Rationale-Guided Knowledge Distillation for Cross-Lingual Stance Detection

# Rationale-Guided Knowledge Distillation for Cross-Lingual Stance Detection
Source: [https://arxiv.org/html/2607.18693](https://arxiv.org/html/2607.18693)
Qiuli Zhou,Jingyuan Yao[2024212237@mail\.hfut\.edu\.cn](https://arxiv.org/html/2607.18693v1/mailto:[email protected])School of Computer Science and Information Engineering, Hefei University of TechnologyHefei230601China,Shengeng Tang[tangsg@hfut\.edu\.cn](https://arxiv.org/html/2607.18693v1/mailto:[email protected])School of Computer Science and Information Engineering, Hefei University of TechnologyHefei230601China,Hongzhi Chen[chenhz313@126\.com](https://arxiv.org/html/2607.18693v1/mailto:[email protected])School of Artificial Intelligence Engineering, Hefei Institute of TechnologyHefei238076China,Jun Tang[958748910@qq\.com](https://arxiv.org/html/2607.18693v1/mailto:[email protected])School of Foreign Studies, Hefei University of TechnologyHefei230601ChinaandRichang Hong[hongrc@hfut\.edu\.cn](https://arxiv.org/html/2607.18693v1/mailto:[email protected])School of Computer Science and Information Engineering, Hefei University of TechnologyHefei230601China

\(2026\)

###### Abstract\.

Stance detection aims to identify whether a text expresses a favorable or opposing attitude toward a given target, and serves as an important task for various downstream applications\. Although existing studies have achieved strong performance in monolingual settings, especially in English, many low\-resource languages such as Catalan still lack sufficient annotated data for training effective models\. Cross\-lingual stance detection alleviates this problem by transferring stance knowledge from resource\-rich languages to low\-resource languages\. However, most existing methods mainly rely on semantic alignment between texts and targets, while ignoring the reasoning process required for reliable stance inference\. Although Large Language Models provide strong reasoning ability, their high computational cost and inference latency limit practical deployment\. To address these limitations, we propose a rationale\-guided knowledge distillation framework for cross\-lingual stance detection\. Specifically, we use Chain\-of\-Thought prompting to guide Large Language Models in generating informative rationales, and distill the resulting reasoning knowledge into a compact student model\. We further design a dual\-path distillation mechanism to align rationale\-enhanced and rationale\-free representations, together with their prediction distributions\. In addition, two contrastive learning strategies are introduced to improve stance discrimination\. Experiments on multilingual benchmarks demonstrate that our method consistently outperforms competitive baselines\.

Cross\-Lingual, Stance Detection, Large Language Model

Research results of Anhui Province Philosophy and Social Science Planning Project \(No\. AHSKQ2024D044\)\.

††copyright:none††journalyear:2026††journal:TALLIP††ccs:Computing methodologies Natural language processing††ccs:Information systems Multilingual and cross\-lingual retrieval††ccs:Computing methodologies Machine learning## 1\.Introduction

With the rapid growth of social media platforms, stance detection has attracted increasing attention in natural language processing\. This task aims to identify a user’s attitude, such as favor, neutral, or against, toward a specific target, such as a topic or claim, from a given text\(AlDayel and Magdy,[2021](https://arxiv.org/html/2607.18693#bib.bib5)\)\. It supports a wide range of downstream applications\. For example, in recommendation systems, stance detection can help analyze user preferences and manage content diversity\. By identifying whether users support or oppose specific content, recommendation models can better balance viewpoint diversity and reduce the risk of reinforcing selective exposure\(Alamet al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib1); Ianaet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib2)\)\.

Most existing studies focus on monolingual settings, especially English\(Hardalovet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib3); Küçük and Can,[2020](https://arxiv.org/html/2607.18693#bib.bib4); Allaway and McKeown,[2020](https://arxiv.org/html/2607.18693#bib.bib43)\)\. In contrast, many other languages have limited annotated data and cover only a narrow range of topics, which restricts the performance and transfer ability of stance detection models\(Hardalovet al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib6)\)\. To address this data imbalance, cross\-lingual stance detection transfers stance knowledge from resource\-rich languages to low\-resource languages\(Zhanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib9),[a](https://arxiv.org/html/2607.18693#bib.bib7),[2024](https://arxiv.org/html/2607.18693#bib.bib8)\)\. For example,\(Zhanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib9)\)proposed a cross\-lingual and cross\-target distillation framework, namely CCSD, which mines category information from target representations to reduce target inconsistency and constructs cross\-lingual stance templates for knowledge transfer\. In addition,\(Zhanget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib8)\)introduced external background knowledge and inference knowledge from resource\-rich languages\. However, this strategy depends on external resources that may be difficult to generalize to unseen targets, because constructing comprehensive knowledge bases for diverse topics remains challenging\.

![Refer to caption](https://arxiv.org/html/2607.18693v1/LLM_prompt.png)Figure 1\.Illustrative example of Chain\-of\-Thought \(CoT\) reasoning for stance detection in English\.Example prompt and Chain\-of\-Thought rationale used for stance detection\.Recent advances in Large Language Models have shown strong reasoning capabilities, providing a new direction for cross\-lingual knowledge transfer and stance inference\(Weiet al\.,[2022b](https://arxiv.org/html/2607.18693#bib.bib11); Magisteret al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib10); Weiet al\.,[2022a](https://arxiv.org/html/2607.18693#bib.bib14); Zhouet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib58)\)\. Different from previous neural methods that mainly improve performance through model architecture design\(Zhanget al\.,[2023a](https://arxiv.org/html/2607.18693#bib.bib7)\), human stance judgment often requires identifying implicit emotions, attitudes, and reasoning cues behind the surface text\. Motivated by this observation, we use the Large Language Model \(LLM\) to extract stance related evidence and generate rationales that explain the author’s attitude toward the target\. Prior studies have shown that such reasoning abilities can be elicited through Chain\-of\-Thought prompting, which guides language models to decompose a complex task into intermediate reasoning steps\(Weiet al\.,[2022b](https://arxiv.org/html/2607.18693#bib.bib11); Magisteret al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib10); Hsiehet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib13); Liet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib12)\)\. In this work, we combine Chain\-of\-Thought prompting with In\-Context Learning\(Donget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib15); Minet al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib16)\)to generate coherent rationales for stance detection, as shown in Fig[1](https://arxiv.org/html/2607.18693#S1.F1)\. Specifically, we provide 8 demonstrations covering diverse topics, such as health and political policy, to encourage the model to produce consistent and transferable reasoning patterns\.

Although Large Language Models are effective in reasoning, directly deploying them for stance detection is often impractical due to high computational cost and inference latency\. Their large model size makes it difficult to satisfy real\-time response requirements in practical applications\(Xiaet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib19)\)\. To improve efficiency, we adopt knowledge distillation\(Hintonet al\.,[2015](https://arxiv.org/html/2607.18693#bib.bib18); Gouet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib17)\)to transfer the reasoning ability of Large Language Models to a compact model\. Since the BERT family\(Devlinet al\.,[2019](https://arxiv.org/html/2607.18693#bib.bib20); Pireset al\.,[2019](https://arxiv.org/html/2607.18693#bib.bib21)\)has achieved strong performance on many downstream tasks, and its encoder architecture is suitable for cloze\-style stance prediction, we use multilingual BERT \(mBERT\) as the student model\. Specifically, we fine\-tune mBERT with rationales generated by the LLM, enabling the student model to learn reasoning\-aware representations for stance detection\.

However, directly distilling knowledge from Large Language Models to mBERT is nontrivial\. Large Language Models are usually based on decoder\-only or encoder\-decoder architectures, whereas mBERT follows an encoder only structure\. This architecture gap makes direct output alignment difficult\(Hofstätteret al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib22)\)\. To address this problem, we design a dual\-branch hierarchical distillation framework, which contains a rationale\-enhanced path and a standard path\. The rationale\-enhanced path takes the target, text, and LLM\-generated rationale as input, while the standard path only takes the target and text as input\. By aligning both hidden representations and prediction distributions between the two paths, the student model can learn reasoning knowledge during training while maintaining efficient rationale\-free inference at test time\.

![Refer to caption](https://arxiv.org/html/2607.18693v1/mismatch.png)Figure 2\.Representation and structural gaps in knowledge distillation between decoder\-only Large Language Models and encoder\-only mBERT\.Diagram showing representation and architecture gaps between decoder\-only language models and encoder\-only mBERT\.To further improve the discrimination between opposing stances, we introduce two contrastive learning strategies in the source\-language training stage and the teacher\-student distillation stage\(Gaoet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib23); Chenet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib25); Hanet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib55); Tianet al\.,[2019](https://arxiv.org/html/2607.18693#bib.bib24)\)\. In the source\-language stage, where sufficient labeled data are available, we construct class prototypes from support samples and optimize query representations according to their distances to the stance prototypes\. In the target\-language stage, where training data are limited, we avoid constructing unstable prototypes\. Instead, we design a dual constraint contrastive alignment objective that pulls together the two path representations of the same sample and pushes apart representations from different samples\. This design helps the student model capture stance discriminative features under low\-resource conditions\. We conduct extensive experiments on multilingual datasets covering diverse topics\. The results show that our method consistently outperforms competitive baselines, demonstrating the effectiveness of rationale\-guided knowledge distillation for cross\-lingual stance detection\.

The main contributions of our work are as follows:

- •A rationale\-guided cross\-lingual reasoning framework\.We propose a framework that leverages the reasoning ability of Large Language Models through Chain\-of\-Thought prompting\. By using instruction\-based rationale generation, the proposed method provides explicit reasoning supervision for low\-resource cross\-lingual stance detection\.
- •Dual\-branch hierarchical distillation for architecture compatibility\.To bridge the structural gap between decoder\-based Large Language Models and encoder\-based mBERT, we design a dual\-branch distillation mechanism\. This framework enables the student model to learn from rationale\-enhanced inputs during training while supporting efficient rationale\-free inference through representation\-level and response\-level alignment\.
- •Enhanced stance transfer with contrastive constraints\.We introduce different contrastive learning objectives for the source\-language and target\-language stages\. These constraints complement label supervision and improve the model’s ability to distinguish between opposing stances\.

## 2\.Related Work

This section reviews three lines of research that are closely related to our work: stance detection, knowledge distillation, and LLM reasoning\. These three areas correspond to the task background, the transfer mechanism, and the source of reasoning knowledge in our framework, respectively\. First, stance detection provides the fundamental task setting, where the main research focus has gradually shifted from supervised monolingual classification to cross\-target and cross\-lingual scenarios under data scarcity\. This line of work motivates the need for effective knowledge transfer in low\-resource languages\. Second, knowledge distillation offers a practical solution for transferring useful knowledge from large or specialized teacher models to compact student models, which is essential for improving efficiency and reducing the deployment cost of stance detection systems\. Third, recent studies on LLM reasoning show that Large Language Models can generate intermediate rationales through In\-Context Learning and Chain\-of\-Thought prompting, providing richer logical supervision than conventional label\-based learning\. Taken together, these studies motivate our rationale\-guided knowledge distillation framework: LLM reasoning supplies explicit stance rationales, knowledge distillation transfers such reasoning knowledge into an efficient mBERT\-based student model, and cross\-lingual stance detection serves as the target application\.

