Committee

arXiv:2502.16903v2 Announce Type: replace Abstract: Despite the growing interest in jailbreak methods as an effective red-teaming tool for building safe and responsible large language models (LLMs), flawed evaluation system designs have led to significant discrepancies in their effectiveness assessments. We conduct a systematic measurement study based on 37 jailbreak studies since 2022, focusing on both the methods and the evaluation systems they employ. We find that existing evaluation systems lack case-specific criteria, resulting in misleading conclusions about their effectiveness and safety implications. This paper advocates a shift to a more nuanced, case-by-case evaluation paradigm. We introduce GuidedBench, a novel benchmark comprising a curated harmful question dataset, detailed case-by-case evaluation guidelines and an evaluation system integrated with these guidelines — GuidedEval. Experiments demonstrate that GuidedBench offers more accurate measurements of jailbreak performance, enabling meaningful comparisons across methods and uncovering new insights overlooked in previous evaluations. GuidedEval reduces inter-evaluator variance by at least 76.03%. Furthermore, we observe that incorporating guidelines can enhance the effectiveness of jailbreak methods themselves, offering new insights into both attack strategies and evaluation paradigms.

arXiv:2507.05557v1 Announce Type: new Abstract: Test-time scaling has emerged as a promising paradigm in language modeling, leveraging additional computational resources at inference time to enhance model performance. In this work, we introduce R2-LLMs, a novel and versatile hierarchical retrieval-augmented reasoning framework designed to improve test-time scaling in large language models (LLMs) without requiring distillation from more advanced models to obtain chain-of-thought (CoT) training data. R2-LLMs enhances inference-time generalization by integrating dual-level retrieval-based in-context learning: (1) At the coarse level, our approach extracts abstract templates from complex reasoning problems and retrieves similar problem-answer pairs to facilitate high-level in-context learning; (2) At the fine level, during Monte Carlo Tree Search (MCTS), R2-LLMs efficiently retrieves analogous intermediate solution steps from reference mathematical problem datasets, refining step-wise reasoning with the aid of a process reward model (PRM) for scoring. R2-LLMs is a robust hierarchical reasoning-augmentation method that enhances in-context-level reasoning while seamlessly integrating with step-level tree search methods. Utilizing PRM, it refines both candidate generation and decision-making for improved reasoning accuracy. Empirical evaluations on the MATH500, GSM8K, and OlympiadBench-TO datasets achieve substantial relative improvement with an increase of up to 16% using LLaMA-3.1-8B compared to the baselines, showcasing the effectiveness of our approach in complex reasoning tasks.

arXiv:2507.03483v2 Announce Type: replace Abstract: In this paper, we introduce BMMR, a large-scale bilingual, multimodal, multi-disciplinary reasoning dataset for the community to develop and evaluate large multimodal models (LMMs). BMMR comprises 110k college-level questions spanning 300 UNESCO-defined subjects, spanning diverse formats-multiple-choice, fill-in-the-blank, and open-ended QA-and sourced from both print and digital media such as books, exams, and quizzes. All data are curated and filtered via a human-in-the-loop and scalable framework, and each instance is paired with a high-quality reasoning path. The dataset is organized into two parts: BMMR-Eval that comprises 20,458 high-quality instances to comprehensively assess LMMs’ knowledge and reasoning across multiple disciplines in both Chinese and English; and BMMR-Train that contains 88,991 instances to support further research and development, extending the current focus on mathematical reasoning to diverse disciplines and domains. In addition, we propose the process-based multi-discipline verifier (i.e., BMMR-Verifier) for accurate and fine-grained evaluation of reasoning paths. Extensive experiments on 24 models reveal that (i) even SOTA models (e.g., o3 and Gemini-2.5-Pro) leave substantial headroom on BMMR-Eval; (ii) reasoning models exhibit discipline bias and outperform LMMs only on specific subjects; (iii) open-source models still trail their proprietary counterparts; and (iv) fine-tuning on BMMR-Train narrows this gap. Additionally, we conduct reasoning-chain analyses using BMMR-Verifier and other in-depth studies, uncovering the challenges LMMs currently face in multidisciplinary reasoning. We will release the data, and we hope our work can offer insights and contributions to the community.

