Fine-tuning plays an essential role in improving the performance of large language models (LLMs) on specific tasks. A central challenge lies in designing data-efficient strategy to achieve better fine-tuning performance. Curriculum learning, which organizes data from easy to hard, has become a widely adopted technique in LLMs training. However, existing methods for curriculum learning focus only on the difficulty of samples, while neglecting their contribution to improving model performance, making them vulnerable when applied to fine-tuning LLMs. To address this, we propose Difficulty-Utility Curriculum Learning (DUCL), a curriculum learning framework that jointly considers difficulty and utility. DUCL introduces a novel scoring method, Difficulty-Utility Evaluation (DUE), and a soft scheduling strategy called Window Ordering, which together promote efficient and effective fine-tuning. Our method not only improves convergence and final performance with negligible computational overhead, but is also broadly applicable across a wide range of tasks, making it a practical and scalable solution for LLMs fine-tuning.

Large Language Models (LLMs) achieve remarkable success, but their static nature leads to inherent limitations and persistent mistakes in dynamic real-world scenarios. While Continual Instruction Tuning (CIT) and Continual Pre-training (CPT) are primary continual learning approaches, they struggle with scalable knowledge acquisition and maintaining model capabilities. To address these, we propose the Continue Evolving from Mistakes (CEM) method, a novel and data-efficient framework for continuous LLM evolution. Inspired by human learning, CEM establishes an iterative process: it efficiently collects targeted CPT data by robustly identifying LLM mistakes and uncertainties (via an Ambiguity-Aware Knowledge Collection (AAKC) algorithm), and employs a novel joint training paradigm that leverages CIT and CPT to assimilate knowledge efficiently while maintaining existing capabilities and mitigating catastrophic forgetting. Extensive experiments confirm CEM's effectiveness, yielding substantial accuracy gains for multiple models, increasing accuracy by up to 29.63%. Code and datasets are available on GitHub https://anonymous.4open.science/r/cem-BB25.

Recent advancements in large language models (LLMs) have demonstrated that progressive refinement, rather than providing a single answer, results in more accurate and thoughtful outputs. However, existing methods often rely heavily on supervision signals to evaluate previous responses, making it difficult to effectively assess output quality in more open-ended scenarios. Additionally, these methods are typically designed for specific tasks, which limits their generalization to new domains. To address these limitations, we propose Progressive Thought Refinement (PTR), a framework that enables LLMs to progressively refine their responses. PTR operates in two phases: (1) Thought data construction stage: We propose a weak and strong model collaborative selection strategy to build a high-quality progressive refinement dataset to ensure logical consistency from thought to answers, and the answers are gradually refined in each round. (2) Thought-Mask Fine-Tuning Phase: We design a training structure to mask the "thought" and adjust loss weights to encourage LLMs to refine prior thought, teaching them to implicitly understand "how to improve" rather than "what is correct." Experimental results show that PTR significantly enhances LLM performance across ten diverse tasks (avg. from 49.6% to 53.5%) without task-specific fine-tuning. Notably, in more open-ended tasks, LLMs also demonstrate substantial improvements in the quality of responses beyond mere accuracy, suggesting that PTR truly teaches LLMs to self-improve over time. Our work is now open-source. https://github.com/cydu24/Progressive-Thought-Refinement

Existing Large Language Models (LLMs) generate texts through unidirectional autoregressive decoding methods to respond to various queries. These methods tend to consider token selection in a simple sequential manner, making it easy to fall into suboptimal options when encountering uncertain tokens, referred to as chaotic points in this paper. Many chaotic points exist in the texts generated by LLMs, and they often significantly affect the quality of subsequently generated tokens, which would interfere with LLMs’ generation. In this paper, we propose DEval, a decoding framework for enhancing LLMs’ generation. Analogous to human decision-making, DEval integrates speculation and evaluation into decoding, allowing LLMs to make more careful decisions, and thus optimize token selection at chaotic points. Experiments across various tasks using different LLMs demonstrate DEval’s effectiveness. The code and more technical details are available at our repository1.