Zhiyu Guo

Yijie Xu, Huizai Yao, Zhiyu Guo et al. · tsinghua

Large language models (LLMs) are increasingly deployed in specialized domains such as finance, medicine, and agriculture, where they face significant distribution shifts from their training data. Domain-specific fine-tuning can mitigate this challenge but relies on high-quality labeled data that is expensive and slow to collect in expertise-limited settings. We study label-free test-time adaptation for language models and present SyTTA, an inference-time framework that adapts models on-the-fly without additional supervision. SyTTA couples two complementary uncertainty signals that arise under distribution shift: input-side perplexity, indicating mismatch with domain-specific terminology and patterns, and output-side predictive entropy, indicating diffuse and unstable token probabilities during generation. Across diverse model architectures and domain-specific benchmarks, SyTTA delivers consistent gains. Notably, on agricultural question answering, SyTTA improves Rouge-LSum by over 120% on Qwen-2.5-7B with only 4 extra tokens per query. These results show that effective test-time adaptation for language models is achievable without labeled examples, supporting deployment in label-scarce domains. The code will be made available upon acceptance.

Zhiyu Guo, Hidetaka Kamigaito, Taro Wanatnabe

The rapid advancement in Large Language Models (LLMs) has markedly enhanced the capabilities of language understanding and generation. However, the substantial model size poses hardware challenges, affecting both memory size for serving and inference latency for token generation. To address those challenges, we propose Dependency-aware Semi-structured Sparsity (DaSS), a novel method for the recent prevalent GLU-based LLMs pruning, which incorporates structural dependency into the weight magnitude-based unstructured pruning. We introduce an MLP-specific pruning metric that evaluates the importance of each weight by jointly considering its magnitude and its corresponding MLP intermediate activation norms. DaSS facilitates a balance between the adaptability offered by unstructured pruning and the structural consistency inherent in dependency-based structured pruning. Empirical evaluations on LLaMA2, Mistral, and Gemma model families demonstrate that DaSS not only outperforms both SparseGPT and Wanda in achieving hardware-friendly N:M sparsity patterns but also maintains the computational efficiency of Wanda.

Senhao Liu, Zhiyu Guo, Zhiyuan Ji et al.

Typical financial risk management involves distinct phases for pre-service risk assessment and in-service default detection, often modeled separately. This paper proposes a novel framework, Multi-Granularity Knowledge Distillation (abbreviated as MGKD), aimed at improving pre-service risk prediction through the integration of in-service user behavior data. MGKD follows the idea of knowledge distillation, where the teacher model, trained on historical in-service data, guides the student model, which is trained on pre-service data. By using soft labels derived from in-service data, the teacher model helps the student model improve its risk prediction prior to service activation. Meanwhile, a multi-granularity distillation strategy is introduced, including coarse-grained, fine-grained, and self-distillation, to align the representations and predictions of the teacher and student models. This approach not only reinforces the representation of default cases but also enables the transfer of key behavioral patterns associated with defaulters from the teacher to the student model, thereby improving the overall performance of pre-service risk assessment. Moreover, we adopt a re-weighting strategy to mitigate the model's bias towards the minority class. Experimental results on large-scale real-world datasets from Tencent Mobile Payment demonstrate the effectiveness of our proposed approach in both offline and online scenarios.

Zhiyu Guo, Hidetaka Kamigaito, Taro Watanabe

Scaling the context size of large language models (LLMs) enables them to perform various new tasks, e.g., book summarization. However, the memory cost of the Key and Value (KV) cache in attention significantly limits the practical applications of LLMs. Recent works have explored token pruning for KV cache reduction in LLMs, relying solely on attention scores as a token importance indicator. However, our investigation into value vector norms revealed a notably non-uniform pattern questioning their reliance only on attention scores. Inspired by this, we propose a new method: Value-Aware Token Pruning (VATP) which uses both attention scores and the $ \ell_{1} $ norm of value vectors to evaluate token importance. Extensive experiments on LLaMA2-7B-chat and Vicuna-v1.5-7B across 16 LongBench tasks demonstrate that VATP outperforms attention-score-only baselines in over 12 tasks, confirming the effectiveness of incorporating value vector norms into token importance evaluation of LLMs.