In the absence of extensive human-annotated data for complex reasoning tasks, self-improvement -- where models are trained on their own outputs -- has emerged as a primary method for enhancing performance. Recently, the approach to self-improvement has shifted toward a more dynamic, online fashion through iterative training processes. However, the critical factors underlying the mechanism of these self-improving methods remain poorly understood, such as under what conditions self-improvement is effective, and what are the bottlenecks in the current iterations. In this work, we identify and propose methods to monitor two pivotal factors in this iterative process: (1) the model's ability to explore and generate high-quality responses among multiple candidates (exploration); and (2) the reliability of external rewards in selecting the best responses from the generated outputs (exploitation). These factors are inherently moving targets throughout the self-improvement cycles, yet their dynamics are rarely discussed in prior research -- It remains unclear what impedes continual model enhancement after only a few iterations. Using mathematical reasoning as a case study, we begin with a quantitative analysis to track the dynamics of exploration and exploitation, discovering that a model's exploratory capabilities rapidly deteriorate over iterations, and the effectiveness of exploiting external rewards diminishes as well due to shifts in distribution from the original policy. Motivated by these findings, we introduce B-STaR, a Self-Taught Reasoning framework that autonomously adjusts configurations across iterations to Balance exploration and exploitation, thereby optimizing the self-teaching effectiveness based on the current policy model and available rewards. Our experiments in mathematical reasoning demonstrate that B-STaR not only enhances the model's exploratory capabilities throughout training but also achieves a more effective balance between exploration and exploitation, leading to superior performance. Crucially, this work deconstructs the opaque nature of self-training algorithms, elucidating the interpretable dynamics throughout the process and highlighting current limitations for future research to address.

Language model pretraining involves training on extensive corpora, where data quality plays a pivotal role. In this work, we aim to directly estimate the contribution of data during pretraining and select pretraining data in an efficient manner. Specifically, we draw inspiration from recent findings showing that compression efficiency (i.e., normalized loss) of diverse models on certain text correlates strongly with their downstream performance, when the text domain aligns with the downstream benchmarks (Huang et al., 2024). Building on this observation, we hypothesize that data on which model losses are predictive of downstream abilities also contribute effectively to learning, which shares similar intuition with Thrush et al. (2024). To leverage this insight, we introduce predictive data selection (PreSelect), a lightweight and efficient data selection method that requires training and deploying only a fastText-based scorer. Through comprehensive experiments with 1B and 3B parameter models, we demonstrate that models trained on 30B tokens selected with PreSelect surpass the performance of the vanilla baseline trained on 300B tokens, achieving a 10x reduction in compute requirements. Furthermore, PreSelect significantly outperforms other competitive data selection baselines, such as DCLM and FineWeb-Edu on a scale of 3B models trained on 100B tokens. We open-source our trained data selection scorer along with the curated datasets at https://github.com/hkust-nlp/PreSelect.

New NLP benchmarks are urgently needed to align with the rapid development of large language models (LLMs). We present C-Eval, the first comprehensive Chinese evaluation suite designed to assess advanced knowledge and reasoning abilities of foundation models in a Chinese context. C-Eval comprises multiple-choice questions across four difficulty levels: middle school, high school, college, and professional. The questions span 52 diverse disciplines, ranging from humanities to science and engineering. C-Eval is accompanied by C-Eval Hard, a subset of very challenging subjects in C-Eval that requires advanced reasoning abilities to solve. We conduct a comprehensive evaluation of the most advanced LLMs on C-Eval, including both English- and Chinese-oriented models. Results indicate that only GPT-4 could achieve an average accuracy of over 60%, suggesting that there is still significant room for improvement for current LLMs. We anticipate C-Eval will help analyze important strengths and shortcomings of foundation models, and foster their development and growth for Chinese users.