Papers with instruction-tuning
Improving and Assessing the Fidelity of Large Language Models Alignment to Online Communities (2025.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown promise in representing individuals and communities, but evaluating their fidelity remains a challenge. |
| Approach: | They propose a framework for aligning large language models with online communities via instruction-tuning and comprehensively evaluating alignment across various aspects of language. |
| Outcome: | The proposed framework shows that it can be used to create high-fidelity representations of people and communities. |
OpinionGPT: Modelling Explicit Biases in Instruction-Tuned LLMs (2024.naacl-demo)
Copied to clipboard
| Challenge: | Current research seeks to de-bias such models, or suppress potentially biased answers. |
| Approach: | They present a web demo to test the biases of instruction-tuned Large Language Models . they identify 11 different biase based on a corpus of data . |
| Outcome: | The proposed demo shows that biases in instruction-tuning are explicit and transparent . the demo shows how the model was trained and showcases the web application . |
Comparing Specialised Small and General Large Language Models on Text Classification: 100 Labelled Samples to Achieve Break-Even Performance (2025.emnlp-main)
Copied to clipboard
| Challenge: | specialised small models outperform general large models with few labelled samples (on average 100) performance variance is taken into account when comparing the number of labelled sample required to tune a specialised model with a larger number of samples. |
| Approach: | They find that specialised small models need only few labelled samples to outperform general large models with limited labelled data. |
| Outcome: | The proposed models outperform general large models with few labelled samples and take performance variance into account. |
PROM: Pivoted and Regulated Optimization for Multilingual Instruction Learning (2025.naacl-short)
Copied to clipboard
| Challenge: | Existing solutions to large language models (LLMs) are English-centric, hindering their application to 6500+ existing languages. |
| Approach: | They propose to append English tuning data with its translated pair to solve this problem . they identify English as an internal pivot language and propose to regulate between them . |
| Outcome: | The proposed model is able to generalize on multiple benchmarks across different languages. |
Aligning Instruction Tasks Unlocks Large Language Models as Zero-Shot Relation Extractors (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent work has shown that fine-tuning large language models on large instruction-following datasets improves their performance on a wide range of NLP tasks, but they fail to outperform small LMs on relation extraction (RE), a fundamental information extraction task. |
| Approach: | They propose a framework that aligns RE with question answering (QA), a predominant task in instruction-tuning datasets. |
| Outcome: | The proposed framework outperforms small LLMs on relation extraction (RE), a fundamental information extraction task, by a large margin. |
Where is the answer? An empirical study of positional bias for parametric knowledge extraction in language model (2025.naacl-long)
Copied to clipboard
| Challenge: | Language model (LM) stores diverse factual knowledge in their parameters, which is learned during self-supervised training on unlabeled documents. |
| Approach: | They investigate the issue of "perplexity curse" in the continued training of language model (LM) they find that all studied LMs suffer from positional bias in the training document . |
| Outcome: | The proposed model is able to extract information from multiple questions with diverse queries. |
Datasets and Recipes for Video Temporal Grounding via Reinforcement Learning (2025.emnlp-industry)
Copied to clipboard
Ruizhe Chen, Tianze Luo, Zhiting Fan, Heqing Zou, Zhaopeng Feng, Guiyang Xie, Hansheng Zhang, Zhuochen Wang, Zuozhu Liu, Zhang Huaijian
| Challenge: | Existing methods for video temporal grounding suffer from limited temporal awareness and poor generalization. |
| Approach: | They propose a two-stage training framework that integrates supervised fine-tuning with reinforcement learning to improve both the accuracy and robustness of VTG models. |
| Outcome: | The proposed training framework outperforms existing models on multiple benchmarks on open-domain and challenging scenarios. |
CodeT5+: Open Code Large Language Models for Code Understanding and Generation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing code LLMs adopt a specific architecture or rely on a unified encoder-decoder network for downstream tasks, lacking flexibility to operate in the optimal architecture for a particular task. |
| Approach: | They propose to initialize code LLMs with frozen off-the-shelf LLM and explore instruction-tuning to align with natural language instructions. |