### 2\.1\.Stance Detection

Stance detection initially focused on single\-language settings\(Sobhaniet al\.,[2017](https://arxiv.org/html/2607.18693#bib.bib27); Mohammadet al\.,[2016](https://arxiv.org/html/2607.18693#bib.bib28); Glandtet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib29); Confortiet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib30)\), with early research primarily established within highly constrained supervised learning environments\. In these configurations, models are typically trained and tested on a predefined and fixed set of targets\(Mohammadet al\.,[2016](https://arxiv.org/html/2607.18693#bib.bib28); Duet al\.,[2017](https://arxiv.org/html/2607.18693#bib.bib31); Zhouet al\.,[2017](https://arxiv.org/html/2607.18693#bib.bib32)\)\.

During the supervised learning phase, stance targets were generally confined to noun phrases representing specific entities\. This is exemplified by foundational datasets such as SemEval\-2016\(Mohammadet al\.,[2016](https://arxiv.org/html/2607.18693#bib.bib28)\), which focuses on predefined entities, and WT\-WT\(Confortiet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib30)\), which centers on financial mergers\. A critical limitation of these early datasets is their heavy reliance on specific targets, leading to significant performance degradation when models encounter unseen topics\. This bottleneck eventually prompts the research community to pivot toward more flexible cross\-target configurations to enhance model generalization\(Fuet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib57); Weiet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib64); Chenet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib74); Jinet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib75); Tianet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib56); Yeet al\.,[2025a](https://arxiv.org/html/2607.18693#bib.bib76),[b](https://arxiv.org/html/2607.18693#bib.bib79)\)\.

Cross\-target stance detection aims to leverage knowledge gained from a source target to identify sentiments regarding a different, yet related, destination target\(Sunet al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib60)\)\. To achieve this, researchers have developed various strategies to transcend target\-specific constraints\. For example,\(Zhanget al\.,[2023a](https://arxiv.org/html/2607.18693#bib.bib7)\)utilizes a target relation graph to capture semantic associations between targets both within and across languages\. By employing a graph\-based approach and relation alignment strategies, the model can effectively bridge the gap between different language domains\(Wanget al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib80); Liuet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib89); Zhaoet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib90)\)\.

The development of cross\-lingual stance detection primarily focuses on overcoming the imbalance of linguistic resources, undergoing a paradigm shift from early machine translation alignment to multilingual pre\-trained models and adversarial learning\(Zhanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib9),[a](https://arxiv.org/html/2607.18693#bib.bib7); Tianet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib56); Zhanget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib8); Bharathi and Zubiaga,[2025](https://arxiv.org/html/2607.18693#bib.bib36); Liet al\.,[2025a](https://arxiv.org/html/2607.18693#bib.bib37); Huanget al\.,[2023a](https://arxiv.org/html/2607.18693#bib.bib61),[b](https://arxiv.org/html/2607.18693#bib.bib62); Maet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib65); Songet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib77); Chenet al\.,[2026a](https://arxiv.org/html/2607.18693#bib.bib85),[b](https://arxiv.org/html/2607.18693#bib.bib86)\)\. In recent years, with the rise of Large Language Models, cross\-lingual stance detection has been evolving from simple label alignment toward logical reasoning alignment\. Modern frameworks like KEAR\(Zhanget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib8)\)have broken the dependence on labeled data in target languages by inducing language\-agnostic external knowledge \(such as background facts\) from the LLM to serve as a bridge, enhancing prediction accuracy in zero\-shot settings\(Wanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib72); Chenet al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib73); Jinet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib75); Liet al\.,[2025b](https://arxiv.org/html/2607.18693#bib.bib78)\)\.

Despite these advances, most existing stance detection methods still emphasize target\-level or language\-level alignment through labels, templates, or representations, while the intermediate reasoning process that connects the text, target, and stance label is rarely modeled explicitly\. In contrast, our work introduces LLM\-generated rationales as transferable reasoning supervision, enabling the student model to learn reasoning\-aware stance representations rather than relying only on surface semantic alignment\.

### 2\.2\.Knowledge Distillation

Knowledge distillation focuses on transferring expertise from a large, complex model to a more lightweight counterpart\(Hintonet al\.,[2015](https://arxiv.org/html/2607.18693#bib.bib18); Yeet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib54); Gouet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib17); Do and Lee,[2023](https://arxiv.org/html/2607.18693#bib.bib63)\)\. In practical scenarios where computing resources are limited and low latency is required, these techniques are widely used to improve model efficiency, support multilingual and cross\-target transfer, and alleviate bias issues in training data\. The core idea is to use soft labels or intermediate representations generated by high\-performance teacher models to guide the learning process of lightweight or domain\-specific student models\.

In stance detection, knowledge distillation has been adopted to transfer semantic and task\-specific knowledge into more efficient architectures\. For instance, BERTtoCNN\(Liet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib38)\)combines traditional distillation loss with similarity\-preserving loss to inject BERT’s semantic understanding ability into a Text\-CNN architecture, significantly improving inference speed on benchmark tasks such as SemEval\-2016\. Furthermore, CCSD\(Zhanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib9)\)adopts dual\-knowledge distillation with a “Cross\-lingual Teacher” and a “Cross\-target Teacher”, which extract stance category information from source\-language data and provide pseudo labels for training the target\-language student model\(Wanget al\.,[2023a](https://arxiv.org/html/2607.18693#bib.bib81); Chenet al\.,[2026c](https://arxiv.org/html/2607.18693#bib.bib82); Qinet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib87); Guoet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib88)\)\.

Nevertheless, existing distillation methods for stance detection mainly transfer logits, intermediate representations, or pseudo labels from teacher classifiers, which makes it difficult to convey the logical evidence behind stance decisions\. Our framework addresses this limitation by distilling rationale\-enhanced knowledge: LLM\-generated explanations enrich the target\-language training inputs, while the dual\-path representation\-level and response\-level alignment enables mBERT to internalize such reasoning knowledge while remaining rationale\-free at inference time\(Yeet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib54)\)\.

### 2\.3\.LLM reasoning

The evolution of reasoning abilities in Large Language Models has driven a shift in stance detection from traditional supervised classification toward generative logical reasoning\(Aiyappaet al\.,[2026](https://arxiv.org/html/2607.18693#bib.bib40); Taranukhinet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib41); Lanet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib42)\)\. The transition primarily lies in In\-Context Learning\(Donget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib15); Magisteret al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib10); Brownet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib44); Xieet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib45); Liuet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib66); Chenet al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib67); Huanget al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib68); Linet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib71)\), a mechanism that allows models to perform specific tasks through zero\-shot or few\-shot prompting according to input instructions without updating parameters\. Instruction\-tuning techniques further bolster the ability of models to follow complex logical intents, enabling them to capture subtle semantic associations between texts and targets through transfer learning from large\-scale corpora\(Xiaet al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib19); Zhanget al\.,[2025](https://arxiv.org/html/2607.18693#bib.bib69); Liet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib70); Fuet al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib57)\)\.

In practical applications, Large Language Models employ Chain\-of\-Thought\(Weiet al\.,[2022b](https://arxiv.org/html/2607.18693#bib.bib11); Kojimaet al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib46)\)reasoning to generate intermediate analytical steps, which not only significantly enhances prediction accuracy on complex topic sets like VAST\(Allaway and McKeown,[2020](https://arxiv.org/html/2607.18693#bib.bib43)\), but also provides interpretable logical support for decision\-making\. In cross\-lingual settings, the reasoning power of Large Language Models is employed for knowledge mining and alignment; for instance, the RCCA\(Liet al\.,[2025a](https://arxiv.org/html/2607.18693#bib.bib37)\)framework utilizes reinforcement learning to perform cross\-lingual Chain\-of\-Thought alignment, ensuring logical consistency when processing texts from diverse cultural backgrounds\(Zhanget al\.,[2026](https://arxiv.org/html/2607.18693#bib.bib83); Wanget al\.,[2026](https://arxiv.org/html/2607.18693#bib.bib84)\)\.

However, many LLM reasoning approaches either deploy Large Language Models directly for prediction or require additional reasoning alignment during inference, which can introduce high latency and computational cost in practical low\-resource multilingual scenarios\. Our method uses the LLM only as an rationale generator, and then transfers the resulting Chain\-of\-Thought supervision into a compact mBERT student, preserving the benefit of explicit reasoning without requiring LLM inference during deployment\.

## 3\.Methods

Our proposed method first prepares LLM\-generated rationales, followed by a training step in the source language and a teacher\-student distillation step, as illustrated in Fig[3](https://arxiv.org/html/2607.18693#S3.F3)and Fig[4](https://arxiv.org/html/2607.18693#S3.F4)\. During the rationale preparation phase, we employ an In\-Context Learning strategy\(Donget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib15); Magisteret al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib10); Brownet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib44); Xieet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib45)\)to elicit inferential rationales from a frozen Large Language Model \(LLM\), which are subsequently utilized to guide distillation in the target language\. In the source\-language training stage, we partition the dataset into a support set and a query set with a 4:1 ratio\. The mBERT model is fine\-tuned using cross\-lingual templates, where class prototypes are constructed based on the output embeddings of the support set for each category\. By calculating the distances between query samples and these class prototypes, we derive a contrastive loss to optimize both mBERT and the classifier, thereby enhancing the model’s ability to distinguish between opposing stances\. In the knowledge distillation stage, to address the challenge of data scarcity, we augment the limited training data with previously obtained LLM rationales to facilitate reasoning\-level knowledge transfer\. Within this stage, we design a contrastive representation distillation mechanism that pulls together embeddings of the same sample from dual paths while pushing apart those from different samples, thus maximizing the utility of low\-resource target\-language data\. In addition, we design a response\-level distillation to enable knowledge transfer of the soft labels\.

![Refer to caption](https://arxiv.org/html/2607.18693v1/step0.png)Figure 3\.The pipeline of reasoning rationale preparation via In\-Context Learning \(ICL\)\.Pipeline for preparing reasoning rationales with In\-Context Learning\.### 3\.1\.Task Definition

We define the training dataset in the source language asDsrc=\{ti,ci,yi\}i=1Ns​r​cD\_\{\\text\{src\}\}=\\\{t\_\{i\},c\_\{i\},y\_\{i\}\\\}\_\{i=1\}^\{N\_\{src\}\}, wheretit\_\{i\}denotes the target,cic\_\{i\}represents the input text,yi∈\{favor, against\}y\_\{i\}\\in\\\{\\text\{favor, against\}\\\}designates the ground\-truth stance label, andNs​r​cN\_\{src\}is the total number of source samples\. To facilitate contrastive learning, we further partition the source training data into a support setDs=\{ti,ci,yi\}i=1NsupD\_\{s\}=\\\{t\_\{i\},c\_\{i\},y\_\{i\}\\\}\_\{i=1\}^\{N\_\{\\text\{sup\}\}\}and a query setDq=\{ti,ci,yi\}i=1NqD\_\{q\}=\\\{t\_\{i\},c\_\{i\},y\_\{i\}\\\}\_\{i=1\}^\{N\_\{\\text\{q\}\}\}, maintaining a sample ratio of 4:1 \(support samples are divided evenly for each category\) betweenNsupN\_\{\\text\{sup\}\}andNqN\_\{q\}\.

In target languages, however, a significant challenge is the scarcity of high\-quality labeled resources\. To mitigate this, we augment the target\-language training data by generating a comprehensive stance reasoning process for each training pair\(ti,ci\)\(t\_\{i\},c\_\{i\}\), thereby providing enriched supervision signals for mBERT fine\-tuning\. The target language dataset is denoted asDtgt=\{ti,ci,ri,yi\}i=1NtD\_\{\\text\{tgt\}\}=\\\{t\_\{i\},c\_\{i\},r\_\{i\},y\_\{i\}\\\}\_\{i=1\}^\{N\_\{t\}\}, whererir\_\{i\}refers to the inferential rationale generated by the LLM\. Notably, the target language scaleNtN\_\{t\}is significantly smaller than that of the source language, typically by a factor of about 8 to 10\. We employ Qwen\(Qwen Team,[2026](https://arxiv.org/html/2607.18693#bib.bib53)\)as the base Large Language Model \(LLM\) to generate these inferential rationales, keeping its parameters frozen throughout the inference phase\.

