Challenge: Existing methods to optimize language model pre-training data mixtures are difficult due to the complexity of the data mixture.
Approach: They propose a method to optimize language model pre-training data mixtures by approximating cross-entropy loss via a Mixture of Data Experts (MDE).
Outcome: The proposed method improves performance on a slimPajama dataset with a mixture of data experts.

Similar Papers

Mixture-of-Linguistic-Experts Adapters for Improving and Interpreting Pre-trained Language Models (2023.findings-emnlp)

Copied to clipboard

Challenge: In recent years, pre-trained language models have become the de facto instrument for the field of natural language processing (NLP).
Approach: They propose a method that injects linguistic structures into pre-trained language models in the parameter-efficient fine-tuning setting.
Outcome: The proposed approach outperforms state-of-the-art methods with a comparable number of parameters.
MixCE: Training Autoregressive Language Models by Mixing Forward and Reverse Cross-Entropies (2023.acl-long)

Copied to clipboard

Challenge: Autoregressive language models are trained by minimizing the cross-entropy of the model distribution Q relative to the data distribution P. However, these systems still struggle in many openended generation settings, where they are asked to produce a long text following a short prompt.
Approach: They propose to combine forward and reverse cross-entropy to train autoregressive language models by minimizing the cross-Entropy of the model distribution Q relative to the data distribution P.
Outcome: The proposed model overgeneralizes and produces non-human-like text without complex decoding strategies.
Parameter-Efficient Mixture-of-Experts Architecture for Pre-trained Language Models (2022.coling-1)

Copied to clipboard

Challenge: Recent results show that the mix-of-experts architecture is parameter inefficient . large-scale pre-trained language models can achieve excellent performance in many NLP tasks.
Approach: They propose to build a parameter-efficient mix-of-experts architecture by sharing information across experts.
Outcome: The proposed architecture increases model capacity without increasing computation costs.
Rethinking Data Mixture for Large Language Models: A Comprehensive Survey and New Perspectives (2026.findings-eacl)

Copied to clipboard

Challenge: Existing methods for data mixture improve the generalization capability of large language models (LLMs) on downstream tasks.
Approach: They propose a fine-grained categorization of existing methods and propose three subtypes of offline and online methods.
Outcome: The proposed methods extend beyond offline and online classifications and highlight key challenges in the field of data mixture.
MoEBERT: from BERT to Mixture-of-Experts via Importance-Guided Adaptation (2022.naacl-main)

Copied to clipboard

Challenge: Existing methods for training pre-trained language models have limited practicality due to latency requirements.
Approach: They propose a method that uses a Mixture-of-Experts structure to increase model capacity and inference speed.
Outcome: The proposed method outperforms existing distillation methods on natural language understanding and question answering tasks.
Pretraining Language Models Using Translationese (2024.emnlp-main)

Copied to clipboard

Challenge: a recent study shows that large language models perform well in low-resource languages . a vast majority of languages don't have comparable data as compared to English .
Approach: They propose to use Translationese as synthetic data for pre-training language models for low-resource languages.
Outcome: The proposed method reduces performance of LMs trained on clean data in Indian languages . the proposed model performs better in English than in other languages, but is not comparable to English.
SimSMoE: Toward Efficient Training Mixture of Experts via Solving Representational Collapse (2025.findings-naacl)

Copied to clipboard

Challenge: Sparse mixture of experts (SMoE) is an effective approach for scaling large language models with a constant computational cost.
Approach: They propose a similarity-based sparse mixture of experts (SimSMoE) that guarantees a solution to address the representation collapse problem between experts given a fixed FLOPs budget.
Outcome: The proposed method outperforms existing methods and outperformed other SMoE routing methods in performance.
NeKo: Cross-Modality Post-Recognition Error Correction with Tasks-Guided Mixture-of-Experts Language Model (2025.acl-industry)

Copied to clipboard

Challenge: Existing methods to train a model on a mixture of domain datasets require separate correction language models.
Approach: They propose a multi-task correction MoE that trains experts to become an "expert" of speech-to-text, language-totext and vision-to text datasets by learning to route each dataset’s tokens to its mapped expert.
Outcome: The proposed model outperforms GPT-3.5 and Claude-3.5-Sonnet on the Open ASR Leaderboard and reaches an average relative 5.0% WER reduction and substantial improvements in BLEU scores.
Recipes for Sequential Pre-training of Multilingual Encoder and Seq2Seq Models (2023.findings-acl)

Copied to clipboard

Challenge: Pre-trained encoder-only and sequence-to-sequence models are computationally expensive.
Approach: They propose a recipe to initialize one model from the other to improve pre-training efficiency.
Outcome: The proposed method matches the performance of a from-scratch model with a multilingual encoder while reducing the total compute cost by 27%.
YuLan-Mini: Pushing the Limits of Open Data-efficient Language Model (2025.acl-long)

Copied to clipboard

Challenge: prevailing pre-training approaches for large language models involve several complexities.
Approach: They propose a low-cost training recipe and a robust optimization approach to mitigate training instability . they also propose synthesis, curriculum, and data selection pipelines to integrate data .
Outcome: The proposed model achieves top-tier performance among models with similar parameter scale . it is comparable to industry-leading models that require significantly more data .

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations