Multiple-choice questions (MCQs) play a crucial role in fostering deep thinking and knowledge integration in education. However, previous research has primarily focused on generating MCQs with textual options, but it largely overlooks the visual options. Moreover, generating high-quality distractors remains a major challenge due to the high cost and limited scalability of manual authoring. To tackle these problems, we propose a Cross-modal Options Synthesis (CmOS), a novel framework for generating educational MCQs with visual options. Our framework integrates Multimodal Chain-of-Thought (MCoT) reasoning process and Retrieval-Augmented Generation (RAG) to produce semantically plausible and visually similar answer and distractor. It also includes a discrimination module to identify content suitable for visual options. Experimental results on test tasks demonstrate the superiority of CmOS in content discrimination, question generation and visual option generation over existing methods across various subjects and educational levels.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Large Language Models (LLMs) often exhibit gender bias, resulting in unequal treatment of male and female subjects across different contexts. To address this issue, we propose a novel data generation framework that fosters exploratory thinking in LLMs. Our approach prompts models to generate story pairs featuring male and female protagonists in structurally identical, morally ambiguous scenarios, then elicits and compares their moral judgments. When inconsistencies arise, the model is guided to produce balanced, gender-neutral judgments. These story-judgment pairs are used to fine-tune or optimize the models via Direct Preference Optimization (DPO). Experimental results show that our method significantly reduces gender bias while preserving or even enhancing general model capabilities. We will release the code and generated data.
Conv-Basis: A New Paradigm for Efficient Attention Inference and Gradient Computation in Transformers
PDF ↗The self-attention mechanism is key to the success of transformers in recent large language models (LLMs). However, the quadratic computational cost, O(n^2), with respect to the input sequence length n poses a significant obstacle to further improvement and scalability in longer contexts.In this work, we leverage the convolution-like structure of attention matrices to develop an efficient approximation method for attention computation using convolution matrices. We propose a \mathsf{conv} basis system, analogous to the rank basis, and show that any lower triangular matrix can be decomposed as a sum of structured convolution matrices in this basis. We then design a fast algorithm to approximate the attention matrix using a sum of k convolution matrices. This enables us to compute attention during inference via Fast Fourier Transforms (FFT) in O(knd \log n) time, where d is the hidden dimension, achieving nearly linear time complexity, n^{1+o(1)}, in practical scenarios where kd = n^{o(1)}. Furthermore, both training forward and backward gradient computations can be performed in n^{1+o(1)} time as well.We provide theoretical guarantees on runtime and approximation error and conduct preliminary experiments to evaluate the effectiveness of our approach. We hope this new paradigm for accelerating attention computation in transformer models facilitates their application to longer contexts.
Generative large language models ( LLMs) have achieved remarkable success in various industrial applications, owing to their promising In-Context Learning capabilities. However, the issue of long context in complex tasks poses a significant barrier to their wider adoption, manifested in two main aspects: (i) The excessively long context leads to high costs and inference delays. (ii) A substantial amount of task-irrelevant information introduced by long contexts exacerbates the “lost in the middle” problem. Existing methods compress context by removing redundant tokens using metrics such as self-information or perplexity ( PPL ), which is inconsistent with the objective of retaining the most important tokens when conditioning on a given query. In this study, we introduce information bottleneck theory (IB) to model the problem, offering a novel perspective that thoroughly addresses the essential properties required for context compression. Additionally, we propose a cross-attention-based approach to approximate mutual information in IB, which can be flexibly replaced with suitable alternatives in different scenarios. Extensive experiments on four datasets demonstrate that our method achieves a 25% increase in compression rate compared to the state-of-the-art, while maintaining question answering performance. In particular, the context compressed by our method even outperform the full context in some cases.
While the Mamba architecture demonstrates superior inference efficiency and competitive performance on short-context natural language processing (NLP) tasks, empirical evidence suggests its capacity to comprehend long contexts is limited compared to transformer-based models. In this study, we investigate the long-context efficiency issues of the Mamba models and propose ReMamba, which enhances Mamba’s ability to comprehend long contexts. ReMamba incorporates selective compression and adaptation techniques within a two-stage re-forward process, incurring minimal additional inference costs overhead. Experimental results on the LongBench and L-Eval benchmarks demonstrate ReMamba’s efficacy, improving over the baselines by 3.2 and 1.6 points, respectively, and attaining performance almost on par with same-size transformer models.