To ensure the quality and consistency of the generated rationales, we adopt an In\-Context Learning \(ICL\) strategy\(Donget al\.,[2024](https://arxiv.org/html/2607.18693#bib.bib15); Magisteret al\.,[2023](https://arxiv.org/html/2607.18693#bib.bib10); Brownet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib44); Xieet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib45)\)to guide the LLM’s inferential path\. Specifically, we manually construct 8 diverse prompt examples in the target language\. The inquiry prompt is structured as follows \(translated for clarity\): “The stance of ¡text¿ toward ¡target¿ is ¡stance¿\. Why? Please cite specific keywords or phrases from the source text to support your analysis, and explain the stance by considering the author’s attitude or tone\.” For the model’s responses, we design a Chain\-of\-Thought \(CoT\) reasoning paradigm, exemplified by: “Let’s think step by step\. First, the text mentions …, which implies … and reflects the author’s … attitude\. Second, …” \(a specific example is illustrated in Fig[1](https://arxiv.org/html/2607.18693#S1.F1)\)\. Notably, to enhance the generalization capability of the student model \(mBERT\), these 8 examples cover a broad spectrum of topics\.

![Refer to caption](https://arxiv.org/html/2607.18693v1/step12.png)Figure 4\.Overall architecture of the proposed framework\. Step 1 trains the source\-language mBERT model with cross\-lingual templates and query\-centered contrastive learning\. Step 2 trains the target\-language student model by dual\-path teacher\-student distillation, where LLM\-generated rationales provide additional supervision and both representation\-level and response\-level knowledge are transferred\.Two\-step architecture for source\-language training and target\-language teacher\-student distillation\.
### 3\.2\.Source\-Language Training

In cross\-lingual stance detection, it is essential to first fine\-tune a model on source\-language data in order to transfer relevant knowledge to the target language\. We adopt multilingual BERT \(mBERT\)\(Devlinet al\.,[2019](https://arxiv.org/html/2607.18693#bib.bib20); Pireset al\.,[2019](https://arxiv.org/html/2607.18693#bib.bib21)\)as our backbone due to its strong contextual understanding and its demonstrated effectiveness on downstream tasks such as stance detection\. To promote cross\-lingual representational capability from the outset, we fine\-tune mBERT simultaneously on both source\-language and cross\-lingual templates\. The template structure follows the pattern \(illustrated for clarity\):“The stance of ¡text¿ towards ¡target¿ is \[MASK\]”, where¡text¿and¡target¿serve as slot placeholders, and the\[MASK\]token corresponds to the stance label to be predicted\. For the target language, the instruction portion of the template is translated accordingly while the prediction mechanism over the masked token remains unchanged\.

To further improve the model’s ability to discriminate between opposing stances, we partition the source training data into a support setDSD\_\{S\}and a query setDqD\_\{q\}\. Class prototypes are computed from the support samples of each stance category to characterize category centroids, which are subsequently employed to define a contrastive loss on the query samples\.

Specifically, we first extract the last hidden state corresponding to the\[MASK\]token from the source\-language template:

\(1\)𝐡is=mBERT​\(xis\)\[MASK\]\\mathbf\{h\}\_\{i\}^\{s\}=\\text\{mBERT\}\(x\_\{i\}^\{s\}\)\_\{\[\\text\{MASK\}\]\}wherexxcorresponds to the template sequence\. Analogously from the cross\-lingual template:

\(2\)𝐡it=mBERT​\(xit\)\[MASK\]\\mathbf\{h\}\_\{i\}^\{t\}=\\text\{mBERT\}\(x\_\{i\}^\{t\}\)\_\{\[\\text\{MASK\}\]\}
Within each batch, we then aggregate the support samples belonging to the same stance category to compute a category prototype, defined as the mean of the corresponding embedding vectors\. For theFavorcategory we obtain:

\(3\)𝐂Favor=1\|DS,favor\|​∑𝐡i∈DS,favor𝐡i\\mathbf\{C\}\_\{\\text\{Favor\}\}=\\frac\{1\}\{\|D\_\{S,\\text\{favor\}\}\|\}\\sum\_\{\\mathbf\{h\}\_\{i\}\\in D\_\{S,\\text\{favor\}\}\}\\mathbf\{h\}\_\{i\}and for theAgainstcategory:

\(4\)𝐂Against=1\|DS,against\|​∑𝐡i∈DS,against𝐡i\\mathbf\{C\}\_\{\\text\{Against\}\}=\\frac\{1\}\{\|D\_\{S,\\text\{against\}\}\|\}\\sum\_\{\\mathbf\{h\}\_\{i\}\\in D\_\{S,\\text\{against\}\}\}\\mathbf\{h\}\_\{i\}where𝐡i\\mathbf\{h\}\_\{i\}collectively denotes both𝐡is\\mathbf\{h\}\_\{i\}^\{s\}and𝐡it\\mathbf\{h\}\_\{i\}^\{t\}, and\|DS,category\|\|D\_\{S,\\text\{category\}\}\|indicates the number of support samples of the given category within the current batch\.

Once the category centroids are established, we measure the similarity between each query sample𝐡q∈Dq\\mathbf\{h\}\_\{q\}\\in D\_\{q\}and the prototypes via cosine similarity:

\(5\)s​\(𝐡q,𝐂k\)=𝐡q⋅𝐂k‖𝐡q‖​‖𝐂k‖,k∈\{Favor, Against\}s\(\\mathbf\{h\}\_\{q\},\\mathbf\{C\}\_\{k\}\)=\\frac\{\\mathbf\{h\}\_\{q\}\\cdot\\mathbf\{C\}\_\{k\}\}\{\\\|\\mathbf\{h\}\_\{q\}\\\|\\,\\\|\\mathbf\{C\}\_\{k\}\\\|\},\\quad k\\in\\\{\\text\{Favor, Against\}\\\}wheres​\(𝐡q,𝐂k\)s\(\\mathbf\{h\}\_\{q\},\\mathbf\{C\}\_\{k\}\)represents the cosine similarity between the query embedding and the prototype of categorykk\. The contrastive loss for a query sample is then defined as its similarity to the prototype of theoppositestance:

\(6\)ℒc​\(𝐡q\)=\{s​\(𝐡q,𝐂Against\),if​yq=Favors​\(𝐡q,𝐂Favor\),if​yq=Against\\mathcal\{L\}\_\{\\text\{c\}\}\(\\mathbf\{h\}\_\{q\}\)=\\begin\{cases\}s\(\\mathbf\{h\}\_\{q\},\\mathbf\{C\}\_\{\\text\{Against\}\}\),&\\text\{if \}y\_\{q\}=\\text\{Favor\}\\\\ s\(\\mathbf\{h\}\_\{q\},\\mathbf\{C\}\_\{\\text\{Favor\}\}\),&\\text\{if \}y\_\{q\}=\\text\{Against\}\\end\{cases\}withyqy\_\{q\}denoting the ground\-truth label of the query sample\.

In parallel with the contrastive module, the embeddings derived from mBERT are passed through a classifier to produce predicted stance probabilities\. In our framework, the classifier is implemented as a two\-layer feed\-forward neural network:

\(7\)y^i=Classifier​\(𝐡i\)\\hat\{y\}\_\{i\}=\\text\{Classifier\}\(\\mathbf\{h\}\_\{i\}\)wherey^i\\hat\{y\}\_\{i\}refers to the prediction for either the source or cross\-lingual template \(i\.e\.,y^is\\hat\{y\}\_\{i\}^\{s\}ory^it\\hat\{y\}\_\{i\}^\{t\}\)\.

To enforce consistency between the predictions from the two templates, we introduce a Kullback–Leibler \(KL\) divergence loss:

\(8\)ℒkl=1Ns​∑i=1Ns\(KL​\(y^it∥y^is\)\+KL​\(y^is∥y^it\)\)\\mathcal\{L\}\_\{\\text\{kl\}\}=\\frac\{1\}\{N\_\{s\}\}\\sum\_\{i=1\}^\{N\_\{s\}\}\\Bigl\(\\text\{KL\}\(\\hat\{y\}\_\{i\}^\{t\}\\\|\\hat\{y\}\_\{i\}^\{s\}\)\+\\text\{KL\}\(\\hat\{y\}\_\{i\}^\{s\}\\\|\\hat\{y\}\_\{i\}^\{t\}\)\\Bigr\)whereNsN\_\{s\}denotes the number of support samples per batch\.

Finally, the standard cross\-entropy loss is applied to the predicted labels:

\(9\)ℒce=−1Ns​∑i=1Ns\[yi​log⁡\(y^i\)\+\(1−yi\)​log⁡\(1−y^i\)\]\\mathcal\{L\}\_\{\\text\{ce\}\}=\-\\frac\{1\}\{N\_\{s\}\}\\sum\_\{i=1\}^\{N\_\{s\}\}\\bigl\[y\_\{i\}\\log\(\\hat\{y\}\_\{i\}\)\+\(1\-y\_\{i\}\)\\log\(1\-\\hat\{y\}\_\{i\}\)\\bigr\]withyiy\_\{i\}representing the ground\-truth label of each sample andNsN\_\{s\}again being the batch size of the support set\.

The overall training objective for the source\-language stage is the weighted combination of the above components:

\(10\)ℒtotal=ℒces\+ℒcet\+ℒkl\+α⋅ℒc\\mathcal\{L\}\_\{\\text\{total\}\}=\\mathcal\{L\}\_\{\\text\{ce\}\}^\{s\}\+\\mathcal\{L\}\_\{\\text\{ce\}\}^\{t\}\+\\mathcal\{L\}\_\{\\text\{kl\}\}\+\\alpha\\cdot\\mathcal\{L\}\_\{c\}whereℒces\\mathcal\{L\}\_\{\\text\{ce\}\}^\{s\}andℒcet\\mathcal\{L\}\_\{\\text\{ce\}\}^\{t\}correspond to the cross\-entropy losses computed from the source\-language and cross\-lingual templates, respectively, andα\\alphais a hyperparameter that controls the contribution of the contrastive loss\.

### 3\.3\.Teacher\-Student Distillation

Having fine\-tuned mBERT on the source language, the model has already acquired foundational knowledge for stance detection\. The subsequent stage aims to adapt the model to the target\-language environment using the available target\-language resources\. However, such resources are typically scarce, often an order of magnitude smaller than those in the source language\. Consequently, we refrain from constructing separate support and query sets in this stage, as the reliable estimation of class prototypes would require a substantial number of samples that are unavailable in this low\-resource setting\.

To mitigate the challenge of limited data, we augment the target\-language training corpus with inferential rationales generated by a Large Language Model \(LLM\)\. For the student model, we continue to employ the mBERT instance that was initialized and pre\-fine\-tuned in Source\-Language Training step, thereby ensuring low computational cost and inference latency\. In this stage, the LLM is used as a frozen rationale generator\. The distillation is performed between two mBERT\-based branches: a rationale\-enhanced teacher branch that takes\(ti,ci,ri\)\(t\_\{i\},c\_\{i\},r\_\{i\}\)as input, whererir\_\{i\}denotes the rationale, and a standard student branch that takes only\(ti,ci\)\(t\_\{i\},c\_\{i\}\)as input\. In this way, the student branch learns from the reasoning\-aware representations and predictions of the teacher branch while preserving rationale\-free inference\.

