for Evaluating Large Language Models in Multi-Turn Dialogues (2024.acl-long)
Copied to clipboard
Ge Bai, Jie Liu, Xingyuan Bu, Yancheng He, Jiaheng Liu, Zhanhui Zhou, Zhuoran Lin, Wenbo Su, Tiezheng Ge, Bo Zheng, Wanli Ouyang
| Challenge: | Large Language Models (LLMs) have greatly enhanced dialogue systems, but evaluation of their capabilities remains a challenge. |
| Approach: | They propose a model to evaluate the fine-grained abilities of Large Language Models in multi-turn dialogues. |
| Outcome: | The proposed model evaluates 21 popular chatbots based on MT-Bench-101 . it includes 3 overarching abilities and 13 distinct tasks within multi-turn dialogue scenarios. |
Deconstructing In-Context Learning: Understanding Prompts via Corruption (2024.lrec-main)
Copied to clipboard
| Challenge: | Prior work examined how modifying different elements of the prompt can affect model performance, but this limited number of elements made replication challenging. |
| Approach: | They decompose the entire prompt into four components: task description, demonstration inputs, labels, and inline instructions provided for each demonstration. |
| Outcome: | The proposed model is robust to minor prompt modifications, but its underlying pre-trained backbone is brittle . previous studies focused on models with fewer than 15 billion parameters or exclusively examined black-box models like GPT-3 or PaLM, making replication challenging. |
Aligning to What? Limits to RLHF Based Alignment (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing studies on RLHF and covert and overt biases in large language models are unclear . et al. analyzed off-the-shelf language models to evaluate their overt and cover racial biase . |
| Approach: | They evaluate the relationship between reinforcement learning from human feedback and biases in large language models. |
| Outcome: | The proposed approach can be used to mitigat covert biases, the authors show . they found that the RLHF approach calcifies model biase . |
FaST: Feature-aware Sampling and Tuning for Personalized Preference Alignment with Limited Data (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have focused on personalizing conversational assistants to meet specific user preferences. |
| Approach: | They propose to use a dataset to analyze a problem where only a small set of preference annotations can be collected per user. |
| Outcome: | The proposed approach leverages high-level features discovered from the data, achieving the best overall performance. |
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. |
The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm (2024.emnlp-main)
Copied to clipboard
null Aakanksha, Arash Ahmadian, Beyza Ermis, Seraphina Goldfarb-Tarrant, Julia Kreutzer, Marzieh Fadaee, Sara Hooker
| Challenge: | Existing approaches to safety alignment focus on homogeneous monolingual settings . preference training and safety measures often overfit to harms common in Western-centric datasets . |
| Approach: | They propose to use human annotated red teaming prompts to identify global and local harms. |
| Outcome: | The proposed approach can address and optimize for a non-homogeneous set of languages and cultural preferences while minimizing both global and local harms. |
HealthAlignSumm : Utilizing Alignment for Multimodal Summarization of Code-Mixed Healthcare Dialogues (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Collaboration between doctors and AI scientists is leading to personalized models to stream-line healthcare tasks and improve productivity. |
| Approach: | They propose to use alignment techniques to combine a doctor-patient dialogue with a visual component of the BART model. |
| Outcome: | The proposed model in-tegrates visual components with the BART ar-chitecture. |
Evaluating Transformers for OCR Post-Correction in Early Modern Dutch Theatre (2025.coling-main)
Copied to clipboard
| Challenge: | a new study examines the effectiveness of two types of transformer models for OCR post-correction in early modern Dutch plays. |
| Approach: | They propose to use large generative models and sequence-to-sequence models for OCR post-correction in early modern Dutch plays. |
| Outcome: | The proposed model outperforms generative models on the OCR post-correction task . the model outpersforms the model with the lowest error rate on the historical English dataset . |
RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Preference optimization is a widely adopted post-training technique to align large language models with human preferences. |
| Approach: | They propose a method for generating multilingual feedback data to balance data coverage. |
| Outcome: | The proposed method achieves 54.4% win-rate against current state-of-the-art multilingual LLM in its parameter class and 69.5% win- rate or higher against widely used models like Gemma, Mistral and Llama 3. |
Margin Matching Preference Optimization: Enhanced Model Alignment with Granular Feedback (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for large language models rely on binary labels that fail to capture the subtle differences in relative quality between pairs. |
| Approach: | They propose a method that incorporates relative quality margins into optimization to improve LLM policies and reward models. |
| Outcome: | The proposed approach outperforms baseline methods on popular benchmarks including MT-bench and RewardBench. |
Controlling Distributional Bias in Multi-Round LLM Generation via KL-Optimized Fine-Tuning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing evaluation methods focus on single-round inference, but this view is problematic in real-world applications. |
| Approach: | They propose a framework that couples Steering Token Calibration with Semantic Alignment to ensure that LLMs are correctly aligned across gender, race, and sentiment. |