HEAL: Hybrid Enhancement with LLM-based Agents for Text-attributed Hypergraph Self-supervised Representation Learning
PDF ↗This paper studies the problem of text-attributed hypergraph self-supervised representation learning, which aims to generate discriminative representations of hypergraphs without any annotations for downstream tasks. However, real-world hypergraphs could contain incomplete signals, which could deteriorate the representation learning procedure, especially under label scarcity. Towards this end, we introduce a new perspective that leverages large language models to enhance hypergraph self-supervised learning and propose a novel data-centric approach named Hybrid Hypergraph Enhancement with LLM-based Agents (HEAL). The core of our HEAL is to generate informative nodes and hyperedges through multi-round interaction with LLM-based agents. In particular, we first retrieve similar samples for each node to facilitate the node expansion agent for different views. To generate challenging samples, we measure the gradients for each augmented view and select the most informative one using an evaluation agent. From the structural view, we adopt a topology refinement agent to incorporate new hyperedges for the recovery of missing structural signals. The enhanced hypergraphs would be incorporated into a self-supervised learning framework for discriminative representations. Extensive experiments on several datasets validate the effectiveness of our HEAL in comparison with extensive baselines.
Controversial discourse is abundant on social media. Understanding how controversial problems are framed in online discourse is crucial for gaining insights into public opinion formation and for addressing misinformation and polarization. This paper presents a novel method for discovering and articulating framing of controversial problems, enabling the investigation of how controversy is framed across several diverse topics. The promising results, made possible by recent advances in Large Language Models, indicate that discovering framings across topics is feasible. The discovered frames offer valuable insights into how and why controversial problems are discussed on social media.
The use of large language models (LLMs) as judges, particularly in preference comparisons, has become widespread, but this reveals a notable bias towards longer responses, undermining the reliability of such evaluations. To better understand such bias, we propose to decompose the preference evaluation metric, specifically the win rate, into two key components: desirability and information mass, where the former is length-independent and related to trustworthiness such as correctness, toxicity, and consistency, and the latter is length-dependent and represents the amount of information in the response. We empirically demonstrated the decomposition through controlled experiments and found that response length impacts evaluations by influencing information mass. To derive a reliable evaluation metric that assesses content quality without being confounded by response length, we propose AdapAlpaca, a simple yet effective adjustment to win rate measurement. Specifically, AdapAlpaca ensures a fair comparison of response quality by aligning the lengths of reference and test model responses under equivalent length intervals.
Large language models (LLMs) have achieved remarkable success across various natural language processing (NLP) tasks. However, recent studies suggest that they still face challenges in performing fundamental NLP tasks essential for deep language understanding, particularly syntactic parsing. In this paper, we conduct an in-depth analysis of LLM parsing capabilities, delving into the underlying causes of why LLMs struggle with this task and the specific shortcomings they exhibit. We find that LLMs may be limited in their ability to fully leverage grammar rules from existing treebanks, restricting their capability to generate syntactic structures. To help LLMs acquire knowledge without additional training, we propose a self-correction method that leverages grammar rules from existing treebanks to guide LLMs in correcting previous errors. Specifically, we automatically detect potential errors and dynamically search for relevant rules, offering hints and examples to guide LLMs in making corrections themselves. Experimental results on three datasets using various LLMs demonstrate that our method significantly improves performance in both in-domain and cross-domain settings.
The application of large language models (LLMs) in healthcare holds significant promise for enhancing clinical decision-making, medical research, and patient care. However, their integration into real-world clinical settings raises critical concerns around trustworthiness, particularly around dimensions of truthfulness, privacy, safety, robustness, fairness, and explainability. These dimensions are essential for ensuring that LLMs generate reliable, unbiased, and ethically sound outputs. While researchers have recently begun developing benchmarks and evaluation frameworks to assess LLM trustworthiness, the trustworthiness of LLMs in healthcare remains underexplored, lacking a systematic review that provides a comprehensive understanding and future insights. This survey addresses that gap by providing a comprehensive review of current methodologies and solutions aimed at mitigating risks across key trust dimensions. We analyze how each dimension affects the reliability and ethical deployment of healthcare LLMs, synthesize ongoing research efforts and identify critical gaps in existing approaches. We also identify emerging challenges posed by evolving paradigms, such as multi-agent collaboration, multi-modal reasoning, and the development of small open-source medical models. Our goal is to guide future research toward more trustworthy, transparent, and clinically viable LLMs.
Aspect-Opinion Pair Extraction (AOPE) and Aspect Sentiment Triplet Extraction (ASTE) have drawn growing attention in NLP. However, most existing approaches extract aspects and opinions independently, optionally adding pairwise relations, often leading to error propagation and high time complexity. To address these challenges and being inspired by transition-based dependency parsing, we propose the first transition-based model for AOPE and ASTE that performs aspect and opinion extraction jointly, which also better captures position-aware aspect-opinion relations and mitigates entity-level bias. By integrating contrastive-augmented optimization, our model delivers more accurate action predictions and jointly optimizes separate subtasks in linear time. Extensive experiments on four commonly used ASTE/AOPE datasets show that, our proposed transition-based model outperform previous models on two out of the four datasets when trained on a single dataset. When multiple training sets are used, our proposed method achieves new state-of-the-art results on all datasets. We show that this is partly due to our model’s ability to benefit from transition actions learned from multiple datasets and domains.Our code is available at https://github.com/Paparare/trans_aste.