During training, the standard path processes the\(ti,ci\)\(t\_\{i\},c\_\{i\}\)tuple through the encoder to produce the\[C​L​S\]\[CLS\]embedding of the final hidden layer, denoted asFiF\_\{i\}:

\(11\)Fi=mBERT​\(\(t​a​r​g​e​ti,t​e​x​ti\)\)F\_\{i\}=\\text\{mBERT\}\(\(target\_\{i\},text\_\{i\}\)\)Analogously, the reasoning path processes the augmented tuple\(ti,ci,ri\)\(t\_\{i\},c\_\{i\},r\_\{i\}\)to yield the corresponding\[C​L​S\]\[CLS\]embeddingFr​iF\_\{ri\}:

\(12\)Fr​i=mBERT​\(\(t​a​r​g​e​ti,t​e​x​ti,r​e​a​s​o​ni\)\)F\_\{ri\}=\\text\{mBERT\}\(\(target\_\{i\},text\_\{i\},reason\_\{i\}\)\)
The embeddingsFiF\_\{i\}andFr​iF\_\{ri\}are subsequently fed into a Contrastive Representation Distillation module, which serves dual purposes\. From a contrastive perspective, it aligns the feature representations originating from the two different paths for the same sample while simultaneously pushing apart the representations of distinct samples\. In terms of distillation, the module distills the rich representation information from the teacher branch to the student branch\. This dual constraint facilitates effective knowledge distillation and enhances the model’s discriminative capacity\.

First, we compute the cosine similarity between embeddings from the standard path and those from the reasoning path:

\(13\)sim​\(i,j\)=Fi⋅Fr​j‖Fi‖​‖Fr​j‖\\text\{sim\}\(i,j\)=\\frac\{F\_\{i\}\\cdot F\_\{rj\}\}\{\\\|F\_\{i\}\\\|\\,\\\|F\_\{rj\}\\\|\}whereFiF\_\{i\}denotes the embedding of theii\-th sample from the standard path, andFr​jF\_\{rj\}denotes the embedding of thejj\-th sample from the reasoning path\.

Based on these similarities, we formulate a contrastive loss using the InfoNCE objective:

\(14\)ℒc=−∑i=1Nlog⁡exp⁡\(sim​\(i,i\)/τ\)∑j=1Nexp⁡\(sim​\(i,j\)/τ\)\\mathcal\{L\}\_\{c\}=\-\\sum\_\{i=1\}^\{N\}\\log\\frac\{\\exp\(\\text\{sim\}\(i,i\)/\\tau\)\}\{\\sum\_\{j=1\}^\{N\}\\exp\(\\text\{sim\}\(i,j\)/\\tau\)\}whereτ\\tauis a temperature hyperparameter andNNis the batch size\.

The embeddings from both paths are then passed through the classifier, which retains the same two\-layer architecture employed in the Source\-Language Training step:

\(15\)y^i=Classifier​\(Fi\)\\hat\{y\}\_\{i\}=\\text\{Classifier\}\(F\_\{i\}\)wherey^i\\hat\{y\}\_\{i\}represents the predicted stance probability\. The same classifier is applied to the reasoning path embeddingsFr​iF\_\{ri\}in an identical manner\.

Besides the representation\-level distillation, we design a response\-level distillation which transfers the soft label information from the teacher to the student, which assists with the representation distillation to achieve a more efficient knowledge transfer:

\(16\)ℒkl=1Ns​∑i=1Ns\(KL​\(y^it∥y^is\)\+KL​\(y^is∥y^it\)\)\\mathcal\{L\}\_\{\\text\{kl\}\}=\\frac\{1\}\{N\_\{s\}\}\\sum\_\{i=1\}^\{N\_\{s\}\}\\Bigl\(\\text\{KL\}\(\\hat\{y\}\_\{i\}^\{t\}\\\|\\hat\{y\}\_\{i\}^\{s\}\)\+\\text\{KL\}\(\\hat\{y\}\_\{i\}^\{s\}\\\|\\hat\{y\}\_\{i\}^\{t\}\)\\Bigr\)wherey^it\\hat\{y\}\_\{i\}^\{t\}andy^is\\hat\{y\}\_\{i\}^\{s\}refer to the prediction of the teacher model and the student model, respectively\.

Finally, the student model is supervised using the ground\-truth stance labels via the cross\-entropy loss:

\(17\)ℒc​e=−1N​∑i=1N\[yi​log⁡\(y^is\)\+\(1−yi\)​log⁡\(1−y^is\)\]\\mathcal\{L\}\_\{ce\}=\-\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\bigl\[y\_\{i\}\\log\(\\hat\{y\}\_\{i\}^\{s\}\)\+\(1\-y\_\{i\}\)\\log\(1\-\\hat\{y\}\_\{i\}^\{s\}\)\\bigr\]whereyiy\_\{i\}is the ground\-truth label of theii\-th sample\.

The overall training objective for the target\-language stage is the weighted sum of the individual loss components:

\(18\)ℒtotal=ℒc​e\+ℒk​l\+β⋅ℒc\\mathcal\{L\}\_\{\\text\{total\}\}=\\mathcal\{L\}\_\{ce\}\+\\mathcal\{L\}\_\{kl\}\+\\beta\\cdot\\mathcal\{L\}\_\{c\}Here,β\\betais a trade\-off hyperparameter that balances the contribution of the contrastive constraint\.

## 4\.Experiments and Results

This section presents a comprehensive evaluation of our proposed method alongside detailed experimental configurations\. We begin by introducing the datasets utilized in our study in Section[4\.1](https://arxiv.org/html/2607.18693#S4.SS1), which include the X\-stance, CIC, and VaxxStance datasets\. Section[4\.2](https://arxiv.org/html/2607.18693#S4.SS2)delineates the comparative baselines, spanning both monolingual and cross\-lingual stance detection approaches\. In Section[4\.3](https://arxiv.org/html/2607.18693#S4.SS3), we report the primary experimental results across all datasets, providing a performance comparison with baselines to verify the effectiveness of our framework\. Furthermore, Section[4\.5](https://arxiv.org/html/2607.18693#S4.SS5)investigates the impact of rationale quality on model performance by varying the number of in\-context examples\. In Section[4\.6](https://arxiv.org/html/2607.18693#S4.SS6), we justify the 4:1 ratio for constructing support and query sets in the source\-language training phase through empirical analysis\. The influence of various hyperparameters on experimental outcomes is discussed in Section[4\.7](https://arxiv.org/html/2607.18693#S4.SS7)\. Additionally, we provide qualitative analysis in Section[4\.8](https://arxiv.org/html/2607.18693#S4.SS8)to offer deeper insights into the model’s behavior\. Finally, Section[4\.9](https://arxiv.org/html/2607.18693#S4.SS9)explores the effects of different reasoning modes within the Qwen model on the overall performance\.

### 4\.1\.Dataset

\(1\)X\-stance\(Vamvas and Sennrich,[2020](https://arxiv.org/html/2607.18693#bib.bib47)\): X\-stance is a large\-scale multilingual and multi\-target stance detection dataset focusing on Swiss politics\. It encompasses comments in German, French, and Italian, where each sample consists of a candidate’s response to a specific target question under various political themes \(e\.g\., society, education, economy\), labeled as either “favor” or “against\.” In our experimental setup, we designate German as the source language and French as the target language\. We select four primary topics: Society, Economy, Foreign Policy, and Education\. The training set includes 3,430, 4,076, 2,521, and 4,604 samples in German for each respective topic, while approximately 500 French samples are provided for each as target\-language data\. Furthermore, to evaluate the cross\-target generalization capability of our model, we ensure that each topic contains at least 15 distinct targets\. \(2\)CIC\(Zotovaet al\.,[2020](https://arxiv.org/html/2607.18693#bib.bib48)\): CIC is a multilingual stance detection dataset specifically curated to study the movement for Catalan independence within the linguistic context of Spanish and Catalan\. It consists of social media posts \(primarily from Twitter\) where each entry represents a user’s perspective on the “Independence of Catalonia”\. Each post is annotated with a stance label of “favor” or “against”\. In our settings, we utilize Spanish as the source language and Catalan as the target language, with the training set comprising 4,935 source samples and 510 target samples\. \(3\)VaxxStance\(Agerriet al\.,[2021](https://arxiv.org/html/2607.18693#bib.bib49)\): VaxxStance is a multilingual dataset centered on the anti\-vaccination debate\. Each sample consists of a tweet directed at specific targets related to the vaccination movement \(e\.g\., mandatory vaccination, pharmaceutical companies\) within the domain of healthcare policy\. The labels are categorized into “favor” and “against\.” This dataset is particularly suitable for evaluating stance detection in low\-resource and cross\-lingual transfer scenarios\. In our configuration, we adopt a cross\-dataset transfer setting: the model is pre\-trained on the German “Society” topic from X\-stance \(source\) and then adapted to the Spanish “vaccines” target in VaxxStance \(target\), which contains 430 training samples\.

### 4\.2\.Comparative Baselines

To evaluate the effectiveness of our proposed method, we compare it against several state\-of\-the\-art baselines, including both monolingual cross\-target and specialized cross\-lingual stance detection frameworks:

TAN\(Target\-specific Attention Network\)\(Duet al\.,[2017](https://arxiv.org/html/2607.18693#bib.bib31)\)is a classic stance detection framework that utilizes an attention mechanism to capture semantic interactions between a post and its target\. By prioritizing target\-relevant information, TAN effectively identifies stance\-bearing expressions to improve classification accuracy\.

CrossNet\(Xuet al\.,[2018](https://arxiv.org/html/2607.18693#bib.bib50)\)is a self\-attention\-based model specifically developed for cross\-target stance detection\. It captures latent connections and shared linguistic patterns between a source and a destination target, enabling the transfer of stance\-related knowledge to unseen targets\.

JointCL\(Lianget al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib51)\)is a framework for zero\-shot stance detection that integrates stance contrastive learning with target\-aware prototypical graph contrastive learning\. It generalizes stance representations to unseen targets by leveraging prototypical graphs to share reasoning patterns across different domains\.

CLKD\(Cross\-Lingual Knowledge Distillation\)\(Xu and Yang,[2017](https://arxiv.org/html/2607.18693#bib.bib52)\)is a cross\-lingual distillation framework that trains classifiers for unlabeled target languages\. It uses soft predictions from a source\-language teacher model as supervision on parallel corpora, combining distillation with adversarial adaptation to bridge the language gap\.

mBERT\-FT\(Devlinet al\.,[2019](https://arxiv.org/html/2607.18693#bib.bib20)\)serves as a fundamental baseline where mBERT is directly fine\-tuned on the target\-language training data\. Unlike our method, this baseline does not incorporate dual training paths or specialized stance detection templates\.

CCSD\(Cross\-Lingual Cross\-Target Stance Detection\)\(Zhanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib9)\)is a framework designed to address both language barriers and target inconsistency by transferring linguistic and target\-oriented knowledge\. The original model employs dual knowledge distillation framework consisting of a cross\-lingual teacher and a cross\-target teacher\. To ensure a fair comparison within our few\-shot experimental setting, we adapt the CCSD framework by providing a small amount of labeled target\-language data to its student model for training, aligning it with the data availability of our proposed method\.