| Outcome: | The proposed model outperforms open-source LLMs on 20 code-related benchmarks. |
QUDSELECT: Selective Decoding for Questions Under Discussion Parsing (2024.emnlp-main)
Copied to clipboard
| Challenge: | Question Under Discussion (QUD) uses implicit questions to reveal discourse relationships between sentences. |
| Approach: | They propose a framework that selectively decodes the QUD dependency structures considering the QUC criteria. |
| Outcome: | The proposed framework outperforms the state-of-the-art baseline models by 9% in human evaluation and 4% in automatic evaluation. |
Mitigating Visual Knowledge Forgetting in MLLM Instruction-tuning via Modality-decoupled Gradient Descent (2025.findings-emnlp)
Copied to clipboard
Junda Wu, Yuxin Xiong, Xintong Li, Yu Xia, Ruoyu Wang, Yu Wang, Tong Yu, Sungchul Kim, Ryan A. Rossi, Lina Yao, Jingbo Shang, Julian McAuley
| Challenge: | Existing fine-tuning and continual learning methods compress visual representations and emphasize task alignment over visual retention. |
| Approach: | They propose a modality-decoupled gradient descent (MDGD) that regulates gradient updates to preserve effective rank of visual features and explicitly disentangles visual learning from task-specific alignment. |
| Outcome: | The proposed model reduces visual forgetting and improves visual retention . it disentangles visual learning from task-specific alignment and preserves effective rank . |
Instructed Language Models with Retrievers Are Powerful Entity Linkers (2023.emnlp-main)
Copied to clipboard
| Challenge: | Generative approaches powered by large language models have demonstrated emergent abilities in tasks that require complex reasoning abilities. |
| Approach: | They propose a sequence-to-sequence training objective with instruction-tuning that enables casual language models to perform entity linking over knowledge bases. |
| Outcome: | The proposed framework outperforms existing approaches with +6.8 F1 points gain on average and huge advantage in training data efficiency and compute consumption. |
Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model (2025.acl-long)
Copied to clipboard
Gregor Geigle, Florian Schneider, Carolin Holtermann, Chris Biemann, Radu Timofte, Anne Lauscher, Goran Glavaš
| Challenge: | Existing models for large vision-language tasks are trained on English data, which makes them struggle to understand non-English input and fail to generate output in the desired target language. |
| Approach: | They conduct multi-stage experiments on 13 vision-language tasks and 43 languages . they find that one can include as many as 100 training languages simultaneously with as little as 25-50% of non-English data . |
| Outcome: | The proposed model outperforms existing models in 14 tasks and 56 languages. |
What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects (2026.findings-eacl)
Copied to clipboard
Naihao Deng, Sheng Zhang, Henghui Zhu, Shuaichen Chang, Jiani Zhang, Alexander Hanbo Li, Chung-Wei Hang, Hideo Kobayashi, Yiqun Hu, Patrick Ng
| Challenge: | a series of paradigm shifts have come with distinct characteristics and challenges associated with table modeling. |
| Approach: | They propose to replicate four table LLMs by instruction-tuning three foundation models on four existing datasets. |
| Outcome: | The results show that base model choice plays a more dominant role than training data itself. |
An Empirical Study of Instruction-tuning Large Language Models in Chinese (2023.findings-emnlp)
Copied to clipboard
| Challenge: | emergence of ChatGPT validates the potential of large language models (LLMs) in artificial general intelligence (AGI) however, the closed source of LLMs coupled with the requirement for massive computing resources has deterred researchers from reaching the LLM training stage. |
| Approach: | They propose to use Chinese instruction-tuning LLMs as a cookbook for customizing LLM models that can better respond to Chinese instructions. |
| Outcome: | The proposed LLM can be used to customize Chinese LLMs that can better respond to Chinese instructions. |
Instruction-tuned Language Models are Better Knowledge Learners (2024.acl-long)
Copied to clipboard
Zhengbao Jiang, Zhiqing Sun, Weijia Shi, Pedro Rodriguez, Chunting Zhou, Graham Neubig, Xi Lin, Wen-tau Yih, Srini Iyer
| Challenge: | Large language models store factual knowledge in parameters, but it can become outdated as the work evolves . pre-instruction-tuning improves ability of LLMs to absorb knowledge from new documents . |
| Approach: | They propose a method that instruction-tunes on questions prior to training on documents . they propose to use QA pairs to update factual knowledge of large language models . |
| Outcome: | The proposed method outperforms instruction-tuning on documents by 17.8%. |
ALPACA AGAINST VICUNA: Using LLMs to Uncover Memorization of LLMs (2025.naacl-long)
Copied to clipboard
Aly M. Kassem, Omar Mahmoud, Niloofar Mireshghallah, Hyunwoo Kim, Yulia Tsvetkov, Yejin Choi, Sherif Saad, Santu Rana
