Dynamic sampling is a technique employed in reinforcement Learning for Large Language Models to mitigate training instability, often caused by the gradient-decreasing problem. However, existing dynamic sampling approaches are predominantly designed for tasks with binary rewards (success/failure) and struggle to adapt to the complex reward structures of tool learning. Such task involves interdependent sub-tasks of varying difficulty and yields fine-grained, multi-faceted rewards. To bridge this gap, we introduce Dynamic Sampling with Curriculum Learning (DSCL), an algorithm tailored for the intricate dynamics of tool learning. DSCL integrates two core components: a Reward-Based Dynamic Sampler, which leverages multi-dimensional reward statistics to prioritize high-value training data; and a Task-Based Dynamic Curriculum Learning method to overcome the credit assignment problem caused by the fine-grained rewards. Extensive experiments on widely used tool learning benchmarks demonstrate the efficacy of our approach. DSCL significantly improves model performance, outperforming strong baselines by 4.75% on the BFCL V3 dataset and 4.02% on the API-Bank dataset.

Large Language Models (LLMs) have demonstrated exceptional performance across a broad spectrum of cross-lingual Natural Language Processing (NLP) tasks. However, previous methods predominantly focus on leveraging parallel corpus to conduct instruction data for continuing pre-training or fine-tuning. They ignored the state of parallel data on the hidden layers of LLMs. In this paper, we demonstrate Word-level Cross-lingual Structure (WCS) of LLM which proves that the word-level embedding on the hidden layers are isomorphic between languages. We find that the hidden states of different languages’ input on the LLMs hidden layers can be aligned with an orthogonal matrix on word-level. We prove this conclusion in both mathematical and downstream task ways on two representative LLM foundations, LLaMA2 and BLOOM. Besides, we propose an Isomorphism-based Data Augmentation (IDA) method to apply the WCS on a downstream cross-lingual task, Bilingual Lexicon Induction (BLI), in both supervised and unsupervised ways. The experiment shows the significant improvement of our proposed method over all the baselines, especially on low-resource languages.

Despite their progress in high-resource language settings, unsupervised bilingual lexicon induction (UBLI) models often fail on corpora with low-resource distant language pairs due to insufficient initialization. In this work, we propose a cross-lingual feature extraction (CFE) method to learn the cross-lingual features from monolingual corpora for low-resource UBLI, enabling representations of words with the same meaning leveraged by the initialization step. By integrating cross-lingual representations with pre-trained word embeddings in a fully unsupervised initialization on UBLI, the proposed method outperforms existing state-of-the-art methods on low-resource language pairs (EN-VI, EN-TH, EN-ZH, EN-JA). The ablation study also proves that the learned cross-lingual features can enhance the representational ability and robustness of the existing embedding model.