Selected work

Representative Papers

Zero-Shot Stance Detection in the Wild: Dynamic Target Generation and Multi-Target Adaptation

Jan 27, 2026

This study addresses the challenge of stance detection in real-world social media, where targets are often undefined and dynamically evolving, rendering traditional methods ineffective. The work introduces, for the first time, an open-domain zero-shot stance detection task that leverages large language models (LLMs) to dynamically generate stance targets and adapt to multiple targets without requiring prior target knowledge. Key contributions include the construction of the first Chinese social media stance dataset with multidimensional evaluation metrics and the design of both integrated and two-stage fine-tuning frameworks. Experimental results demonstrate that the two-stage fine-tuned Qwen2.5-7B achieves a composite score of 66.99% in target identification, while the integrated fine-tuned DeepSeek-R1-Distill-Qwen-7B attains an F1 score of 79.26% in stance detection.

0 citationsRead paper

Adversarial Alignment: Ensuring Value Consistency in Large Language Models for Sensitive Domains

Jan 19, 2026

This work addresses the pervasive issue of value inconsistencies and biases exhibited by large language models (LLMs) in sensitive domains such as race, society, and politics. To mitigate this, the authors propose an adversarial alignment framework featuring a novel Attacker-Actor-Critic tripartite architecture: the Attacker generates contentious queries, the Actor produces value-aligned responses, and the Critic filters out low-quality outputs. The framework is optimized through a combination of continual pretraining, instruction tuning, and adversarial training. Additionally, the study introduces the first bilingual (Chinese–English) benchmark dataset for evaluating value alignment in sensitive contexts. Experimental results demonstrate that the resulting model, VC-LLM, significantly outperforms prevailing LLMs in both languages, achieving markedly improved value consistency in sensitive scenarios.

0 citationsRead paper

MSME: A Multi-Stage Multi-Expert Framework for Zero-Shot Stance Detection

Dec 04, 2025

Zero-shot stance detection faces three key challenges: (1) highly dynamic background knowledge, (2) targets often being composite entities or events requiring explicit semantic modeling of their relationships with stance labels, and (3) rhetorical devices (e.g., irony) obscuring the speaker’s true intent. To address these, we propose MSME—a Multi-Stage Multi-Expert framework comprising three phases: knowledge preparation, expert reasoning, and decision aggregation. MSME introduces three specialized experts: a *knowledge expert* integrating large language models with retrieval-augmented generation; a *label expert* performing fine-grained semantic modeling of stance labels; and a *pragmatic expert* equipped with irony-aware reasoning. A meta-arbitrator fuses their outputs. Evaluated on three public benchmarks, MSME significantly outperforms existing zero-shot methods, achieving new state-of-the-art performance. Results demonstrate that multi-source, collaborative reasoning substantially enhances stance identification in linguistically and semantically complex scenarios.

0 citationsRead paper

Enhancing Cross-Lingual Transfer through Reversible Transliteration: A Huffman-Based Approach for Low-Resource Languages

Sep 22, 2025

Low-resource languages—particularly those using non-Latin scripts—suffer from poor cross-lingual transfer performance in large language models (LLMs), while existing transliteration methods lack systematic integration into model training and inference. To address this, we propose the first lightweight input representation framework that tightly integrates reversible transliteration with Huffman coding. Our method achieves lossless, fully invertible transliteration compression at the character level, requiring no vocabulary expansion or additional model parameters. Its key innovation lies in the first application of Huffman coding to compress transliterated sequences, simultaneously improving storage efficiency (50% reduction in file size), computational efficiency (50–80% fewer tokens), and multilingual scalability. Extensive experiments on text classification, machine reading comprehension, and machine translation demonstrate substantial performance gains for low-resource languages, without compromising accuracy on high-resource languages.

0 citationsRead paper

Recent publications

Latest Papers

Zero-Shot Stance Detection in the Wild: Dynamic Target Generation and Multi-Target Adaptation

Jan 27, 2026

This study addresses the challenge of stance detection in real-world social media, where targets are often undefined and dynamically evolving, rendering traditional methods ineffective. The work introduces, for the first time, an open-domain zero-shot stance detection task that leverages large language models (LLMs) to dynamically generate stance targets and adapt to multiple targets without requiring prior target knowledge. Key contributions include the construction of the first Chinese social media stance dataset with multidimensional evaluation metrics and the design of both integrated and two-stage fine-tuning frameworks. Experimental results demonstrate that the two-stage fine-tuned Qwen2.5-7B achieves a composite score of 66.99% in target identification, while the integrated fine-tuned DeepSeek-R1-Distill-Qwen-7B attains an F1 score of 79.26% in stance detection.

0 citationsRead paper

Adversarial Alignment: Ensuring Value Consistency in Large Language Models for Sensitive Domains

Jan 19, 2026

This work addresses the pervasive issue of value inconsistencies and biases exhibited by large language models (LLMs) in sensitive domains such as race, society, and politics. To mitigate this, the authors propose an adversarial alignment framework featuring a novel Attacker-Actor-Critic tripartite architecture: the Attacker generates contentious queries, the Actor produces value-aligned responses, and the Critic filters out low-quality outputs. The framework is optimized through a combination of continual pretraining, instruction tuning, and adversarial training. Additionally, the study introduces the first bilingual (Chinese–English) benchmark dataset for evaluating value alignment in sensitive contexts. Experimental results demonstrate that the resulting model, VC-LLM, significantly outperforms prevailing LLMs in both languages, achieving markedly improved value consistency in sensitive scenarios.

0 citationsRead paper

MSME: A Multi-Stage Multi-Expert Framework for Zero-Shot Stance Detection

Dec 04, 2025

Zero-shot stance detection faces three key challenges: (1) highly dynamic background knowledge, (2) targets often being composite entities or events requiring explicit semantic modeling of their relationships with stance labels, and (3) rhetorical devices (e.g., irony) obscuring the speaker’s true intent. To address these, we propose MSME—a Multi-Stage Multi-Expert framework comprising three phases: knowledge preparation, expert reasoning, and decision aggregation. MSME introduces three specialized experts: a *knowledge expert* integrating large language models with retrieval-augmented generation; a *label expert* performing fine-grained semantic modeling of stance labels; and a *pragmatic expert* equipped with irony-aware reasoning. A meta-arbitrator fuses their outputs. Evaluated on three public benchmarks, MSME significantly outperforms existing zero-shot methods, achieving new state-of-the-art performance. Results demonstrate that multi-source, collaborative reasoning substantially enhances stance identification in linguistically and semantically complex scenarios.

0 citationsRead paper

Enhancing Cross-Lingual Transfer through Reversible Transliteration: A Huffman-Based Approach for Low-Resource Languages

Sep 22, 2025

Low-resource languages—particularly those using non-Latin scripts—suffer from poor cross-lingual transfer performance in large language models (LLMs), while existing transliteration methods lack systematic integration into model training and inference. To address this, we propose the first lightweight input representation framework that tightly integrates reversible transliteration with Huffman coding. Our method achieves lossless, fully invertible transliteration compression at the character level, requiring no vocabulary expansion or additional model parameters. Its key innovation lies in the first application of Huffman coding to compress transliterated sequences, simultaneously improving storage efficiency (50% reduction in file size), computational efficiency (50–80% fewer tokens), and multilingual scalability. Extensive experiments on text classification, machine reading comprehension, and machine translation demonstrate substantial performance gains for low-resource languages, without compromising accuracy on high-resource languages.

0 citationsRead paper