### 4\.3\.Results and Analysis

Table 1\.Performance comparison between our method and competitive baselines on the X\-stance, CIC, and VaxxStance datasets\. Accuracy \(Acc\) and macroF1F\_\{1\}\-score \(F1F\_\{1\}\) are reported\. “improve \(vs\. CCSD\)” denotes the absolute performance gain of our best setting over the strongest baseline CCSD\.We evaluate our proposed framework across the X\-stance, CIC, and VaxxStance datasets, covering six distinct topics\. Performance is measured using Accuracy and the MacroF1F\_\{1\}\-score as the primary evaluation metrics\. The LLM utilized in our approach is Qwen 3\.5 Flash\(Qwen Team,[2026](https://arxiv.org/html/2607.18693#bib.bib53)\), which is tested in both non\-thinking and thinking modes\. To demonstrate the superiority of our method, we compare its performance against several competitive monolingual and cross\-lingual stance detection baselines\.

Among the monolingual baselines, the prototypical graph mechanism in JointCL\(Lianget al\.,[2022](https://arxiv.org/html/2607.18693#bib.bib51)\)bridges the gap between known and unknown targets more effectively than attention\-based architectures like TAN\(Duet al\.,[2017](https://arxiv.org/html/2607.18693#bib.bib31)\)and CrossNet\(Xuet al\.,[2018](https://arxiv.org/html/2607.18693#bib.bib50)\)\. This advantage suggests that extracting high\-level semantic features is crucial for mitigating target inconsistency\. In cross\-lingual scenarios, direct fine\-tuning of mBERT yields competitive results, leveraging its robust pre\-trained multilingual representations\. Similarly, CLKD\(Xu and Yang,[2017](https://arxiv.org/html/2607.18693#bib.bib52)\)exhibits strong performance, validating that cross\-lingual knowledge distillation is a highly viable strategy in the absence of target\-language labels\. Furthermore, the CCSD\(Zhanget al\.,[2023b](https://arxiv.org/html/2607.18693#bib.bib9)\)framework, which also employs mBERT as its backbone, serves as a strong baseline for joint cross\-lingual and cross\-target tasks\.

The comprehensive experimental results are summarized in Table[1](https://arxiv.org/html/2607.18693#S4.T1)\. Our method consistently outperforms all baselines, achieving a performance gain of approximately 2% to 3% over the state\-of\-the\-art CCSD framework\. Specifically, on the X\-stance dataset, our method demonstrates significant improvements in the Society and Economy topics, with an average increase of 2\.48% in Accuracy and 2\.24% inF1F\_\{1\}\-score\. On the CIC dataset, our approach substantially exceeds the CCSD baseline, particularly in terms ofF1F\_\{1\}\-score, which sees a marked improvement of 3\.91%\. In the VaxxStance dataset, we observe a steady improvement of 1\.37% in Accuracy over CCSD\.

Compared with previous approaches, our framework derives more robust representations by constructing class prototypes and optimizing the model via calculating the distance between query samples and class centroids\. This prototypical approach mitigates the impact of individual sample noise on the overall representation\. Additionally, by incorporating LLM\-generated rationales to augment the target\-language data, we guide mBERT to focus on critical linguistic features and emulate the systematic reasoning process of the LLM\. In the following sections, we provide a detailed ablation study to further verify the effectiveness of the individual components within our framework\.

### 4\.4\.The Effect of LLM\-Generated Rationales

![Refer to caption](https://arxiv.org/html/2607.18693v1/reason_acc.png)\(a\)
![Refer to caption](https://arxiv.org/html/2607.18693v1/reason_f1.png)\(b\)

Figure 5\.Effectiveness of the LLM reasoning on four X\-stance topics\.\(a\)Accuracy comparison of w/o Rationales, Ours \(Qwen\), and Ours \(Qwen \+ Thinking\)\.\(b\)F1F\_\{1\}\-score comparison of the same three settings\.Accuracy and $F\_\{1\}$ comparisons showing the effect of LLM\-generated rationales across X\-stance topics\.One of the core designs of our framework is the reasoning path integrated into the Teacher\-Student Distillation stage\. In this section, we investigate the empirical effectiveness of incorporating LLM\-generated rationales into our method\.

To evaluate the contribution of these rationales, we conducted ablation experiments by removing the reasoning path in the distillation step, thereby retaining only the standard training path\. These experiments were performed across the X\-Society, X\-Economy, X\-Education, and X\-Policy datasets, with the comparative results illustrated in Fig[5](https://arxiv.org/html/2607.18693#S4.F5)\. As indicated by the gray line in the figure which represents the framework performance in the absence of reasoning\-enhanced data, the exclusion of rationales leads to a noticeable decline in overall performance\. Specifically, compared to the framework utilizing Qwen in non\-thinking mode, the Accuracy andF1F\_\{1\}\-score drop by approximately 1\.0% and 1\.5%, respectively\.

A more granular analysis reveals that the performance degradation is most pronounced in the X\-Society dataset, where Accuracy decreases by 1\.11% compared to the Qwen non\-thinking mode and by 1\.63% compared to the thinking mode\. Similarly, theF1F\_\{1\}\-score in this dataset falls by 1\.44% and 1\.66% relative to the two LLM modes\. In contrast, the X\-Education dataset exhibits a more moderate decline, with Accuracy andF1F\_\{1\}\-score dropping by 0\.68% and 0\.61%, respectively\. These results consistently demonstrate that the removal of LLM rationales leads to a significant reduction in model efficacy, thereby validating the critical role and effectiveness of the reasoning path in our target\-language framework\.

### 4\.5\.The Effect of Rationale Quality

Table 2\.Ablation study on the number of ICL demonstrations used for rationale generation with Qwen thinking mode\. The numbers in parentheses denote the absolute change relative to the 8\-sample setting, showing that 8 demonstrations provide the best overall trade\-off between effectiveness and prompt length\.In the rationale preparation stage, we employ In\-Context Learning \(ICL\) to generate high\-quality rationales rich in inferential information, following the Chain\-of\-Thought paradigm: “Let’s think step by step\. First, the word in the text indicates that… Second, …”\. In this section, we investigate the optimal number of in\-context examples required to achieve this effect while balancing computational efficiency\. We conducted five sets of experiments, comparing 2, 4, 6, 8, and 10 examples using the Qwen thinking mode across the X\-Society, X\-Education, X\-Economy, and X\-Policy datasets\. The results are summarized in Table[2](https://arxiv.org/html/2607.18693#S4.T2)\.

As shown in the table, model performance generally improves as the number of in\-context examples increases, reaching a plateau at approximately 8 samples\. In the 2\-sample configuration, the performance across the four datasets is relatively suboptimal; for instance, the Accuracy in X\-Society is 70\.62%, which is 0\.85% lower than that of the 8\-sample setting\. Increasing the number of examples to four yields substantial gains, with Accuracy andF1F\_\{1\}\-score in X\-Society improving by 0\.48% and 0\.49%, respectively\. While performance continues to grow when expanding to 6 samples, the marginal gains begin to diminish\. For example, the improvement in X\-Society is only 0\.09% for both Accuracy andF1F\_\{1\}\-score, representing a significant slowdown in the growth rate compared to the previous increment\.

Similar trends are observed in other datasets\. At 8 samples, the performance increment remains positive but slight\. In the X\-Economy and X\-Education datasets, the growth in Accuracy further tapers to 0\.24% and 0\.11%, respectively, suggesting that the model is approaching its performance ceiling\. Upon increasing the count to 10 samples, the performance stagnates across most datasets\. Specifically, Accuracy in X\-Society even experiences a slight decrease of 0\.07%, while the improvements in X\-Economy and X\-Policy are negligible \(less than 0\.07%\)\. Although the 10\-sample setting occasionally offers marginal benefits, we ultimately selected the 8\-sample configuration\. This decision strikes an optimal balance between effectiveness and efficiency, as utilizing 8 examples reduces input token consumption by approximately 15% compared to the 10\-sample setting, significantly lowering the overall computational cost\.

### 4\.6\.The Effect of Support\-Query Set Construction

Table 3\.Effect of different support\-query set constructions in source\-language training\. Accuracy \(Acc\) andF1F\_\{1\}\-score \(F1F\_\{1\}\) are reported on the four X\-stance topics under different combinations of support and query samples\. The values in parentheses indicate the performance difference relative to the default setting of 8 support samples per class \+ 4 query samples, which provides the best trade\-off between performance and computational cost\.The construction of the support and query sets within the Source\-Language Training path is pivotal, as different configurations can significantly influence model performance\. In this section, we investigate various combinations of these sets to identify the optimal pattern for cross\-lingual stance detection\. Specifically, we vary the number of support samples per class from 4 to 10 to determine the minimum required to construct robust class prototypes\. Concurrently, we evaluate query set sizes ranging from 1 to 4 samples to balance effective model optimization with computational efficiency\.

Initially, we evaluated a configuration consisting of 4 support samples per class and a single query sample \(maintaining a 4:1 ratio\)\. However, the performance was suboptimal; for instance, in the X\-Society dataset, both Accuracy andF1F\_\{1\}\-score remained below 70%\. This deficiency likely stems from the limited number of query samples, where individual noisy samples can disproportionately misguide the model’s optimization\. Increasing the query size to 2 samples yielded a measurable improvement, with Accuracy andF1F\_\{1\}\-score in the X\-Society dataset rising by 1\.09% and 0\.08%, respectively\. Similar gains across other datasets validate our hypothesis that a larger query set mitigates the impact of stochastic noise\. While further increasing the query size to 4 samples continued to enhance performance, the marginal utility began to decrease, with negligible improvements observed in the X\-Education and X\-Policy datasets\.

We speculate that when the support set is limited to 4 samples per class, the representative power of the class prototypes becomes the bottleneck\. To address this, we increased the support set size to 8 samples per class, which resulted in a consistent performance boost of approximately 0\.2%–0\.3% across nearly all datasets\. This indicates that a more substantial support set is essential for constructing robust and stable class prototypes\. Although further expanding the support set to 10 samples per class provided additional gains, the improvement was minor – less than 0\.1% in Accuracy across all datasets while significantly increasing computational overhead\. Consequently, to achieve an optimal balance between empirical performance and computational cost, we designated a configuration of 8 support samples per class and 4 query samples as our standard setting\.

### 4\.7\.The Effects of Hyperparameters

00\.010\.020\.03 \(ours\)0\.040\.050\.060\.070\.080\.090\.1068686969707071717272737370\.6270\.6271\.1671\.1671\.4271\.4271\.5471\.5471\.3871\.3871\.3171\.3171\.0771\.0771\.1471\.1470\.9570\.9570\.7670\.7670\.8370\.8369\.4569\.4569\.6869\.6869\.9369\.9370\.0570\.0569\.8969\.8969\.8369\.8369\.5969\.5969\.6669\.6669\.4769\.4769\.2969\.2969\.3669\.36α\\alphaPerformance \(%\)AccF1F\_\{1\}

\(a\)
00\.010\.020\.030\.040\.05 \(ours\)0\.060\.070\.080\.090\.1068686969707071717272737370\.6970\.6970\.7570\.7571\.2671\.2671\.2471\.2471\.4371\.4371\.5471\.5471\.2171\.2170\.9770\.9771\.0971\.0970\.8470\.8470\.8170\.8169\.1169\.1169\.2769\.2769\.5869\.5869\.7569\.7569\.9469\.9470\.0570\.0569\.7269\.7269\.4969\.4969\.6169\.6169\.2969\.2969\.3369\.33β\\betaPerformance \(%\)AccF1F\_\{1\}

\(b\)

Figure 6\.Effect of the contrastive loss weight on model performance\.\(a\)Performance variation in the source\-language training stage under different weights of the query\-centered contrastive loss\.\(b\)Performance variation in the teacher\-student distillation stage under different weights of the contrastive representation distillation loss\. In both subfigures, Accuracy \(Acc\) andF1F\_\{1\}\-score \(F1F\_\{1\}\) are reported, and the best\-performing setting is highlighted\.Bar charts showing performance sensitivity to the alpha and beta contrastive loss weights\.Our proposed framework involves two key hyperparameters:α\\alphaandβ\\beta, which represent the weights of the contrastive losses in the Source\-Language and Teacher\-Student Distillation stages, respectively\. All sensitivity analyses were conducted on the X\-Society dataset utilizing the thinking mode of Qwen, with results illustrated in Fig[6](https://arxiv.org/html/2607.18693#S4.F6)\.