Safety-aligned LLMs respond to prompts with either compliance or refusal, each corresponding to distinct directions in the model’s activation space. Recent studies have shown that initializing attacks via self-transfer from other prompts significantly enhances their performance. However, the underlying mechanisms of these initializations remain unclear, and attacks utilize arbitrary or hand-picked initializations. This work presents that each gradient-based jailbreak attack and subsequent initialization gradually converge to a single compliance direction that suppresses refusal, thereby enabling an efficient transition from refusal to compliance. Based on this insight, we propose CRI, an initialization framework that aims to project unseen prompts further along compliance directions. We demonstrate our approach on multiple attacks, models, and datasets, achieving an increased attack success rate (ASR) and reduced computational overhead, highlighting the fragility of safety-aligned LLMs.
Is Safety Standard Same for Everyone? User-Specific Safety Evaluation of Large Language Models
PDF ↗As the use of large language model (LLM) agents continues to grow, their safety vulnerabilities have become increasingly evident. Extensive benchmarks evaluate various aspects of LLM safety by defining the safety relying heavily on general standards, overlooking user-specific standards. However, safety standards for LLM may vary based on a user-specific profiles rather than being universally consistent across all users. This raises a critical research question: Do LLM agents act safely when considering user-specific safety standards? Despite its importance for safe LLM use, no benchmark datasets currently exist to evaluate the user-specific safety of LLMs. To address this gap, we introduce U-SafeBench, a benchmark designed to assess user-specific aspect of LLM safety. Our evaluation of 20 widely used LLMs reveals current LLMs fail to act safely when considering user-specific safety standards, marking a new discovery in this field. To address this vulnerability, we propose a simple remedy based on chain-of-thought, demonstrating its effectiveness in improving user-specific safety.
To investigate the effect of problem presentation on LLMs’ ability to solve optimization problems, we introduce the dataset of Everyday Hard Optimization Problems (EHOP), a collection of NP-hard problems expressed in natural language. EHOP includes problem formulations that could be found in computer science textbooks (e.g., graph coloring), versions that are dressed up as problems that could arise in real life (e.g., party planning), and variants with inverted rules. We find that state-of-the-art LLMs, across multiple prompting strategies, systematically solve textbook problems more accurately than their real-life and inverted counterparts. While reasoning models are more capable, they nonetheless show high variance across problem presentations, suggesting they lack a truly robust reasoning mechanism. We argue that this constitutes evidence that LLMs are still heavily dependent on what was seen in training and struggle to generalize to novel problems.
Patients must possess the knowledge necessary to actively participate in their care. To this end, we developed NoteAid-Chatbot, a conversational AI designed to help patients better understand their health through a novel framework of learning as conversation. We introduce a new learning paradigm that leverages a multi-agent large language model (LLM) and reinforcement learning (RL) framework—without relying on costly human-generated training data. Specifically, NoteAid-Chatbot was built on a lightweight 3-billion-parameter LLaMA 3.2 model using a two-stage training approach: initial supervised fine-tuning on conversational data synthetically generated using medical conversation strategies, followed by RL with rewards derived from patient understanding assessments in simulated hospital discharge scenarios. Our evaluation, which includes comprehensive human-aligned assessments and case studies, demonstrates that NoteAid-Chatbot exhibits key emergent behaviors critical for patient education—such as clarity, relevance, and structured dialogue—even though it received no explicit supervision for these attributes. Our results show that even simple Proximal Policy Optimization (PPO)-based reward modeling can successfully train lightweight, domain-specific chatbots to handle multi-turn interactions, incorporate diverse educational strategies, and meet nuanced communication objectives. Our Turing test demonstrates that NoteAid-Chatbot surpasses non-expert human. Although our current focus is on healthcare, the framework we present illustrates the feasibility and promise of applying low-cost, PPO-based RL to realistic, open-ended conversational domains—broadening the applicability of RL-based alignment methods.