| Challenge: | Existing studies have shown that pre-trained LLMs emit training data up to 150 more often than in regular operation. |
| Approach: | They propose a black-box prompt optimization method where an attacker LLM agent uncovers higher levels of memorization in a victim agent . |
| Outcome: | The proposed method shows 23.7% more overlap with training data compared to state-of-the-art baselines. |
HumanRankEval: Automatic Evaluation of LMs as Conversational Assistants (2024.naacl-long)
Copied to clipboard
| Challenge: | Language models (LMs) are popular conversational assistants, but evaluation of such models is not scalable. |
| Approach: | They propose a task that performs automatic evaluation using human judgement and a large-scale set of questions with multiple answers authored and scored by humans. |
| Outcome: | The proposed task performs well with human judgements and is particularly responsive to model changes following instruction-tuning. |
Getting More from Less: Large Language Models are Good Spontaneous Multilingual Learners (2024.emnlp-main)
Copied to clipboard
Shimao Zhang, Changjiang Gao, Wenhao Zhu, Jiajun Chen, Xin Huang, Xue Han, Junlan Feng, Chao Deng, Shujian Huang
| Challenge: | Large Language Models (LLMs) have shown impressive language capabilities, but most of them have very unbalanced performance across different languages. |
| Approach: | They propose to use question translation data to enhance LLMs' multilingual capabilities by using mechanistic interpretability methods. |
| Outcome: | The proposed method improves multilingual alignment even with unannotated answers in English and a wide range of languages even with instruction-tuned LLMs. |
Instructive Dialogue Summarization with Query Aggregations (2023.emnlp-main)
Copied to clipboard
| Challenge: | Conventional dialogue summarization methods generate summaries without considering user’s specific interests. |
| Approach: | They propose a three-step approach to synthesize high-quality query-based summarization triples by training a unified model on three summarizing datasets with multi-purpose instructive triples. |
| Outcome: | The proposed model outperforms state-of-the-art models and even models with larger sizes on four datasets including dialogue summarization and dialogue reading comprehension. |
LM-Searcher: Cross-domain Neural Architecture Search with LLMs via Unified Numerical Encoding (2025.emnlp-main)
Copied to clipboard
Yuxuan Hu, Jihao Liu, Ke Wang, Jinliang Zheng, Weikang Shi, Manyuan Zhang, Qi Dou, Rui Liu, Aojun Zhou, Hongsheng Li
| Challenge: | Recent advances in Large Language Models have opened new avenues for solving complex optimization problems, including Neural Architecture Search (NAS). |
| Approach: | They propose a framework that leverages LLMs for cross-domain neural architecture optimization without extensive domain-specific tuning. |
| Outcome: | The proposed framework achieves competitive performance in both in-domain and out-of-domain tasks. |
Improving Instruct Models for Free: A Study on Partial Adaptation (2025.emnlp-main)
Copied to clipboard
Ozan Irsoy, Pengxiang Cheng, Jennifer L Chen, Daniel Preotiuc-Pietro, Shiyue Zhang, Duccio Pappadopulo
| Challenge: | Instruct models are deemed superior and more usable but can be eroded by instruction tuning . a recent study shows that instruct models are better at following instructions than base models . |
| Approach: | They scale down the strength of instruction tuning to improve model performance . they show that reducing instruction tuning results in material improvement . |
| Outcome: | The proposed model improves on a few-shot in-context learning benchmark . but it loses some degree of its in-training ability . |
Extending LLMs to New Languages: A Case Study of Llama and Persian Adaptation (2025.coling-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are mainly trained on English data and struggle with low-resource languages. |
| Approach: | They propose to add a new language to Llama to improve classification accuracy for Persian tasks by aligning representations through bilingual pretraining and instruction datasets. |
| Outcome: | The proposed model performs on generation and classification tasks with no adverse impact and sometimes even improvements on English tasks. |
An Intent-based and Annotation-free Method for Duplicate Question Detection in CQA Forums (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for detecting duplicate questions in CQA rely on generic text-pair matching models, overlooking the intent behind the questions. |
| Approach: | They propose a new intent-based duplication detector that leverages intent information to address the problem of duplicate question detection in CQA. |
| Outcome: | The proposed detector leverages the characteristics of CQA forums and extracts training labels to recognize and match intents without human annotation. |
Filter-then-Generate: Large Language Models with Structure-Text Adapter for Knowledge Graph Completion (2025.coling-main)
