论文检索

Baixuan Li, Yunlong Fan, Tianyi Ma, Miao Gao, Chuanqi Shi, Zhiqiang Gao

Complex multi-hop question answering requires large language models (LLMs) not only to retrieve external knowledge but also to reason over the retrieved information in order to arrive at the final solution. This involves two key challenges: (i) how to effectively explore the solution space and generate more potentially correct solution candidates, and (ii) how to select the optimal solution from multiple solution candidates, both of which require a training-free approach without introducing a more powerful teacher model. To address these challenges, we propose Retrieval-Augmented Monte Carlo Tree Self-Play with Reasoning Consistency (RASPberry), which introduces a more flexible action-level sampling granularity compared to existing methods, leverages Monte Carlo Tree Search for efficient solution space exploration, and utilizes an enhanced version of reasoning consistency to guide the selection of the optimal solution. Experimental results demonstrate that our proposed RASPberry effectively tackles the two challenges outlined above, achieving more efficient RAG inference-time scaling. Our code is available at https://github.com/BaixuanLi/RASPberry.

Junling Wang, Anna Rutkiewicz, April Wang, Mrinmaya Sachan

Visuals are valuable tools for teaching math word problems (MWPs), helping young learners interpret textual descriptions into mathematical expressions before solving them.However, creating such visuals is labor-intensive and there is a lack of automated methods to support this process. In this paper, we present Math2Visual, an automatic framework for generating pedagogically meaningful visuals from MWP text descriptions. Math2Visual leverages a pre-defined visual language and a design space grounded in interviews with math teachers, to illustrate the core mathematical relationships in MWPs.Using Math2Visual, we construct an annotated dataset of 1,903 visuals and evaluate Text-to-Image (TTI) models for their ability to generate visuals that align with our design. We further fine-tune several TTI models with our dataset, demonstrating improvements in educational visual generation. Our work establishes a new benchmark for automated generation of pedagogically meaningful visuals and offers insights into key challenges in producing multimodal educational content, such as the misrepresentation of mathematical relationships and the omission of essential visual elements.

Jiho Jin, Woosung Kang, Junho Myung, Alice Oh

Measuring social bias in large language models (LLMs) is crucial, but existing bias evaluation methods struggle to assess bias in long-form generation. We propose a Bias Benchmark for Generation (BBG), an adaptation of the Bias Benchmark for QA (BBQ), designed to evaluate social bias in long-form generation by having LLMs generate continuations of story prompts. Building our benchmark in English and Korean, we measure the probability of neutral and biased generations across ten LLMs. We also compare our long-form story generation evaluation results with multiple-choice BBQ evaluation, showing that the two approaches produce inconsistent results.

Jiyang Qiu, Xinbei Ma, Zhuosheng Zhang, Hai Zhao, Yun Li, Qianren Wang

Large language models (LLMs) have exhibited remarkable versatility and adaptability, while their widespread adoption across various applications also raises critical safety concerns.This paper focuses on the impact of backdoored LLMs. Traditional backdoor injection methods are primarily limited to yes-or-no discriminative tasks, leading users to underestimate the potential risks of backdoored LLMs.Given the inherently generative nature of LLMs, this paper reveals that a generative backdoor injected into LLMs can expose the true safety risks in their applications. We propose an editing-based generative backdoor, named MEGen, aiming to expand the backdoor to generative tasks in a unified format of any text-to any text, leading to natural generations with a specific intention. Experiments show that MEGen achieves a high attack success rate by adjusting only a small set of local parameters with few-shot samples. Notably, we show that the backdoored model, when triggered, can freely output pre-set dangerous information while completing downstream tasks.Our work highlights that MEGen enables backdoors in LLMs to exhibit generative capabilities, causing potential safety risks by altering the generative style. The code is available at [https://github.com/MonoQ-hub/MEGen](https://github.com/MonoQ-hub/MEGen).