arXiv:2503.13575v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) possess encompassing capabilities that can process diverse language-related tasks. However, finetuning on LLMs will diminish this general skills and continual finetuning will further cause severe degradation on accumulated knowledge. Recently, Continual Learning (CL) in Large Language Models (LLMs) arises which aims to continually adapt the LLMs to new tasks while maintaining previously learned knowledge and inheriting general skills. Existing techniques either leverage previous data to replay, leading to extra computational costs, or utilize a single parameter-efficient module to learn the downstream task, constraining new knowledge absorption with interference between different tasks. Toward these issues, this paper proposes Analytic Subspace Routing(ASR) to address these challenges. For each task, we isolate the learning within a subspace of deep layers’ features via low-rank adaptation, eliminating knowledge interference between different tasks. Additionally, we propose an analytic routing mechanism to properly utilize knowledge learned in different subspaces. Our approach employs Recursive Least Squares to train a multi-task router model, allowing the router to dynamically adapt to incoming data without requiring access to historical data. Also, the router effectively assigns the current task to an appropriate subspace and has a non-forgetting property of previously learned tasks with a solid theoretical guarantee. Experimental results demonstrate that our method achieves near-perfect retention of prior knowledge while seamlessly integrating new information, effectively overcoming the core limitations of existing methods. Our code will be released after acceptance.

arXiv:2507.05939v1 Announce Type: new Abstract: Nowadays, misinformation articles, especially multimodal ones, are widely spread on social media platforms and cause serious negative effects. To control their propagation, Multimodal Misinformation Detection (MMD) becomes an active topic in the community to automatically identify misinformation. Previous MMD methods focus on supervising detectors by collecting offline data. However, in real-world scenarios, new events always continually emerge, making MMD models trained on offline data consistently outdated and ineffective. To address this issue, training MMD models under online data streams is an alternative, inducing an emerging task named continual MMD. Unfortunately, it is hindered by two major challenges. First, training on new data consistently decreases the detection performance on past data, named past knowledge forgetting. Second, the social environment constantly evolves over time, affecting the generalization on future data. To alleviate these challenges, we propose to remember past knowledge by isolating interference between event-specific parameters with a Dirichlet process-based mixture-of-expert structure, and anticipate future environmental distributions by learning a continuous-time dynamics model. Accordingly, we induce a new continual MMD method DAEDCMD. Extensive experiments demonstrate that DAEDCMD can consistently and significantly outperform the compared methods, including six MMD baselines and three continual learning methods.

arXiv:2507.05305v1 Announce Type: cross Abstract: Frontier Large language models (LLMs) like ChatGPT and Gemini can decipher cryptic compiler errors for novice programmers, but their computational scale, cost, and tendency to over-assist make them problematic for widespread pedagogical adoption. This work demonstrates that smaller, specialised language models, enhanced via Supervised Fine-Tuning (SFT), present a more viable alternative for educational tools. We utilise a new dataset of 40,000 C compiler error explanations, derived from real introductory programming (CS1/2) student-generated programming errors, which we used to fine-tune three open-source models: Qwen3-4B, Llama-3.1-8B, and Qwen3-32B. We performed a dual evaluation, combining expert human reviews with a large-scale automated analysis of 8,000 responses using a validated LLM-as-judge ensemble. Our results show that SFT significantly boosts the pedagogical quality of smaller models, achieving performance comparable to much larger models. We analyse the trade-offs between model size and quality, confirming that fine-tuning compact, efficient models on high-quality, domain-specific data is a potent strategy for creating specialised models to drive educational tools. We provide a replicable methodology to foster broader access to generative AI capabilities in educational contexts.