| Outcome: | The proposed framework outperforms baseline methods in achieving precise distributional control in attribute generation tasks. |
Can We Entrust Justice to AI?: How Persona Traps Contaminate Reasoning in Criminal Investigation (2026.findings-acl)
Copied to clipboard
| Challenge: | a study injected personas into neutralized murder mystery scenarios to examine reasoning stability of large language models . adrian s. gupta: if models are used to evaluate suspects, are they free from the trap of implicit bias? he says current alignment techniques focus on identity-based bias while neglecting relationship-based ones . |
| Approach: | a study systematically injected personas into neutralized murder mystery scenarios . it found that implicit bias propagation was observed across all models . authors propose stability evaluation should encompass outputs and reasoning processes . |
| Outcome: | The proposed pipeline can analyze evidence and evaluate suspects in murder mysteries . it shows that models outwardly state "that information is irrelevant to the judgment" the proposed pipeline could be extended to include reasoning processes, authors say . |
Path Drift in Large Reasoning Models: How First-Person Commitments Override Safety (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies on prompt injection and jailbreak attacks primarily target the surface structure of input prompts. |
| Approach: | They propose a three-stage approach to mitigate the risk of Long-CoT reasoning drift . they propose 'path-level defense' strategy that incorporates role attribution correction and metacognitive reflection . |
| Outcome: | The proposed framework reduces refusal rates and ethical evaporation, while ethical escalation and layered disclaimers progressively steer models toward unsafe completions. |
Benchmarking Multi-National Value Alignment for Large Language Models (2025.findings-acl)
Copied to clipboard
Chengyi Ju, Weijie Shi, Chengzhong Liu, Jiaming Ji, Jipeng Zhang, Ruiyuan Zhang, Jiajie Xu, Yaodong Yang, Sirui Han, Yike Guo
| Challenge: | Existing studies on large language models focus on ethical reviews, failing to capture the diversity of national values. |
| Approach: | They propose a national value extraction pipeline to efficiently construct value assessment datasets and a model-based model with instruction tagging to process raw data sources. |
| Outcome: | The proposed benchmark evaluates the alignment of LLMs with the values of five major nations: China, the United States, the UK, France, and Germany. |
Prompting Large Language Models for Counterfactual Generation: An Empirical Study (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have made remarkable progress in a wide range of natural language understanding and generation tasks, but their ability to generate counterfactuals has not been examined systematically. |
| Approach: | They propose a framework to evaluate LLMs' ability to generate counterfactuals based on key factors including intrinsic properties and prompt design. |
| Outcome: | The proposed framework examines the strengths and weaknesses of large language models (LLMs) and identifies factors that influence their ability to generate counterfactuals. |
Reasoning Hijacking: The Fragility of Reasoning Alignment in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Current LLM safety research focuses on mitigating **Goal Hijacking**, preventing attackers from redirecting a model’s high-level objective. |
| Approach: | They propose a new adversarial prompt attack paradigm that subverts model judgments by injecting spurious decision criteria without altering the high-level task goal. |
| Outcome: | The proposed model subverts model judgments by injecting spurious decision criteria without altering the high-level task goal. |
Mitigating Over-Refusal in Aligned Large Language Models via Inference-Time Activation Energy (2026.acl-long)
Copied to clipboard
Eric Hanchen Jiang, Weixuan Ou, Run Liu, Shengyuan Pang, Guancheng Wan, Ranjie Duan, Wei Dong, Kai-Wei Chang, XiaoFeng Wang, Ying Nian Wu, Xinfeng Li
| Challenge: | Existing safety alignment techniques prioritize mitigating harmful responses at the expense of overcautious behavior, leading models to incorrectly refuse benign requests. |
| Approach: | They propose a fine-tuning free framework to improve safety and reduce false refusals by dynamic, inference-time intervention. |
| Outcome: | The proposed framework raises compliance on the ORB-H benchmark from 57.3% to 82.6% while maintaining the baseline safety performance. |
Towards Unified Multimodal Large Language Models: A survey (2026.findings-acl)
Copied to clipboard
| Challenge: | unified multimodal large language models (MLLMs) are emerging but lack a systematic framework to connect them and situate current trends within a broader landscape. |
| Approach: | They present a systematic review of unified Multimodal Large Language Models . they outline the foundational concepts and prerequisites for understanding them . |
| Outcome: | The present review provides a systematic and systematic overview of unified MLLMs . it discusses persistent challenges and identify promising directions for future research . |
PO-KGQA: Preference Optimization for Low-Resource Complex Knowledge Graph Question Answering (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing low-resource in-context learning-based knowledge graph question answering methods rely heavily on large language models to convert natural language questions into logical forms. |
| Approach: | They propose a low-resource in-context learning-based knowledge graph question answering (KGQA) that uses large language models to convert a natural language question into its corresponding logical form. |
| Outcome: | The proposed method outperforms other methods on complex benchmarks by approximately 9% (avg). |