Kepu Zhang, Weijie Yu, Sunhao Dai, Jun Xu

In this paper, we propose CitaLaw, the first benchmark designed to evaluate LLMs’ ability to produce legally sound responses with appropriate citations. CitaLaw features a diverse set of legal questions for both laypersons and practitioners, paired with a comprehensive corpus of law articles and precedent cases as a reference pool. This framework enables LLM-based systems to retrieve supporting citations from the reference corpus and align these citations with the corresponding sentences in their responses. Moreover, we introduce syllogism-inspired evaluation methods to assess the legal alignment between retrieved references and LLM-generated responses, as well as their consistency with user questions. Extensive experiments on 2 open-domain and 7 legal-specific LLMs demonstrate that integrating legal references substantially enhances response quality. Furthermore, our proposed syllogism-based evaluation method exhibits strong agreement with human judgments.

Haoke Zhang, Xiaobo Liang, Cunxiang Wang, Juntao Li, Min Zhang

The OpenAI o1-series models have demonstrated that leveraging long-form Chain of Thought (CoT) can substantially enhance performance. However, the recursive thinking capabilities of Large Language Models (LLMs) remain limited, particularly in the absence of expert-curated data for distillation. In this paper, we propose AvR: Alignment via Refinement, a novel method aimed at unlocking the potential of LLMs for recursive reasoning through long-form CoT. AvR introduces a refinement process that integrates criticism and improvement actions, guided by differentiable learning techniques to optimize refinement-aware rewards. As a result, the synthesized multi-round data can be organized as a long refinement thought, further enabling test-time scaling. Experimental results show that AvR significantly outperforms conventional preference optimization methods. Notably, with only 3k synthetic samples, our method boosts the performance of the LLaMA-3-8B-Instruct model by over 20% in win rate on AlpacaEval 2.0. Our code is available at Github .

Chenchen Yuan, Zheyu Zhang, Shuo Yang, Bardh Prenkaj, Gjergji Kasneci

Large Language Models (LLMs) have shown impressive moral reasoning abilities. Yet they often diverge when confronted with complex, multi-factor moral dilemmas. To address these discrepancies, we propose a framework that synthesizes multiple LLMs’ moral judgments into a collectively formulated moral judgment, realigning models that deviate significantly from this consensus. Our aggregation mechanism fuses continuous moral acceptability scores (beyond binary labels) into a collective probability, weighting contributions by model reliability. For misaligned models, a targeted embedding-optimization procedure fine-tunes token embeddings for moral philosophical theories, minimizing JS divergence to the consensus while preserving semantic integrity. Experiments on a large-scale social moral dilemma dataset show our approach builds robust consensus and improves individual model fidelity. These findings highlight the value of data-driven moral alignment across multiple models and its potential for safer, more consistent AI systems.

Yuanhe Zhang, Zhenhong Zhou, Wei Zhang, Xinyue Wang, Xiaojun Jia, Yang Liu, Sen Su

Large Language Models (LLMs) have demonstrated remarkable performance across diverse tasks yet still are vulnerable to external threats, particularly LLM Denial-of-Service (LLM-DoS) attacks. Specifically, LLM-DoS attacks aim to exhaust computational resources and block services. However, existing studies predominantly focus on white-box attacks, leaving black-box scenarios underexplored. In this paper, we introduce Auto-Generation for LLM-DoS (\textbf{AutoDoS}) attack, an automated algorithm designed for black-box LLMs. AutoDoS constructs the DoS Attack Tree and expands the node coverage to achieve effectiveness under black-box conditions. By transferability-driven iterative optimization, AutoDoS could work across different models in one prompt.Furthermore, we reveal that embedding the Length Trojan allows AutoDoS to bypass existing defenses more effectively.Experimental results show that AutoDoS significantly amplifies service response latency by over \textbf{250}\times\uparrow, leading to severe resource consumption in terms of GPU utilization and memory usage. Our work provides a new perspective on LLM-DoS attacks and security defenses.