Copied to clipboard
| Challenge: | Empirical evidence suggests that LLMs perform worse than conventional KGC approaches. |
| Approach: | They propose a filter-then-generate paradigm and a multiple-choice question format to harness the capability of LLMs while mitigating the issue casused by hallucinations. |
| Outcome: | The proposed method achieves substantial performance gain compared to existing state-of-the-art methods. |
AdaLomo: Low-memory Optimization with Adaptive Learning Rate (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models require substantial memory for training, thereby setting a high hardware threshold. |
| Approach: | They propose a low-memory optimization technique that reduces memory footprint . they propose an adaptive learning rate for each parameter and a grouped update normalization to stabilize convergence . |
| Outcome: | The proposed low-memory optimization performs better than the prevailing algorithm for large language models, AdamW. |
Automatic Expert Discovery in LLM Upcycling via Sparse Interpolated Mixture-of-Experts (2025.acl-long)
Copied to clipboard
| Challenge: | Sparse Interpolated Mixture-of-Experts (SIMoE) instruction-tuning is an end-to-end algorithm designed to fine-tune a dense pre-trained Large Language Model (LLM) into a MoE-style model that possesses capabilities in multiple specialized domains. |
| Approach: | They propose an algorithm to fine-tune a dense pre-trained Large Language Model into a MoE-style model that possesses capabilities in multiple specialized domains. |
| Outcome: | The proposed algorithm achieves state-of-the-art on common instruction-tuning benchmarks while maintaining an optimal performance-compute trade-off compared to baselines. |
Multi-Task Transfer Matters During Instruction-Tuning (2024.findings-acl)
Copied to clipboard
| Challenge: | Instruction-tuning improves a model’s ability to learn in-context, but the mechanisms that drive in-constext learning are poorly understood. |
| Approach: | They propose to train a model on hundreds of tasks to improve its ability to learn in-context. |
| Outcome: | The proposed methods improve model transfer and in-context generalization, suggesting catastrophic forgetting may impact in-constext learning. |
Membership and Memorization in LLM Knowledge Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in Knowledge Distillation (KD) aim to mitigate the high computational demands of Large Language Models (LLMs). |
| Approach: | They characterize and investigate membership privacy risks inherent in six LLM KD techniques . they use instruction-tuning settings that span seven NLP tasks and three teacher model families and various size student models to examine the extent of privacy risks. |
| Outcome: | The proposed methods carry membership and memorization privacy risks from the teacher to students, but differ across different techniques. |
Multi-Objective Forward Reasoning and Multi-Reward Backward Refinement for Product Review Summarization (2024.lrec-main)
Copied to clipboard
| Challenge: | Product review summarization aims to generate a concise summary based on product reviews . factual accuracy, aspect comprehensiveness, and content relevance are challenges . |
| Approach: | They propose an FB-Thinker framework to improve product review summarization ability . they propose two Chinese product review summary datasets for instruction-tuning and evaluation . |
| Outcome: | The proposed framework improves product review summarization with forward reasoning and backward refinement. |
Optimizing Language Augmentation for Multilingual Large Language Models: A Case Study on Korean (2024.lrec-main)
Copied to clipboard
ChangSu Choi, Yongbin Jeong, Seoyoon Park, Inho Won, HyeonSeok Lim, SangMin Kim, Yejee Kang, Chanhyuk Yoon, Jaewan Park, Yiseul Lee, HyeJin Lee, Younggyun Hahm, Hansaem Kim, KyungTae Lim
| Challenge: | Large language models (LLMs) use pretraining to predict the subsequent word, but less-resourced languages are being overlooked. |
| Approach: | They propose to expand the MLLM vocabularies to enhance expressiveness and use bilingual data for pretraining to align the high- and less-resourced languages. |
| Outcome: | The proposed model outperforms existing models in qualitative analyses compared to Korean monolingual models. |
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. |
Exploring the Impact of Instruction-Tuning on LLM’s Susceptibility to Misinformation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing studies highlight that large language models are receptive to external information that contradicts their parametric knowledge, but little research has been conducted on the direct impact of instruction-tuning on this phenomenon. |
| Approach: | They examine how instruction-tuning influences LLMs' susceptibility to misinformation, particularly in knowledge conflict situations. |
| Outcome: | The proposed model is more user-oriented and more likely to accept misinformation when it is presented by the user. |