Papers with continual pretraining
HYPEROFA: Expanding LLM Vocabulary to New Languages via Hypernetwork-Based Embedding Initialization (2025.acl-srw)
Copied to clipboard
| Challenge: | Pre-trained language models exhibit suboptimal performance on mid- and low-resource languages due to limited exposure to these languages during pre-training. |
| Approach: | They propose a similarity-based subword embedding initialization heuristic that introduces new tokens specific to target languages, initializes their embedders, and applies continual pre-training on target-language data. |
| Outcome: | The proposed method outperforms random initialization baseline and matches or exceeds OFA in both continual pre-training convergence and downstream task performance. |
Continual Pretraining on Encrypted Synthetic Data for Privacy-Preserving LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to protect PII from training on small corpora are difficult to implement in real-world applications. |
| Approach: | They propose an entity-based framework that synthesizes encrypted training data to protect PII. |
| Outcome: | The proposed framework outperforms base models and ensures PII security on limited-scale datasets while exhibiting a modest performance gap compared to models trained on unencrypted synthetic data. |
ELO: Efficient Layer-Specific Optimization for Continual Pretraining of Multilingual LLMs (2026.eacl-industry)
Copied to clipboard
Hangyeol Yoo, ChangSu Choi, Minjun Kim, Seohyun Song, SeungWoo Song, Inho Won, Jongyoul Park, Cheoneum Park, KyungTae Lim
| Challenge: | Recent studies have focused on enhancing multilingual large language models (MLLMs) for specific languages. |
| Approach: | They propose an efficient layer-specific optimization method to enhance continual pretraining (CP) for specific languages in multilingual large language models (MLLMs). |
| Outcome: | The proposed method achieves a training speedup of up to 6.46 times compared to existing methods while improving target language performance by up to 5.2% on qualitative benchmarks. |
Aurora-M: Open Source Continual Pre-training for Multilingual Language and Code (2025.coling-industry)
Copied to clipboard
Taishi Nakamura, Mayank Mishra, Simone Tedeschi, Yekun Chai, Jason T. Stillerman, Felix Friedrich, Prateek Yadav, Tanmay Laud, Vu Minh Chien, Terry Yue Zhuo, Diganta Misra, Ben Bogin, Xuan-Son Vu, Marzena Karpinska, Arnav Varma Dantuluri, Wojciech Kusa, Tommaso Furlanello, Rio Yokota, Niklas Muennighoff, Suhas Pai, Tosin Adewumi, Veronika Laippala, Xiaozhe Yao, Adalberto Barbosa Junior, Aleksandr Drozd, Jordan Clive, Kshitij Gupta, Liangyu Chen, Qi Sun, Ken Tsui, Nour Moustafa-Fahmy, Nicolo Monti, Tai Dang, Ziyang Luo, Tien-Tung Bui, Roberto Navigli, Virendra Mehta, Matthew Blumberg, Victor May, Hiep Nguyen, Sampo Pyysalo
| Challenge: | Pretrained language models are integral part of AI applications, but their high computational cost limits accessibility. |
| Approach: | They evaluate Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. |
| Outcome: | The proposed model outperforms existing models on English, Finnish, Hindi, Japanese, Vietnamese, and code. |
Unsupervised Domain Adaptation for Sparse Retrieval by Filling Vocabulary and Word Frequency Gaps (2022.aacl-main)
Copied to clipboard
| Challenge: | IR models with a pretrained language model outperform lexical approaches like BM25 for vocabulary mismatch. |
| Approach: | They propose an unsupervised domain adaptation method by filling vocabulary gaps by expanding queries and documents through an MLM. |
| Outcome: | The proposed method outperforms the current state-of-the-art domain adaptation method on datasets with a large vocabulary gap from a source domain. |
Efficient Low-Resource Language Adaptation via Multi-Source Dynamic Logit Fusion (2026.acl-long)
Copied to clipboard
| Challenge: | Proxy Tuning offers a logit-level strategy for introducing scaling effects, but it often fails in LRL settings because the large model’s weak LRL competence might overwhelm the knowledge of specialized smaller models. |
| Approach: | They propose a logit-based framework that balances LRL competence from a continually pretrained small model, task competence from high-resource language instruction tuning, and the scaling benefits of large models. |
| Outcome: | Experiments across four model families and eight LRLs show that TriMix outperforms single-model baselines and Proxy Tuning. |
Effective Long-Context Scaling of Foundation Models (2024.naacl-long)
Copied to clipboard
Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, Madian Khabsa, Han Fang, Yashar Mehdad, Sharan Narang, Kshitiz Malik, Angela Fan, Shruti Bhosale, Sergey Edunov, Mike Lewis, Sinong Wang, Hao Ma
| Challenge: | Large language models (LLMs) are rapidly deployed and continue to evolve through scaling. |
| Approach: | They propose a method to train strong long-context LLMs that are capable of utilizing massive context windows of up to 32,000 tokens. |
| Outcome: | The proposed model can surpass gpt-3.5-turbo-16k's overall performance on long-context benchmarks with a cost-effective instruction tuning procedure that is free of expensive annotations. |
Nanda Family: Open-Weights Generative Large Language Models for Hindi (2026.eacl-long)
Copied to clipboard
Aaryamonvikram Singh, Debopriyo Banerjee, Dhruv Sahnan, Monojit Choudhury, Shivam Chauhan, Rocktim Jyoti Das, Xudong Han, Haonan Li, Alok Anil Jadhav, Utkarsh Agarwal, Mukund Choudhary, Fajri Koto, Junaid Hamid Bhat, Awantika Shukla, Samujjwal Ghosh, Samta Kamboj, Onkar Pandit, Lalit Pradhan, Rahul Pal, Sunil Kumar Sahu, Parvez Mullah, Ali El Filali, Zainul Abedien Ahmed Quraishi, Neha Sengupta, Gokulakrishnan Ramakrishnan, Rituraj Joshi, Gurpreet Gosal, Avraham Sheinin, Natalia Vassilieva, Preslav Nakov
