Papers with alignment strategies
SEA: Supervised Embedding Alignment for Token-Level Visual-Textual Integration in MLLMs (2025.emnlp-main)
Copied to clipboard
Yuanyang Yin, Yaqi Zhao, Yajie Zhang, Yuanxing Zhang, Ke Lin, Jiahao Wang, Xin Tao, Pengfei Wan, Wentao Zhang, Feng Zhao
| Challenge: | Multimodal Large Language Models (MLLMs) integrate visual and textual inputs, yet modality alignment remains one of the most challenging aspects. |
| Approach: | They propose a token-level supervision alignment method that enables more precise visual-text alignment during pretraining. |
| Outcome: | The proposed method improves performance across various model sizes, with smaller models benefiting the most. |
Foot-In-The-Door: A Multi-turn Jailbreak for LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly integrated into real-world applications, requiring a high level of safety and alignment. |
| Approach: | They propose a multi-turn jailbreak method that leverages foot-in-the-door principles to escalate malicious intent of user queries through intermediate bridge prompts and aligns the model’s response by itself to induce toxic responses. |
| Outcome: | The proposed method achieves an average attack success rate of 94% across seven widely used models outperforming existing state-of-the-art methods. |
Semantically Inspired AMR Alignment for the Portuguese Language (2020.emnlp-main)
Copied to clipboard
| Challenge: | Abstract Meaning Representation (AMR) parsers require alignment between nodes and words of the sentence. |
| Approach: | They propose to use a more semantically matched word-concept pair to align graphs with words in Portuguese . they performed intrinsic and extrinsic evaluations and found it outperforms the English alignment strategies. |
| Outcome: | The proposed method outperforms the existing methods for English and achieves competitive results with a parser designed for the Portuguese language. |
Aligning Multilingual Embeddings for Improved Code-switched Natural Language Understanding (2022.coling-1)
Copied to clipboard
| Challenge: | a recent study has shown that multilingual models can be effective on monolingual data but need additional training to work well with code-switched text. |
| Approach: | They propose to train multilingual models with alignment objectives using parallel text . they find such an explicit alignment step improves performance on code-switched NLP tasks . |
| Outcome: | The proposed model improves on Hindi-English Sentiment Analysis, Named Entity Recognition and Question Answering tasks. |
Uncertainty Aware Learning for Language Model Alignment (2024.acl-long)
Copied to clipboard
| Challenge: | Existing alignment strategies that focus on diverse and high-quality data often overlook the intrinsic uncertainty of tasks, learning all data samples equally. |
| Approach: | They propose to introduce the sample uncertainty into the alignment of different task scenarios by a simple fashion by setting the label smoothing value of training according to the uncertainty of individual samples. |
| Outcome: | The proposed model outperforms standard supervised fine-tuning on high-entropy tasks and complex low-entropic tasks. |
SemanticCamo: Jailbreaking Large Language Models through Semantic Camouflage (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have made safety issues of LLMs more prominent and critical. |
| Approach: | They propose a framework which attacks LLMs through semantic camouflage and replaces unsafe content with semantic features to conceal malicious intent . |
| Outcome: | The proposed framework outperforms existing models in over 80% of cases and is highly effective against various defenses. |
In-Context Representation Hijacking (2026.acl-long)
Copied to clipboard
| Challenge: | In-context representation hijacking attacks on large language models are insufficient to prevent harm, yet the mechanisms underlying this behavior remain unclear. |
| Approach: | They propose a simple in-context representation hijacking attack that replaces a harmful keyword with a benign token, and embeds the harmful semantics under a euphemism. |
| Outcome: | The proposed attack is optimized for open-source and achieves strong success rates on closed-source systems, reaching 74% on Llama-3.3-70B-Instruct with a single-sentence context override. |
AED-RAG: Continuous Multi-Granular Context Fusion for Retrieval-Augmented Generation via Adaptive Ensemble Decoding (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing alignment strategies that rely on discrete reranking struggle to address this granularity mismatch or effectively balance external evidence with internal knowledge. |
| Approach: | They propose a framework that synergizes discrete retrieval with continuous reranking to discern the information density differences between unstructured narrative passages and structured knowledge triplets. |
| Outcome: | Extensive experiments on four open-domain QA benchmarks show that AED-RAG significantly outperforms competitive baselines. |
The Behavior Gap: Evaluating Zero-shot LLM Agents in Complex Task-Oriented Dialogs (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent studies show that LLM-based agents struggle to perform in zero-shot scenarios. |
| Approach: | They propose a framework to quantify the behavior gap between AI agents and human experts . they propose to examine discrepancies in dialog acts, tool usage, and knowledge utilization . |
| Outcome: | The proposed framework measures the behavior gap between AI agents and human experts on task-oriented dialogs. |
DeAL: Decoding-time Alignment for Large Language Models (2025.acl-long)
Copied to clipboard
James Y. Huang, Sailik Sengupta, Daniele Bonadiman, Yi-An Lai, Arshit Gupta, Nikolaos Pappas, Saab Mansour, Katrin Kirchhoff, Dan Roth
| Challenge: | Large Language Models (LLMs) are expected to generate content aligned with human preferences. |
| Approach: | They propose a framework that allows the user to customize reward functions and enables Decoding-time Alignment of LLMs (DeAL). |
| Outcome: | The proposed framework allows the user to customize reward functions and enables Decoding-time Alignment of LLMs. |
Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models are increasingly used as conversational agents that adopt personas and role-play characters at user request. |
| Approach: | They propose to examine how persona agreeableness influences sycophancy across 13 small, open-weight language models ranging from 0.6B to 20B parameters. |
| Outcome: | The proposed model consists of 275 personas and exposes them to 4,950 sycophancy-eliciting prompts spanning 33 topic categories. |
Probing Audio-Visual Reasoning in Multimodal Language Models through the Lens of Audio (2026.acl-long)
Copied to clipboard
Kaixiong Gong, Kaituo Feng, Bohao Li, Yibing Wang, Mofan Cheng, Shijia Yang, Jiaming Han, Benyou Wang, Yutong Bai, Zhuoran Yang, Xiangyu Yue
| Challenge: | Recent multimodal large language models lack robust audio-visual integration ability and performance on DeafTest is highly correlated with AV-Odyssey accuracy. |
| Approach: | They propose a benchmarking tool that integrates audio-visual reasoning with audio-video cues to infer solutions. |
| Outcome: | The proposed model performs well on DeafTest, but lacks audio perception in simple audio tasks. |