Maja Stahl, Timon Ziegenbein, Joonsuk Park, Henning Wachsmuth

Training large language models (LLMs) to follow instructions has significantly enhanced their ability to tackle unseen tasks. However, despite their strong generalization capabilities, instruction-following LLMs encounter difficulties when dealing with tasks that require domain knowledge. This work introduces a specialized instruction fine-tuning for the domain of computational argumentation (CA). The goal is to enable an LLM to effectively tackle any unseen CA tasks while preserving its generalization capabilities. Reviewing existing CA research, we crafted natural language instructions for 105 CA tasks to this end. On this basis, we developed a CA-specific benchmark for LLMs that allows for a comprehensive evaluation of LLMs’ capabilities in solving various CA tasks. We synthesized 52k CA-related instructions, adapting the self-instruct process to train a CA-specialized instruction-following LLM. Our experiments suggest that CA-specialized instruction fine-tuning significantly enhances the LLM on both seen and unseen CA tasks. At the same time, performance on the general NLP tasks of the SuperNI benchmark remains stable.

Chen Zhang, Mingxu Tao, Zhiyuan Liao, Yansong Feng

Large language models (LLMs) excel in high-resource languages but struggle with low-resource languages (LRLs), particularly those spoken by minority communities in China, such as Tibetan, Uyghur, Kazakh, and Mongolian. To systematically track the progress in these languages, we introduce MiLiC-Eval, a benchmark designed for minority languages in China, featuring 24K instances across 9 tasks. MiLiC-Eval focuses on underrepresented writing systems. Its parallelism between tasks and languages can provide a faithful and fine-grained assessment of linguistic and problem-solving skills. Our evaluation reveals that open-source LLMs perform poorly on syntax-intensive tasks and multi-script languages. We further demonstrate how MiLiC-Eval can help advance LRL research in handling diverse writing systems and understanding the process of language adaptation.

Yueqi Song, Frank F. Xu, Shuyan Zhou, Graham Neubig

Web browsers are a portal to the internet, where much of human activity is undertaken. Thus, there has been significant research work in AI agents that interact with the internet through web browsing.However, there is also another interface designed specifically for machine interaction with online content: application programming interfaces (APIs). In this paper we ask – *what if we were to take tasks traditionally tackled by Browsing Agents, and give AI agents access to APIs*?To do so, we propose two varieties of agents: (1) an API-calling agent that attempts to perform online tasks through APIs only, similar to traditional coding agents, and (2) a Hybrid Agent that can interact with online data through both web browsing and APIs.In experiments on WebArena, a widely-used and realistic benchmark for web navigation tasks, we find that API-Based Agents outperform web Browsing Agents.Hybrid Agents out-perform both others nearly uniformly across tasks, resulting in a more than 24.0% absolute improvement over web browsing alone, achieving a success rate of 38.9%, the SOTA performance among task-agnostic agents.These results strongly suggest that when APIs are available, they present an attractive alternative to relying on web browsing alone.

Sougata Saha, Monojit Choudhury

Measuring the generalization ability of Large Language Models (LLMs) is challenging due to data contamination. As models grow and computation becomes cheaper, ensuring tasks and test cases are unseen during training phases will become nearly impossible. We argue that knowledge-retrieval and reasoning tasks are not ideal for measuring generalization, as LLMs are not trained for specific tasks. Instead, we propose user behavior prediction, also a key aspect of personalization, as a theoretically sound, scalable, and robust alternative. We introduce a novel framework for this approach and test it on movie and music recommendation datasets for GPT-4o, GPT-4o-mini, and Llama-3.1-8B-Instruct. Results align with our framework’s predictions, showing GPT-4o outperforms GPT-4o-mini and Llama, though all models have much room for improvement, especially Llama.

Zhijun Wang, Jiahuan Li, Hao Zhou, Rongxiang Weng, Jingang Wang, Xin Huang, Xue Han, Junlan Feng, Chao Deng, Shujian Huang

Large language models (LLMs) exhibit remarkable multilingual capabilities despite the extreme language imbalance in the pre-training data. In this paper, we closely examine the reasons behind this phenomenon, focusing on the pre-training corpus. We find that the existence of code-switching, alternating between different languages within a context, is key to multilingual capabilities. We conduct an analysis to investigate code-switching in the pre-training corpus, examining its presence and categorizing it into four types within two quadrants. We then assess its impact on multilingual performance. These types of code-switching data are unbalanced in proportions and demonstrate different effects on facilitating language transfer. To better explore the power of code-switching for language alignment during pre-training, we investigate the strategy of synthetic code-switching. We continuously scale up the synthetic code-switching data and observe remarkable improvements in both benchmarks and representation space. Extensive experiments indicate that incorporating synthetic code-switching data enables better language alignment and generalizes well to high, medium, and low-resource languages with pre-training corpora of varying qualities.

