Yuanhao Zeng

Yuanhao Zeng, Ao Lu, Lufei Li et al.

Generating diverse responses is crucial for test-time scaling of large language models (LLMs), yet standard stochastic sampling mostly yields surface-level lexical variation, limiting semantic exploration. In this paper, we propose Exploratory Sampling (ESamp), a decoding approach that explicitly encourages semantic diversity during generation. ESamp is motivated by the well-known observation that neural networks tend to make lower-error predictions on inputs similar to those encountered before, and incur higher prediction error on novel ones. Building on this property, we train a lightweight Distiller at test time to predict deep-layer hidden representations of the LLM from its shallow-layer representations to model the LLM's depth-wise representation transitions. During decoding, the Distiller continuously adapts to the mappings induced by the current generation context. ESamp uses the prediction error as a novelty signal to reweight candidate token extensions conditioned on the current prefix, thereby biasing decoding toward less-explored semantic patterns. ESamp is implemented with an asynchronous training--inference pipeline, with less than 5% worst case overhead (1.2% in the optimized release). Empirical results show that ESamp significantly boosts the Pass@k efficiency of reasoning models, showing superior or comparable performance to strong stochastic and heuristic baselines. Notably, ESamp achieves robust generalization across mathematics, science, and code generation benchmarks and breaks the trade-off between diversity and coherence in creative writing. Our code has released at: https://github.com/LinesHogan/tLLM.

Yuanhao Zeng, Min Wang, Yihang Wang et al.

Large Language Models (LLMs) have excelled in various tasks but perform better in high-resource scenarios, which presents challenges in low-resource scenarios. Data scarcity and the inherent difficulty of adapting LLMs to specific tasks compound the challenge. To address the twin hurdles, we introduce \textbf{Leverage Learning}. We present a streamlined implement of this methodology called Token-Efficient Leverage Learning (TELL). TELL showcases the potential of Leverage Learning, demonstrating effectiveness across various LLMs and low-resource tasks, ranging from $10^4$ to $10^6$ tokens. It reduces task data requirements by up to nearly an order of magnitude compared to conventional Supervised Fine-Tuning (SFT) while delivering competitive performance. With the same amount of task data, TELL leads in improving task performance compared to SFT. We discuss the mechanism of Leverage Learning, suggesting it aligns with quantization hypothesis and explore its promising potential through empirical testing.