MedCOD: Enhancing English-to-Spanish Medical Translation of Large Language Models Using Enriched Chain-of-Dictionary Framework
PDF ↗We present MedCOD (Medical Chain-of-Dictionary), a hybrid framework designed to improve English-to-Spanish medical translation by integrating domain-specific structured knowledge into large language models (LLMs). MedCOD integrates domain-specific knowledge from both the Unified Medical Language System (UMLS) and the LLM-as-Knowledge-Base (LLM-KB) paradigm to enhance structured prompting and fine-tuning. We constructed a parallel corpus of 2,999 English-Spanish MedlinePlus articles and a 100-sentence test set annotated with structured medical contexts. Four open-source LLMs (Phi-4, Qwen2.5-14B, Qwen2.5-7B, and LLaMA-3.1-8B) were evaluated using structured prompts that incorporated multilingual variants, medical synonyms, and UMLS-derived definitions, combined with LoRA-based fine-tuning. Experimental results demonstrate that MedCOD significantly improves translation quality across all models. For example, Phi-4 with MedCOD and fine-tuning achieved BLEU 44.23, chrF++ 28.91, and COMET 0.863, surpassing strong baseline models like GPT-4o and GPT-4o-mini. Ablation studies confirm that both MedCOD prompting and model adaptation independently contribute to performance gains, with their combination yielding the highest improvements. These findings highlight the potential of structured knowledge integration to enhance LLMs for medical translation tasks.
Bridging the Dynamic Perception Gap: Training-Free Draft Chain-of-Thought for Dynamic Multimodal Spatial Reasoning
PDF ↗While chains-of-thought (CoT) have advanced complex reasoning in multimodal large language models (MLLMs), existing methods remain confined to text or static visual domains, often faltering in dynamic spatial reasoning tasks. To bridge this gap, we present GRASSLAND, a novel maze navigation benchmark designed to evaluate dynamic spatial reasoning. Our experiments show that augmenting textual reasoning chains with dynamic visual drafts, overlaid on input images, significantly outperforms conventional approaches, offering new insights into spatial reasoning in evolving environments. To generalize this capability, we propose D2R (Dynamic Draft-Augmented Reasoning), a training-free framework that seamlessly integrates textual CoT with corresponding visual drafts into MLLMs. Extensive evaluations demonstrate that D2R consistently enhances performance across diverse tasks, establishing a robust baseline for dynamic spatial reasoning without requiring model fine-tuning.
How do Language Models Generate Slang: A Systematic Comparison between Human and Machine-Generated Slang Usages
PDF ↗Slang is a commonly used type of informal language that poses a daunting challenge to NLP systems. Recent advances in large language models (LLMs), however, have made the problem more approachable. While LLM agents are becoming more widely applied to intermediary tasks such as slang detection and slang interpretation, their generalizability and reliability are heavily dependent on whether these models have captured structural knowledge about slang that align well with human attested slang usages. To answer this question, we contribute a systematic comparison between human and machine-generated slang usages. Our evaluative framework focuses on three core aspects: 1) Characteristics of the usages that reflect systematic biases in how machines perceive slang, 2) Creativity reflected by both lexical coinages and word reuses employed by the slang usages, and 3) Informativeness of the slang usages when used as gold-standard examples for model distillation. By comparing human-attested slang usages from the Online Slang Dictionary (OSD) and slang generated by GPT-4o and Llama-3, we find significant biases in how LLMs perceive slang. Our results suggest that while LLMs have captured significant knowledge about the creative aspects of slang, such knowledge does not align with humans sufficiently to enable LLMs for extrapolative tasks such as linguistic analyses.
Unlearning aims to remove copyrighted, sensitive, or private content from large language models (LLMs) without a full retraining. In this work, we develop a multi-task unlearning benchmark LUME that features three tasks: (1) unlearn synthetically generated creative short novels, (2) unlearn synthetic biographies with sensitive information, and (3) unlearn a collection of public biographies. We further release two fine-tuned LLMs of 1B and 7B parameter sizes as the target models. We conduct detailed evaluations of several recently-proposed algorithms and present results on carefully crafted metrics to understand their behavior and limitations.
When Inverse Data Outperforms: Exploring the Pitfalls of Mixed Data in Multi-Stage Fine-Tuning
PDF ↗Existing work has shown that o1-level performance can be achieved with limited data distillation, but most existing methods focus on unidirectional supervised fine-tuning (SFT), overlooking the intricate interplay between diverse reasoning patterns. In this paper, we construct r1k, a high-quality reverse reasoning dataset derived by inverting 1,000 forward examples from s1k, and examine how SFT and Direct Preference Optimization (DPO) affect alignment under bidirectional reasoning objectives. SFT on r1k yields a 1.6%–6.8% accuracy improvement over s1k across evaluated benchmarks. However, naively mixing forward and reverse data during SFT weakens the directional distinction. Although DPO can partially recover this distinction, it also suppresses less preferred reasoning paths by shifting the probability mass toward irrelevant outputs. These findings suggest that mixed reasoning data introduce conflicting supervision signals, underscoring the need for robust and direction-aware alignment strategies. Our code and data are available at: https://github.com/16demi/ReasonAlign-analysis.