| Challenge: | Large language models remain predominantly English-centric, which limits their utility for underrepresented languages. |
| Approach: | They propose to extend Llama’s vocabulary with 20% Hindi-specific tokens, thus halving Hindi tokenization fertility while preserving English efficiency. |
| Outcome: | The proposed models outperform open-weight models of comparable size on a 65B-token corpus and bilingual instruction and safety alignment on . a culturally grounded dataset. |
BTS: Harmonizing Specialized Experts into a Generalist LLM (2025.emnlp-main)
Copied to clipboard
Qizhen Zhang, Prajjwal Bhargava, Chloe Bi, Chris X. Cai, Jakob Nicolaus Foerster, Jeremy Fu, Punit Singh Koura, Ruan Silva, Sheng Shen, Emily Dinan, Suchin Gururangan, Mike Lewis
| Challenge: | Branch-Train-Stitch (BTS) is an efficient and flexible training algorithm for combining independently trained large language model (LLM) experts into a single, capable generalist model. |
| Approach: | They propose an efficient and flexible training algorithm for combining large language model (LLM) experts into a single, capable generalist model using lightweight stitch layers. |
| Outcome: | The proposed model can generalize to new domains despite being frozen . it yields the best generalist performance on a variety of downstream tasks, retaining the specialized capabilities of each of the experts. |
LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding (2024.acl-long)
Copied to clipboard
Mostafa Elhoushi, Akshat Shrivastava, Diana Liskovich, Basil Hosmer, Bram Wasti, Liangzhen Lai, Anas Mahmoud, Bilge Acun, Saurabh Agarwal, Ahmed Roman, Ahmed Aly, Beidi Chen, Carole-Jean Wu
| Challenge: | Large Language Models (LLMs) have been deployed to many applications, yet their high compute and memory requirements lead to high financial and energy costs when deployed to GPU servers. |
| Approach: | They propose an end-to-end solution to speed-up inference of large language models . they apply layer dropout, and show that it increases the accuracy of early exit at earlier layers without adding any auxiliary layers or modules to the model. |
| Outcome: | The proposed method shows speedups of up to 2.16x on summarization for CNN/DM documents, 1.82x on coding, and 2.0x on TOPv2 semantic parsing task. |
NYAYAANUMANA and INLEGALLLAMA: The Largest Indian Legal Judgment Prediction Dataset and Specialized Language Model for Enhanced Decision Analysis (2025.coling-main)
Copied to clipboard
Shubham Kumar Nigam, Deepak Patnaik Balaramamahanthi, Shivam Mishra, Noel Shallum, Kripabandhu Ghosh, Arnab Bhattacharya
| Challenge: | In India, a significant backlog of cases burdens the legal system. |
| Approach: | They present a corpus of 7,02,945 preprocessed Indian legal cases compiled for LJP . they use a domain-specific generative large language model tailored to the intricacies of the legal system . |
| Outcome: | The proposed dataset surpasses existing datasets like PredEx and ILDC, and improves prediction accuracy and comprehensible explanations. |
AfriInstruct: Instruction Tuning of African Languages for Diverse Tasks (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) for African languages perform worse compared to high-resource languages. |
| Approach: | They propose a model that specializes in instruction-tuning of multiple African languages covering various tasks. |
| Outcome: | The proposed model outperforms GPT-3.5-Turbo and other models of similar size in multiple tasks. |
Efficient Domain Continual pretraining by Mitigating the Stability Gap (2025.acl-long)
Copied to clipboard
| Challenge: | Continual pretraining is an important approach for Large Language Models to improve their performance in target domains, learn new topics and languages, and even boost their general capabilities. |
| Approach: | They propose a training strategy that mitigates instability by increasing the number of epochs, along with two data sampling strategies targeting data domain relevance and corpus distribution. |
| Outcome: | The proposed training strategy improves the average medical task performance of the OpenLlama-3B model from 36.2% to 40.7% using only 40% of the original training budget, while also enhancing general task performance without causing forgetting. |
Biomed-Enriched: Data-Efficient Biomedical Pretraining via Paragraph-Level Annotation (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models have demonstrated remarkable capabilities across a wide range of general tasks, from question answering to code generation. |
| Approach: | They use a paragraph-level pipeline to annotate PubMed Central paragraphs . they use XLM-RoBERTa to fine-tune the pipeline and propagate annotations to the full corpus . |
| Outcome: | The proposed approach improves performance on 11 tasks while using 2.5x fewer tokens and only public data. |
Learning Faster with Better Tokens: Parameter-Efficient Vocabulary Adaptation for Specialized Text Summarization (2026.acl-long)
Copied to clipboard
| Challenge: | a new approach to adapt generalist models to expert domains is needed to overcome this problem. |
| Approach: | They propose a parameter-efficient domain adaptation approach that combines vocabulary adaptation with pretraining for LLM-based text summarization. |
| Outcome: | The proposed approach reduces training time by 35-55% over continual pretraining and reduces parameter counts up to 37% w.r.t expansion-only methods. |