arXiv:2507.03473v1 Announce Type: new Abstract: Despite mounting evidence that multilinguality can be easily weaponized against language models (LMs), works across NLP Security remain overwhelmingly English-centric. In terms of securing LMs, the NLP norm of “English first” collides with standard procedure in cybersecurity, whereby practitioners are expected to anticipate and prepare for worst-case outcomes. To mitigate worst-case outcomes in NLP Security, researchers must be willing to engage with the weakest links in LM security: lower-resourced languages. Accordingly, this work examines the security of LMs for lower- and medium-resourced languages. We extend existing adversarial attacks for up to 70 languages to evaluate the security of monolingual and multilingual LMs for these languages. Through our analysis, we find that monolingual models are often too small in total number of parameters to ensure sound security, and that while multilinguality is helpful, it does not always guarantee improved security either. Ultimately, these findings highlight important considerations for more secure deployment of LMs, for communities of lower-resourced languages.

arXiv:2506.21571v2 Announce Type: replace Abstract: Large Reasoning Models (LRMs), which autonomously produce a reasoning Chain of Thought (CoT) before producing final responses, offer a promising approach to interpreting and monitoring model behaviors. Inspired by the observation that certain CoT patterns — e.g., “Wait, did I miss anything?” — consistently emerge across tasks, we explore whether LRMs exhibit human-like cognitive habits. Building on Habits of Mind, a well-established framework of cognitive habits associated with successful human problem-solving, we introduce CogTest, a principled benchmark designed to evaluate LRMs’ cognitive habits. CogTest includes 16 cognitive habits, each instantiated with 25 diverse tasks, and employs an evidence-first extraction method to ensure reliable habit identification. With CogTest, we conduct a comprehensive evaluation of 16 widely used LLMs (13 LRMs and 3 non-reasoning ones). Our findings reveal that LRMs, unlike conventional LLMs, not only exhibit human-like habits but also adaptively deploy them according to different tasks. Finer-grained analyses further uncover patterns of similarity and difference in LRMs’ cognitive habit profiles, particularly certain inter-family similarity (e.g., Qwen-3 models and DeepSeek-R1). Extending the study to safety-related tasks, we observe that certain habits, such as Taking Responsible Risks, are strongly associated with the generation of harmful responses. These findings suggest that studying persistent behavioral patterns in LRMs’ CoTs is a valuable step toward deeper understanding of LLM misbehavior. The code is available at: https://github.com/jianshuod/CogTest.

arXiv:2411.04109v3 Announce Type: replace Abstract: Self-alignment, whereby models learn to improve themselves without human annotation, is a rapidly growing research area. However, existing techniques often fail to improve complex reasoning tasks due to the difficulty of assigning correct rewards. An orthogonal approach that is known to improve correctness is self-consistency, a method applied at inference time based on multiple sampling in order to find the most consistent answer. In this work, we extend the self-consistency concept to help train models. We thus introduce self-consistency preference optimization (ScPO), which iteratively trains consistent answers to be preferred over inconsistent ones on unsupervised new problems. We show ScPO leads to large improvements over conventional reward model training on reasoning tasks such as GSM8K and MATH, closing the gap with supervised training with gold answers or preferences, and that combining ScPO with standard supervised learning improves results even further. On ZebraLogic, ScPO finetunes Llama-3 8B to be superior to Llama-3 70B, Gemma-2 27B, and Claude-3 Haiku.

arXiv:2507.03875v1 Announce Type: new Abstract: The introduction of ChatGPT has garnered significant attention within the NLP community and beyond. Previous studies have demonstrated ChatGPT’s substantial advancements across various downstream NLP tasks, highlighting its adaptability and potential to revolutionize language-related applications. However, its capabilities and limitations in genre prediction remain unclear. This work analyzes three Large Language Models (LLMs) using the MovieLens-100K dataset to assess their genre prediction capabilities. Our findings show that ChatGPT, without fine-tuning, outperformed other LLMs, and fine-tuned ChatGPT performed best overall. We set up zero-shot and few-shot prompts using audio transcripts/subtitles from movie trailers in the MovieLens-100K dataset, covering 1682 movies of 18 genres, where each movie can have multiple genres. Additionally, we extended our study by extracting IMDb movie posters to utilize a Vision Language Model (VLM) with prompts for poster information. This fine-grained information was used to enhance existing LLM prompts. In conclusion, our study reveals ChatGPT’s remarkable genre prediction capabilities, surpassing other language models. The integration of VLM further enhances our findings, showcasing ChatGPT’s potential for content-related applications by incorporating visual information from movie posters.