Moxin Li, Yuantao Zhang, Wenjie Wang, Wentao Shi, Zhuo Liu, Fuli Feng, Tat-Seng Chua

Multi-Objective Alignment (MOA) aims to align LLMs’ responses with multiple human preference objectives, with Direct Preference Optimization (DPO) emerging as a prominent approach. However, we find that DPO-based MOA approaches suffer from widespread preference conflicts in the data, where different objectives favor different responses. This results in conflicting optimization directions, hindering the optimization on the Pareto Front. To address this, we propose to construct Pareto-optimal responses to resolve preference conflicts. To efficiently obtain and utilize such responses, we propose a self-improving DPO framework that enables LLMs to self-generate and select Pareto-optimal responses for self-supervised preference alignment. Extensive experiments on two datasets demonstrate the superior Pareto Front achieved by our framework compared to various baselines

Xin Li, Mengbing Liu, Li Wei, Jiancheng An, Merouane Abdelkader Debbah, Chau Yuen

Large Language Models (LLMs) have achieved impressive results across a broad array of tasks, yet their capacity for complex, domain-specific mathematical reasoning—particularly in wireless communications—remains underexplored. In this work, we introduce WirelessMathBench, a novel benchmark specifically designed to evaluate LLMs on mathematical modeling challenges to wireless communications engineering. Our benchmark consists of 587 meticulously curated questions sourced from 40 state-of-the-art research papers, encompassing a diverse spectrum of tasks ranging from basic multiple-choice questions to complex equation completion tasks, including both partial and full completions, all of which rigorously adhere to physical and dimensional constraints. Through extensive experimentation with leading LLMs, we observe that while many models excel in basic recall tasks, their performance degrades significantly when reconstructing partially or fully obscured equations, exposing fundamental limitations in current LLMs. Even DeepSeek-R1, the best performer on our benchmark, achieves an average accuracy of only 38.05%, with a mere 7.83% success rate in full equation completion. By publicly releasing WirelessMathBench along with the evaluation toolkit, we aim to advance the development of more robust, domain-aware LLMs for wireless system analysis and broader engineering applications.

Anna Wegmann, Dong Nguyen, David Jurgens

Variation in language is ubiquitous and often systematically linked to regional, social, and contextual factors. Tokenizers split texts into smaller units and might behave differently for less common linguistic forms. This might affect downstream LLM performance differently on two types of tasks: Tasks where the model should be robust to language variation (e.g., for semantic tasks like NLI, labels do not depend on whether a text uses British or American spelling) and tasks where the model should be sensitive to language variation (e.g., for form-based tasks like authorship verification, labels depend on whether a text uses British or American spelling). We pre-train BERT base models with the popular Byte-Pair Encoding algorithm to investigate how key tokenization design choices impact the performance of downstream models: the corpus used to train the tokenizer, the pre-tokenizer and the vocabulary size. We find that the best tokenizer varies on the two task types and that the pre-tokenizer has the biggest overall impact on performance. Further, we introduce a new approach to estimate tokenizer impact on downstream LLM performance, showing substantial improvement over metrics like Rényi efficiency. We encourage more work on language variation and its relation to tokenizers and thus LLM performance.

