DreamW1ngs

Large Language Models (LLMs) have shown impressive reasoning capabilities, yet they often rely on spurious reasoning — producing answers from superficial features, leading to failure at robustness and generalization. We propose AdaR framework to enable adaptive reasoning, wherein models rely on problem-solving logic to produce answers. AdaR synthesizes logically equivalent queries by varying variable values, and trains models with RLVR on these data to penalize spurious logic while encouragin...

Spearman correlation coefficient is set as the best metric to train the sentence-level task. Using the DCSQE framework, synthetic data is generated from the WMT2023 parallel corpus for pre-training, and then fine-tuned on the WMT2022 QE EN-DE training set, all implemented with the Fairseq framework. For a detailed description of the DCSQE framework, please refer to the paper: Alleviating Distribution Shift in Synthetic Data for Machine Translation Quality Estimation

Matthews correlation coefficient is set as the best metric to train the word-level task. Using the DCSQE framework, synthetic data is generated from the WMT2023 parallel corpus for pre-training, and then fine-tuned on the WMT2022 QE EN-DE training set, all implemented with the Fairseq framework. For a detailed description of the DCSQE framework, please refer to the paper: Alleviating Distribution Shift in Synthetic Data for Machine Translation Quality Estimation

Matthews correlation coefficient is set as the best metric to train the word-level task. Using the DCSQE framework, synthetic data is generated from the WMT2023 parallel corpus for pre-training, and then fine-tuned on the WMT2022 QE ZH-EN training set, all implemented with the Fairseq framework. For a detailed description of the DCSQE framework, please refer to the paper: Alleviating Distribution Shift in Synthetic Data for Machine Translation Quality Estimation

Spearman correlation coefficient is set as the best metric to train the sentence-level task. Using the DCSQE framework, synthetic data is generated from the WMT2023 parallel corpus for pre-training, and then fine-tuned on the WMT2022 QE ZH-EN training set, all implemented with the Fairseq framework. For a detailed description of the DCSQE framework, please refer to the paper: Alleviating Distribution Shift in Synthetic Data for Machine Translation Quality Estimation