To evaluate the impact of the source\-language contrastive loss, we varyα\\alphafrom 0 to 0\.10 with an interval of 0\.01\. As shown in the bar chart, the Accuracy increases from 70\.62% atα=0\\alpha=0to a peak of 71\.54% atα=0\.03\\alpha=0\.03\. Beyond this threshold, the performance exhibits a downward trend, with Accuracy decreasing to 70\.83% asα\\alphareaches 0\.10\. TheF1F\_\{1\}\-score follows a similar distribution, suggesting that a moderate contrastive penalty effectively regularizes the source representation\.

Then, we examine the contrastive representation distillation loss weightβ\\betain the target language within a range of 0 to 0\.10 with an interval of 0\.01\. The results demonstrate that Accuracy reaches its optimum atβ=0\.05\\beta=0\.05and diminishes when the value deviates from this center\. TheF1F\_\{1\}\-score consistently follows this pattern, confirming that appropriate contrastive alignment in the target language is essential for robust cross\-lingual transfer\.

### 4\.8\.Qualitative Results

![Refer to caption](https://arxiv.org/html/2607.18693v1/qualitative.png)Figure 7\.Qualitative comparison between the baseline model and our method using Integrated Gradients \(IG\)\. Red indicates positive contribution toward Favor, and blue indicates positive contribution toward Against; darker colors denote stronger attribution scores\.Integrated Gradients visualization comparing token\-level stance attribution for the baseline model and the proposed method\.To provide an intuitive demonstration of our framework’s effectiveness, we conducted a qualitative analysis using Integrated Gradients \(IG\)\(Sundararajanet al\.,[2017](https://arxiv.org/html/2607.18693#bib.bib59)\), a gradient\-based attribution method that assigns importance scores to each input token\. In our visualization, red highlights denote positive contributions toward a “Favor” stance, while blue highlights indicate contributions toward “Against”\. The color intensity reflects the model’s confidence and the magnitude of the token’s impact on the final decision\.

The qualitative results are illustrated in Fig[7](https://arxiv.org/html/2607.18693#S4.F7)\. In case 1, we detect the stance of the text“They hold social gatherings, why should they be excluded?”toward the target“ Would you be in favor of introducing the right to vote and election at the local level for foreigners?”\. Our method identifies the phrase“why should they be excluded”as a whole to identify the coherent favor stance, while the baseline model is misled by the single word“excluded”\. In case 2, our model demonstrates its proficiency in integrating contextual information from entire phrases rather than relying on isolated keywords, leading to more robust and trustworthy stance detection\.

### 4\.9\.The Effect of LLM Reasoning Modes

By comparing the performance of the Qwen thinking mode and non\-thinking mode within our framework, we observe that the thinking mode consistently outperforms the non\-thinking mode across nearly all dataset configurations\. Our empirical analysis suggests that this superiority is primarily attributed to the enhanced quality and higher degree of cleanliness in the generated data\. Through manual qualitative inspection of the rationales produced by both modes, we find that the thinking mode yields responses that are typically more accurate and characterized by greater logical depth\. In contrast, rationales generated in the non\-thinking mode occasionally exhibit instability, manifesting as structural inconsistencies, noisy characters, or irrelevant information\. Such artifacts can introduce undesirable noise into the training process, potentially misleading mBERT and undermining its accuracy in stance detection tasks\. Consequently, the advanced reasoning capabilities inherent in the thinking mode provide more robust and reliable supervisory signals for cross\-lingual knowledge transfer\.

### 4\.10\.Implementation Details

All models are optimized using Adam with a learning rate of2×10−52\\times 10^\{\-5\}, and the number of training epochs is set to 15\. We use multilingual BERT \(mBERT\) as the student backbone, while rationale generation is performed by Qwen3\.5\-Flash\(Qwen Team,[2026](https://arxiv.org/html/2607.18693#bib.bib53)\)which supports a maximum context window of 1,000,000 tokens and a maximum output length of 64,000 tokens\. Specifically, we access the hosted API version of Qwen3\.5\-Flash, rather than deploying the model locally, and keep the LLM parameters frozen throughout the rationale generation stage\. We evaluate both the thinking and non\-thinking modes of Qwen3\.5\-Flash to analyze the influence of rationale quality\. Based on the hyperparameter analysis above, the contrastive loss weight in source\-language training is set toα=0\.03\\alpha=0\.03, and the contrastive representation distillation weight in target\-language training is set toβ=0\.05\\beta=0\.05\. For rationale generation, we use 8 in\-context examples, which provide the best trade\-off between effectiveness and prompt length in our experiments\.

## 5\.Discussion

The experimental results presented in this paper provide strong support for our hypothesis that leveraging the reasoning capabilities of Large Language Models through knowledge distillation can significantly improve cross\-lingual stance detection, especially in low\-resource settings\. Our framework consistently outperforms state\-of\-the\-art baselines such as CCSD, CLKD, and JointCL across multiple multilingual datasets \(X\-stance, CIC, and VaxxStance\), with accuracy gains ranging from approximately 1% to 3%\. Such results align with the observations of previous studies that emphasize the importance of logical reasoning over superficial semantic correlations\. Our work demonstrates that incorporating explicit Chain\-of\-Thought rationales generated by a frozen LLM provides richer supervisory signals that help the student model \(mBERT\) to focus on critical linguistic cues and avoid being misled by isolated keywords, as illustrated in our qualitative analysis\.

Our dual\-branch hierarchical distillation framework is designed to overcome the architectural mismatch between decoder\-only Large Language Models and encoder\-only mBERT\. The consistent performance gains validate that aligning both representation\-level and response\-level outputs across the reasoning and standard paths enables effective knowledge transfer without directly aligning LLM outputs with mBERT outputs\. Furthermore, the introduction of a contrastive learning mechanism for both the source language and the target language proves essential\. Our ablation study shows that removing the reasoning path leads to a noticeable decline of approximately 1\.0%–1\.5% in accuracy andF1F\_\{1\}, confirming that the rationales are central to the model’s improved discriminative capacity\. The sensitivity analysis on hyperparametersα\\alphaandβ\\betafurther reveals that moderate contrastive penalties yield optimal results, echoing findings in prior contrastive learning literature that excessive constraints can hurt generalization\.

The broader implications of our findings extend beyond stance detection\. In many real\-world applications such as public opinion mining, market analysis, and content recommendation systems, low\-resource languages are often neglected due to data scarcity\. Our framework offers a practical pathway to transfer reasoning\-rich knowledge from high\-resource to low\-resource languages without requiring large\-scale parallel corpora or expensive LLM deployment at inference time\. The use of Qwen in thinking mode, which produces cleaner and more logically coherent rationales, underscores the importance of reasoning quality; this suggests that even moderately sized Large Language Models can serve as effective teachers when properly prompted with few\-shot examples\. The optimal selection of 8 in\-context examples \(Section[4\.5](https://arxiv.org/html/2607.18693#S4.SS5)\) strikes a practical balance between performance and computational cost, reducing token consumption by 15% compared to 10 examples while maintaining near\-peak accuracy\.

Nevertheless, several limitations and future directions should be noted\. First, our method still relies on a small amount of labeled target\-language data \(approximately 500 samples per topic\)\. Extending to a truly zero\-shot setting where no target\-language labels are available would require additional mechanisms such as unsupervised domain adaptation or more sophisticated prompt engineering\. Second, while we focus on encoder\-only mBERT for efficiency, exploring other compact architectures \(e\.g\., DistilBERT or XLM\-RoBERTa\) could further improve deployment flexibility\. Third, the reasoning rationales are generated by a single LLM \(Qwen\); using LLM ensembles or iterative rationale refinement might yield even higher quality\. Finally, our current framework assumes binary stance labels \(favor/against\); extending to multi\-class or fine\-grained stance detection is a promising avenue\. Future work could also investigate dynamic adjustment of the number of in\-context examples based on topic complexity, as well as incorporating external knowledge bases to further enrich the reasoning process without increasing inference latency\.

## 6\.Conclusions

In this paper, we proposed a novel knowledge distillation framework that leverages the reasoning capabilities of Large Language Models to improve cross\-lingual stance detection, particularly for low\-resource languages\. Our approach addresses two fundamental challenges: the architectural mismatch between decoder\-only Large Language Models and encoder\-only mBERT, and the scarcity of labeled data in target languages\. To overcome these issues, we designed a dual\-branch hierarchical distillation mechanism that transfers both representation\-level and response\-level knowledge from LLM\-generated rationales to the student model\. Additionally, we introduced two complementary contrastive learning strategies: prototype\-based contrastive loss for the source language \(which benefits from abundant data\) and contrastive alignment for the target language \(which operates under few\-shot conditions\)\. Extensive experiments on the X\-stance, CIC, and VaxxStance datasets demonstrated that our method consistently outperforms competitive baselines including CCSD, CLKD, and JointCL, with the largest improvements reaching approximately 2%–3%\. Furthermore, we empirically determined optimal configurations for In\-Context Learning \(8 examples\) and support\-query set construction \(8 support samples per class with 4 query samples\), balancing effectiveness and computational efficiency\. Our qualitative analysis using Integrated Gradients illustrated that the model learns to focus on coherent phrasal reasoning rather than being misled by isolated keywords\. While our framework still requires a small amount of labeled target\-language data, it provides a practical and efficient solution for cross\-lingual stance detection in realistic settings where deploying Large Language Models directly is infeasible due to high inference latency\. Future work will explore zero\-shot scenarios, alternative compact architectures \(e\.g\., XLM\-RoBERTa\), multi\-class stance detection, and dynamic adjustment of in\-context examples based on topic complexity\.