Wenxuan Wang, Kuiyi Gao, Youliang Yuan, Jen-tse Huang, Qiuzhi Liu, Shuai Wang, Wenxiang Jiao, Zhaopeng Tu

Text-based image generation models, such as Stable Diffusion and DALL-E 3, hold significant potential in content creation and publishing workflows, making them the focus in recent years. Despite their remarkable capability to generate diverse and vivid images, considerable efforts are being made to prevent the generation of harmful content, such as abusive, violent, or pornographic material. To assess the safety of existing models, we introduce a novel jailbreaking method called Chain-of-Jailbreak (CoJ) attack, which compromises image generation models through a step-by-step editing process. Specifically, for malicious queries that cannot bypass the safeguards with a single prompt, we intentionally decompose the query into multiple sub-queries. The image generation models are then prompted to generate and iteratively edit images based on these sub-queries. To evaluate the effectiveness of our CoJ attack method, we constructed a comprehensive dataset, CoJ-Bench, including nine safety scenarios, three types of editing operations, and three editing elements. Experiments on four widely-used image generation services provided by GPT-4V, GPT-4o, Gemini 1.5 and Gemini 1.5 Pro, demonstrate that our CoJ attack method can successfully bypass the safeguards of models for over 60% cases, which significantly outperforms other jailbreaking methods (i.e., 14%). Further, to enhance these models’ safety against our CoJ attack method, we also propose an effective prompting-based method, Think-Twice Prompting, that can successfully defend over 95% of CoJ attack. Our dataset and code are included in the supplementary materials and will be made publicly available upon publication.

Naomi Baes, Raphael Merx, Nick Haslam, Ekaterina Vylomova, Haim Dubossarsky

Lexical Semantic Change (LSC) provides insight into cultural and social dynamics. Yet, the validity of methods for measuring different kinds of LSC remains unestablished due to the absence of historical benchmark datasets. To address this gap, we propose LSC-Eval, a novel three-stage general-purpose evaluation framework to: (1) develop a scalable methodology for generating synthetic datasets that simulate theory-driven LSC using In-Context Learning and a lexical database; (2) use these datasets to evaluate the sensitivity of computational methods to synthetic change; and (3) assess their suitability for detecting change in specific dimensions and domains. We apply LSC-Eval to simulate changes along the Sentiment, Intensity, and Breadth (SIB) dimensions, as defined in the SIBling framework, using examples from psychology. We then evaluate the ability of selected methods to detect these controlled interventions. Our findings validate the use of synthetic benchmarks, demonstrate that tailored methods effectively detect changes along SIB dimensions, and reveal that a state-of-the-art LSC model faces challenges in detecting affective dimensions of LSC. LSC-Eval offers a valuable tool for dimension- and domain-specific benchmarking of LSC methods, with particular relevance to the social sciences.

Marta R. Costa-jussà, Bokai Yu, Pierre Andrews, Belen Alastruey, Necati Cihan Camgoz, Joe Chuang, Jean Maillard, Christophe Ropers, Arina Turkatenko, Carleigh Wood

We introduce the first highly multilingual speech and American Sign Language (ASL) comprehension dataset by extending BELEBELE. Our dataset covers 91 spoken languages at the intersection of BELEBELE and FLEURS, and one sign language (ASL). As a by-product we also extend the Automatic Speech Recognition Benchmark, FLEURS, by 20%. We evaluate 2M-BELEBELE dataset for both 5-shot and zero-shot settings and across languages, the speech comprehension accuracy is ≈ 10% average lower compared to reading comprehension.

Yan Yang, Dongxu Li, Haoning Wu, Bei Chen, Liu Liu, Liyuan Pan, Junnan Li

Solving expert-level multimodal tasks is a key milestone in general intelligence. As the capabilities of multimodal large language models (MLLMs) continue to evolve, evaluation of frontier multimodal intelligence becomes necessary yet challenging. In this work, we introduce ProBench, a benchmark of open-ended user queries encapsulating professional expertise and advanced reasoning. ProBench consists of 4,000 high-quality samples independently collected from professionals based on their productivity demands. It spans across 10 fields and 56 sub-fields, including science, arts, humanities, coding, mathematics, and creative writing. Experimentally, we evaluate and compare 24 latest models using MLLM-as-a-Judge. Our results reveal that although the best open-source models rival the proprietary ones, they all face significant challenges in visual perception, textual understanding, domain knowledge, and advanced reasoning. Our benchmark is publicly accessible at TBC.