## References

- R\. Agerri, R\. Centeno, M\. Espinosa, J\. F\. de Landa, and A\. Rodrigo \(2021\)VaxxStance: a dataset for cross\-lingual stance detection on vaccines\.Cited by:[§4\.1](https://arxiv.org/html/2607.18693#S4.SS1.p1.1)\.
- R\. Aiyappa, S\. Senthilmani, J\. An, H\. Kwak, and Y\. Ahn \(2026\)Zero\-shot stance detection in practice: insights on training, prompting, and decoding with a capable lightweight llm\.PeerJ Computer Science12,pp\. e3540\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- M\. Alam, A\. Iana, A\. Grote, K\. Ludwig, P\. Müller, and H\. Paulheim \(2022\)Towards analyzing the bias of news recommender systems using sentiment and stance detection\.InCompanion proceedings of the web conference 2022,pp\. 448–457\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p1.1)\.
- A\. AlDayel and W\. Magdy \(2021\)Stance detection on social media: state of the art and trends\.Information Processing & Management58\(4\),pp\. 102597\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p1.1)\.
- E\. Allaway and K\. McKeown \(2020\)Zero\-shot stance detection: a dataset and model using generalized topic representations\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 8913–8931\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p2.1)\.
- A\. Bharathi and A\. Zubiaga \(2025\)Zero\-shot cross\-lingual stance detection via adversarial language adaptation\.PeerJ Computer Science11,pp\. e2955\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell,et al\.\(2020\)Language models are few\-shot learners\.Advances in neural information processing systems33,pp\. 1877–1901\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1),[§3\.1](https://arxiv.org/html/2607.18693#S3.SS1.p3.1),[§3](https://arxiv.org/html/2607.18693#S3.p1.1)\.
- T\. Chen, S\. Kornblith, M\. Norouzi, and G\. Hinton \(2020\)A simple framework for contrastive learning of visual representations\.InInternational conference on machine learning,pp\. 1597–1607\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p6.1)\.
- W\. Chen, D\. Hong, Z\. Mao, Y\. Cheng, X\. Liu, L\. Zhang, and Y\. Zhang \(2026a\)CreatiParser: generative image parsing of raster graphic designs into editable layers\.arXiv preprint arXiv:2604\.19632\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- W\. Chen, D\. Hong, Y\. Qi, Z\. Han, S\. Wang, L\. Qing, Q\. Huang, and G\. Li \(2022\)Multi\-attention network for compressed video referring object segmentation\.InProceedings of the 30th ACM International Conference on Multimedia,pp\. 4416–4425\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- W\. Chen, G\. Li, X\. Zhang, S\. Wang, L\. Li, and Q\. Huang \(2023\)Weakly supervised text\-based actor\-action video segmentation by clip\-level multi\-instance learning\.ACM Transactions on Multimedia Computing, Communications and Applications19\(1\),pp\. 1–22\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- W\. Chen, G\. Li, X\. Zhang, H\. Yu, S\. Wang, and Q\. Huang \(2021\)Cascade cross\-modal attention network for video actor and action segmentation from a sentence\.InProceedings of the 29th ACM International Conference on Multimedia,pp\. 4053–4062\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1)\.
- W\. Chen, C\. Ye, Z\. Mao, P\. Song, X\. Liu, L\. Zhang, X\. Chang, and Y\. Zhang \(2026b\)FACE\-net: factual calibration and emotion augmentation for retrieval\-enhanced emotional video captioning\.arXiv preprint arXiv:2603\.17455\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- W\. Chen, C\. Ye, P\. Song, L\. Zhang, Y\. Zhang, and Z\. Mao \(2026c\)Subjective\-objective emotion correlated generation network for subjective video captioning\.IEEE Transactions on Image Processing\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p2.1)\.
- C\. Conforti, J\. Berndt, M\. T\. Pilehvar, C\. Giannitsarou, F\. Toxvaerd, and N\. Collier \(2020\)Will\-they\-won’t\-they: a very large dataset for stance detection on twitter\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 1715–1724\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p1.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1)\.
- J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova \(2019\)Bert: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 \(long and short papers\),pp\. 4171–4186\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p4.1),[§3\.2](https://arxiv.org/html/2607.18693#S3.SS2.p1.1),[§4\.2](https://arxiv.org/html/2607.18693#S4.SS2.p6.1),[Table 1](https://arxiv.org/html/2607.18693#S4.T1.10.6.6.12.5.1)\.
- H\. Do and G\. G\. Lee \(2023\)Target\-oriented knowledge distillation with language\-family\-based grouping for multilingual nmt\.ACM Transactions on Asian and Low\-Resource Language Information Processing22\(2\),pp\. 1–18\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p1.1)\.
- Q\. Dong, L\. Li, D\. Dai, C\. Zheng, J\. Ma, R\. Li, H\. Xia, J\. Xu, Z\. Wu, B\. Chang,et al\.\(2024\)A survey on in\-context learning\.InProceedings of the 2024 conference on empirical methods in natural language processing,pp\. 1107–1128\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1),[§3\.1](https://arxiv.org/html/2607.18693#S3.SS1.p3.1),[§3](https://arxiv.org/html/2607.18693#S3.p1.1)\.
- J\. Du, R\. Xu, Y\. He, and L\. Gui \(2017\)Stance classification with target\-specific neural attention networks\.In26th International Joint Conference on Artificial Intelligence, IJCAI 2017,pp\. 3988–3994\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p1.1),[§4\.2](https://arxiv.org/html/2607.18693#S4.SS2.p2.1),[§4\.3](https://arxiv.org/html/2607.18693#S4.SS3.p2.1),[Table 1](https://arxiv.org/html/2607.18693#S4.T1.10.6.6.8.1.1)\.
- F\. Fu, S\. Fang, W\. Chen, and Z\. Mao \(2024\)Sentiment\-oriented transformer\-based variational autoencoder network for live video commenting\.ACM Transactions on Multimedia Computing, Communications and Applications20\(4\),pp\. 1–24\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- T\. Gao, X\. Yao, and D\. Chen \(2021\)Simcse: simple contrastive learning of sentence embeddings\.InProceedings of the 2021 conference on empirical methods in natural language processing,pp\. 6894–6910\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p6.1)\.
- K\. Glandt, S\. Khanal, Y\. Li, D\. Caragea, and C\. Caragea \(2021\)Stance detection in covid\-19 tweets\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 1596–1611\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p1.1)\.
- J\. Gou, B\. Yu, S\. J\. Maybank, and D\. Tao \(2021\)Knowledge distillation: a survey\.International journal of computer vision129\(6\),pp\. 1789–1819\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p1.1)\.
- Y\. Guo, D\. Hong, W\. Chen, Z\. She, C\. Ye, X\. Chang, and Z\. Mao \(2025\)EmoVerse: a mllms\-driven emotion representation dataset for interpretable visual emotion analysis\.arXiv preprint arXiv:2511\.12554\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p2.1)\.
- J\. Han, Q\. Wang, L\. Zhang, W\. Chen, Y\. Song, and Z\. Mao \(2023\)Text style transfer with contrastive transfer pattern mining\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 7914–7927\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p6.1)\.
- M\. Hardalov, A\. Arora, P\. Nakov, and I\. Augenstein \(2021\)Cross\-domain label\-adaptive stance detection\.InProceedings of the 2021 conference on empirical methods in natural language processing,pp\. 9011–9028\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1)\.
- M\. Hardalov, A\. Arora, P\. Nakov, and I\. Augenstein \(2022\)Few\-shot cross\-lingual stance detection with sentiment\-based pre\-training\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.36,pp\. 10729–10737\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1)\.
- G\. Hinton, O\. Vinyals, and J\. Dean \(2015\)Distilling the knowledge in a neural network\.arXiv preprint arXiv:1503\.02531\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p1.1)\.
- S\. Hofstätter, S\. Althammer, M\. Schröder, M\. Sertkan, and A\. Hanbury \(2020\)Improving efficient neural ranking models with cross\-architecture knowledge distillation\.arXiv preprint arXiv:2010\.02666\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p5.1)\.
- C\. Hsieh, C\. Li, C\. Yeh, H\. Nakhost, Y\. Fujii, A\. Ratner, R\. Krishna, C\. Lee, and T\. Pfister \(2023\)Distilling step\-by\-step\! outperforming larger language models with less training data and smaller model sizes\.InFindings of the Association for Computational Linguistics: ACL 2023,pp\. 8003–8017\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1)\.
- H\. Huang, B\. Zhang, Y\. Li, B\. Zhang, Y\. Sun, C\. Luo, and C\. Peng \(2023a\)Knowledge\-enhanced prompt\-tuning for stance detection\.ACM Transactions on Asian and Low\-Resource Language Information Processing22\(6\),pp\. 1–20\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- X\. Huang, W\. Chen, B\. Hu, and Z\. Mao \(2025\)Graph mixture of experts and memory\-augmented routers for multivariate time series anomaly detection\.InProceedings of the AAAI conference on artificial intelligence,Vol\.39,pp\. 17476–17484\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- Y\. Huang, Y\. Liang, Z\. Wu, E\. Zhu, and Z\. Yu \(2023b\)Cross\-lingual sentence embedding for low\-resource chinese\-vietnamese based on contrastive learning\.ACM Transactions on Asian and Low\-Resource Language Information Processing22\(6\),pp\. 1–18\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- A\. Iana, M\. Alam, and H\. Paulheim \(2024\)A survey on knowledge\-aware news recommender systems\.Semantic Web15\(1\),pp\. 21–82\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p1.1)\.
- Y\. Jin, W\. Chen, Y\. Tian, Y\. Song, and C\. Yan \(2024\)Improving radiology report generation with multi\-grained abnormality prediction\.Neurocomputing600,pp\. 128122\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- T\. Kojima, S\. S\. Gu, M\. Reid, Y\. Matsuo, and Y\. Iwasawa \(2022\)Large language models are zero\-shot reasoners\.Advances in neural information processing systems35,pp\. 22199–22213\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p2.1)\.
- D\. Küçük and F\. Can \(2020\)Stance detection: a survey\.ACM Computing Surveys \(CSUR\)53\(1\),pp\. 1–37\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1)\.
- X\. Lan, C\. Gao, D\. Jin, and Y\. Li \(2024\)Stance detection with collaborative role\-infused llm\-based agents\.InProceedings of the international AAAI conference on web and social media,Vol\.18,pp\. 891–903\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- B\. Li, M\. Zou, X\. Zhang, S\. Chen, and Z\. Feng \(2025a\)A reinforcement learning framework for cross\-lingual stance detection using chain\-of\-thought alignment\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 21674–21688\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p2.1)\.
- J\. Li, Z\. Mao, H\. Li, W\. Chen, and Y\. Zhang \(2024\)Exploring visual relationships via transformer\-based graphs for enhanced image captioning\.ACM Transactions on Multimedia Computing, Communications and Applications20\(5\),pp\. 1–23\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- L\. H\. Li, J\. Hessel, Y\. Yu, X\. Ren, K\. Chang, and Y\. Choi \(2023\)Symbolic chain\-of\-thought distillation: small models can also “think” step\-by\-step\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 2665–2679\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1)\.
- Y\. Li, Y\. Sun, and N\. Zhu \(2021\)BERTtoCNN: similarity\-preserving enhanced knowledge distillation for stance detection\.Plos one16\(9\),pp\. e0257130\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p2.1)\.
- Z\. Li, L\. Zhang, K\. Zhang, W\. Chen, Y\. Zhang, and Z\. Mao \(2025b\)Rethinking pseudo word learning in zero\-shot composed image retrieval: from an object\-aware perspective\.InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 833–843\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- B\. Liang, Q\. Zhu, X\. Li, M\. Yang, L\. Gui, Y\. He, and R\. Xu \(2022\)Jointcl: a joint contrastive learning framework for zero\-shot stance detection\.InProceedings of the 60th annual meeting of the association for computational linguistics \(volume 1: long papers\),pp\. 81–91\.Cited by:[§4\.2](https://arxiv.org/html/2607.18693#S4.SS2.p4.1),[§4\.3](https://arxiv.org/html/2607.18693#S4.SS3.p2.1),[Table 1](https://arxiv.org/html/2607.18693#S4.T1.10.6.6.10.3.1)\.
- Z\. Lin, W\. Chen, Y\. Song, and Y\. Zhang \(2024\)Prompting few\-shot multi\-hop question generation via comprehending type\-aware semantics\.InFindings of the Association for Computational Linguistics: NAACL 2024,pp\. 3730–3740\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- C\. Liu, Y\. Tian, W\. Chen, Y\. Song, and Y\. Zhang \(2024\)Bootstrapping large language models for radiology report generation\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 18635–18643\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- X\. Liu, W\. Chen, Z\. Qi, B\. Zhang, and W\. Zhang \(2025\)Matching street view and satellite images via drone imagery and semantic descriptions\.InProceedings of the 3rd International Workshop on UAVs in Multimedia: Capturing the World from a New Perspective,pp\. 4–9\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p3.1)\.
- J\. Ma, Y\. Huang, L\. Wang, X\. Huang, H\. Peng, Z\. Yu, and P\. Yu \(2024\)Augmenting low\-resource cross\-lingual summarization with progression\-grounded training and prompting\.ACM Transactions on Asian and Low\-Resource Language Information Processing23\(9\),pp\. 1–22\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- L\. C\. Magister, J\. Mallinson, J\. Adamek, E\. Malmi, and A\. Severyn \(2023\)Teaching small language models to reason\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),pp\. 1773–1781\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1),[§3\.1](https://arxiv.org/html/2607.18693#S3.SS1.p3.1),[§3](https://arxiv.org/html/2607.18693#S3.p1.1)\.
- S\. Min, X\. Lyu, A\. Holtzman, M\. Artetxe, M\. Lewis, H\. Hajishirzi, and L\. Zettlemoyer \(2022\)Rethinking the role of demonstrations: what makes in\-context learning work?\.InProceedings of the 2022 conference on empirical methods in natural language processing,pp\. 11048–11064\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1)\.
- S\. Mohammad, S\. Kiritchenko, P\. Sobhani, X\. Zhu, and C\. Cherry \(2016\)Semeval\-2016 task 6: detecting stance in tweets\.InProceedings of the 10th international workshop on semantic evaluation \(SemEval\-2016\),pp\. 31–41\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p1.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1)\.
- T\. Pires, E\. Schlinger, and D\. Garrette \(2019\)How multilingual is multilingual bert?\.InProceedings of the 57th annual meeting of the association for computational linguistics,pp\. 4996–5001\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p4.1),[§3\.2](https://arxiv.org/html/2607.18693#S3.SS2.p1.1)\.
- X\. Qin, D\. Hong, W\. Chen, C\. Ye, X\. Liu, P\. Song, and L\. Zhang \(2025\)Query\-based collaborative multimodal token pruning for audio\-visual question answering\.In2025 4th International Conference on Artificial Intelligence, Human\-Computer Interaction and Robotics \(AIHCIR\),pp\. 1–6\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p2.1)\.
- Qwen Team \(2026\)Qwen3\.5: towards native multimodal agents\.External Links:[Link](https://qwen.ai/blog?id=qwen3.5)Cited by:[§3\.1](https://arxiv.org/html/2607.18693#S3.SS1.p2.4),[§4\.10](https://arxiv.org/html/2607.18693#S4.SS10.p1.3),[§4\.3](https://arxiv.org/html/2607.18693#S4.SS3.p1.1)\.
- P\. Sobhani, D\. Inkpen, and X\. Zhu \(2017\)A dataset for multi\-target stance detection\.InProceedings of the 15th conference of the European chapter of the association for computational linguistics: volume 2, short papers,pp\. 551–557\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p1.1)\.
- P\. Song, L\. Zhang, L\. Lan, W\. Chen, D\. Guo, X\. Yang, and M\. Wang \(2025\)Towards efficient partially relevant video retrieval with active moment discovering\.IEEE Transactions on Multimedia\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- Q\. Sun, X\. Xi, J\. Sun, Z\. Wang, and H\. Xu \(2022\)Stance detection with a multi\-target adversarial attention network\.ACM Transactions on Asian and Low\-Resource Language Information Processing22\(2\),pp\. 1–21\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p3.1)\.
- M\. Sundararajan, A\. Taly, and Q\. Yan \(2017\)Axiomatic attribution for deep networks\.InInternational conference on machine learning,pp\. 3319–3328\.Cited by:[§4\.8](https://arxiv.org/html/2607.18693#S4.SS8.p1.1)\.
- M\. Taranukhin, V\. Shwartz, and E\. Milios \(2024\)Stance reasoner: zero\-shot stance detection on social media with explicit reasoning\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),pp\. 15257–15272\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- Y\. Tian, D\. Krishnan, and P\. Isola \(2019\)Contrastive representation distillation\.arXiv preprint arXiv:1910\.10699\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p6.1)\.
- Y\. Tian, W\. Chen, B\. Hu, Y\. Song, and F\. Xia \(2023\)End\-to\-end aspect\-based sentiment analysis with combinatory categorial grammar\.InFindings of the Association for Computational Linguistics: ACL 2023,pp\. 13597–13609\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- J\. Vamvas and R\. Sennrich \(2020\)X\-stance: a multilingual multi\-target dataset for stance detection\.arXiv preprint arXiv:2003\.08385\.Cited by:[§4\.1](https://arxiv.org/html/2607.18693#S4.SS1.p1.1)\.
- C\. Wang, W\. Chen, X\. Cui, Y\. Zhao, Z\. Qi, P\. Huang, X\. Liu, and W\. Zhang \(2025\)Combatting data imbalance and noise in micro\-action recognition\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 14229–14235\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p3.1)\.
- L\. Wang, C\. Ye, W\. Chen, P\. Song, B\. Hu, and Z\. Mao \(2026\)A multi\-agent framework with structured reasoning and reflective refinement for multimodal empathetic response generation\.arXiv preprint arXiv:2604\.18988\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p2.1)\.
- T\. Wang, W\. Chen, J\. Li, Y\. Peng, and Z\. Mao \(2023a\)Contour\-augmented concept prediction network for image captioning\.InInternational Conference on Artificial Neural Networks,pp\. 180–191\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p2.1)\.
- T\. Wang, W\. Chen, Y\. Tian, Y\. Song, and Z\. Mao \(2023b\)Improving image captioning via predicting structured concepts\.InProceedings of the 2023 conference on empirical methods in natural language processing,pp\. 360–370\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- J\. Wei, Y\. Tay, R\. Bommasani, C\. Raffel, B\. Zoph, S\. Borgeaud, D\. Yogatama, M\. Bosma, D\. Zhou, D\. Metzler,et al\.\(2022a\)Emergent abilities of large language models\.arXiv preprint arXiv:2206\.07682\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022b\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in neural information processing systems35,pp\. 24824–24837\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p2.1)\.
- K\. Wei, L\. Jin, Z\. Zhang, Z\. Guo, X\. Li, Q\. Liu, and W\. Feng \(2024\)More than syntaxes: investigating semantics to zero\-shot cross\-lingual relation extraction and event argument role labelling\.ACM Transactions on Asian and Low\-Resource Language Information Processing23\(5\),pp\. 1–21\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1)\.
- R\. Xia, Y\. Ji, Y\. Zhou, H\. Liu, T\. Zhang, and P\. Li \(2025\)From reasoning llms to bert: a two\-stage distillation framework for search relevance\.arXiv preprint arXiv:2510\.11056\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p4.1),[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- S\. M\. Xie, A\. Raghunathan, P\. Liang, and T\. Ma \(2021\)An explanation of in\-context learning as implicit bayesian inference\.arXiv preprint arXiv:2111\.02080\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1),[§3\.1](https://arxiv.org/html/2607.18693#S3.SS1.p3.1),[§3](https://arxiv.org/html/2607.18693#S3.p1.1)\.
- C\. Xu, C\. Paris, S\. Nepal, and R\. Sparks \(2018\)Cross\-target stance classification with self\-attention networks\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),pp\. 778–783\.Cited by:[§4\.2](https://arxiv.org/html/2607.18693#S4.SS2.p3.1),[§4\.3](https://arxiv.org/html/2607.18693#S4.SS3.p2.1),[Table 1](https://arxiv.org/html/2607.18693#S4.T1.10.6.6.9.2.1)\.
- R\. Xu and Y\. Yang \(2017\)Cross\-lingual distillation for text classification\.InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 1415–1425\.Cited by:[§4\.2](https://arxiv.org/html/2607.18693#S4.SS2.p5.1),[§4\.3](https://arxiv.org/html/2607.18693#S4.SS3.p2.1),[Table 1](https://arxiv.org/html/2607.18693#S4.T1.10.6.6.11.4.1)\.
- C\. Ye, W\. Chen, B\. Hu, L\. Zhang, Y\. Zhang, and Z\. Mao \(2025a\)Improving video summarization by exploring the coherence between corresponding captions\.IEEE Transactions on Image Processing\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1)\.
- C\. Ye, W\. Chen, J\. Li, L\. Zhang, and Z\. Mao \(2024\)Dual\-path collaborative generation network for emotional video captioning\.InProceedings of the 32nd ACM International Conference on Multimedia,pp\. 496–505\.Cited by:[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p1.1),[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p3.1)\.
- C\. Ye, W\. Chen, P\. Song, X\. Liu, L\. Zhang, and Z\. Mao \(2025b\)Multi\-round mutual emotion\-cause pair extraction for emotion\-attributed video captioning\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 3320–3329\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p2.1)\.
- H\. Zhang, D\. Hong, M\. Yang, Y\. Cheng, Z\. Zhang, W\. Chen, J\. Shao, X\. Wu, Z\. Wu, and Y\. Jiang \(2025\)Creatidesign: a unified multi\-conditional diffusion transformer for creative graphic design\.arXiv preprint arXiv:2505\.19114\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p1.1)\.
- R\. Zhang, Y\. Tian, P\. Wei, D\. D\. Zeng, and W\. Mao \(2024\)An llm\-enabled knowledge elicitation and retrieval framework for zero\-shot cross\-lingual stance identification\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 12253–12266\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- R\. Zhang, N\. Xu, H\. Yang, Y\. Tian, and W\. Mao \(2023a\)Target\-oriented relation alignment for cross\-lingual stance detection\.InFindings of the Association for Computational Linguistics: ACL 2023,pp\. 6391–6404\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1),[§1](https://arxiv.org/html/2607.18693#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p3.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1)\.
- R\. Zhang, H\. Yang, and W\. Mao \(2023b\)Cross\-lingual cross\-target stance detection with dual knowledge distillation framework\.InProceedings of the 2023 conference on empirical methods in natural language processing,pp\. 10804–10819\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p4.1),[§2\.2](https://arxiv.org/html/2607.18693#S2.SS2.p2.1),[§4\.2](https://arxiv.org/html/2607.18693#S4.SS2.p7.1),[§4\.3](https://arxiv.org/html/2607.18693#S4.SS3.p2.1),[Table 1](https://arxiv.org/html/2607.18693#S4.T1.10.6.6.13.6.1)\.
- Z\. Zhang, P\. Song, J\. Hu, W\. Chen, L\. Ni, and X\. Yang \(2026\)Stimuli\-aware emotion adaptor for enhancing llm in affective explanation captioning\.InICASSP 2026\-2026 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 10662–10666\.Cited by:[§2\.3](https://arxiv.org/html/2607.18693#S2.SS3.p2.1)\.
- B\. Zhao, W\. Chen, B\. Hu, H\. Xie, and Z\. Mao \(2023\)Difference\-aware iterative reasoning network for key relation detection\.In2023 IEEE International Conference on Multimedia and Expo \(ICME\),pp\. 276–281\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p3.1)\.
- Q\. Zhou, J\. Yao, S\. Tang, W\. Chen, L\. Cheng, and J\. Tang \(2025\)Hierarchical knowledge distillation for cross\-lingual stance detection\.In2025 4th International Conference on Artificial Intelligence, Human\-Computer Interaction and Robotics \(AIHCIR\),pp\. 1–5\.Cited by:[§1](https://arxiv.org/html/2607.18693#S1.p3.1)\.
- Y\. Zhou, A\. I\. Cristea, and L\. Shi \(2017\)Connecting targets to tweets: semantic attention\-based model for target\-specific stance detection\.InInternational Conference on Web Information Systems Engineering,pp\. 18–32\.Cited by:[§2\.1](https://arxiv.org/html/2607.18693#S2.SS1.p1.1)\.
- E\. Zotova, R\. Agerri, M\. Nuñez, and G\. Rigau \(2020\)Multilingual stance detection in tweets: the catalonia independence corpus\.InProceedings of the Twelfth Language Resources and Evaluation Conference,pp\. 1368–1375\.Cited by:[§4\.1](https://arxiv.org/html/2607.18693#S4.SS1.p1.1)\.