Research papers on LLM
Quickly access the latest research papers on large language models from arXiv.
Papers and research related to Large Language Model
- arXiv - Artificial Intelligence·Mar 31
QuestBench Reveals LLMs' Struggles in Asking Clarifying Questions for Reasoning Tasks
Description
QuestBench evaluates large language models' (LLMs) ability to ask clarifying questions in underspecified reasoning tasks, revealing significant challenges in information acquisition despite strong performance in well-defined problems.
Key Points
- QuestBench formalizes underspecified reasoning tasks as a constraint satisfaction problem, focusing on the ability of LLMs to identify necessary questions for problem-solving.
- The benchmark includes various tasks like Logic-Q and Planning-Q, where LLMs show only 40-50% accuracy, indicating difficulties in asking the right questions.
- Analysis reveals that even proficient models struggle with information acquisition, often failing to hedge their responses when uncertain, highlighting a critical area for further research.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Unicorn Framework: Cost-Effective Text-Only Data Synthesis for Vision-Language Models
Description
The paper presents Unicorn, a novel framework for synthesizing multimodal training data for vision-language models (VLMs) using only text, eliminating the need for costly image-text pairs.
Key Points
- Unicorn employs a three-stage process to generate two datasets: Unicorn-1.2M for pretraining and Unicorn-471K-Instruction for instruction-tuning, enhancing data diversity and quality.
- The first stage synthesizes 1.2 million diverse captions from sparse seeds using large language models, while the second stage creates multi-turn instruction tasks from 471,000 captions.
- The final stage transforms textual representations into visual representations, allowing for effective training of VLMs without relying on real images, thus providing a scalable and cost-effective solution.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Agent-Centric Framework Enhances Personalized Clustering with Multi-Modal LLMs
Description
The paper presents an agent-centric personalized clustering framework utilizing multi-modal large language models (MLLMs) to enhance user-specific clustering by effectively traversing relational graphs for better alignment with user interests.
Key Points
- The proposed framework improves upon traditional CLIP-based methods by leveraging MLLMs for deeper contextual understanding of user preferences in clustering tasks.
- By constructing a relational graph with user-interest-biased embeddings, the method reduces computational overhead and enhances the efficiency of cluster searches.
- Experimental results demonstrate significant improvements in clustering performance, achieving NMI scores of 0.9667 and 0.9481, surpassing the state-of-the-art model by over 140%.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Multimodal Language Models: Enhancing Visual Assistance for the Visually Impaired
Description
This paper evaluates the use of Multimodal Large Language Models (MLLMs) as assistive technologies for visually impaired users, identifying challenges and areas for improvement in their contextual understanding and usability.
Key Points
- A user survey reveals high adoption rates of MLLMs among visually impaired individuals, but highlights issues with contextual understanding and cultural sensitivity in visual interpretation tasks.
- The study introduces five user-centered tasks, including Optical Braille Recognition, to systematically evaluate the performance of twelve MLLMs in real-world scenarios.
- Findings indicate the need for advancements in multilingual support, Braille comprehension, and assistive object recognition to enhance the reliability and inclusivity of visual assistance technologies.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Enhancing Irony Detection in Historical Texts Using Large Language Models
Description
This study investigates the application of large language models (LLMs) for irony detection in 19th-century Latin American newspapers, focusing on enhancing datasets and improving classification tasks.
Key Points
- The research employs BERT and GPT-4o models to analyze irony through multi-class and binary classification, revealing challenges in capturing subtle nuances.
- A semi-automated annotation process successfully addresses class imbalance, enhancing the dataset with high-quality annotations crucial for historical language analysis.
- The study contributes a new historical Spanish dataset for sentiment analysis and proposes a methodology that integrates human expertise to refine LLM results, emphasizing historical and cultural contexts.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Enhancing Multilingual Capabilities in Visual Language Models to Overcome Fidelity Loss
Description
This research paper presents a method to enhance multilingual capabilities in Visual Language Models (VLMs) by addressing Image-induced Fidelity Loss (IFL) through a continuous multilingual integration strategy.
Key Points
- The study identifies Image-induced Fidelity Loss (IFL) as a challenge in VLMs, where models generate responses primarily in English despite multilingual inputs.
- A continuous multilingual integration strategy is proposed, which incorporates text-only multilingual data during visual instruction tuning, improving linguistic fidelity without compromising visual performance.
- The research also examines model merging techniques that enhance language fidelity but may reduce visual performance, contrasting with the core method that maintains robust multilingual alignment without trade-offs.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Niyama: A Game-Changer in Efficient LLM Inference Serving
Description
Niyama is a novel inference serving system designed to optimize resource utilization for Large Language Models (LLMs) by enabling fine-grained Quality-of-Service (QoS) differentiation and efficient workload co-scheduling.
Key Points
- Niyama addresses inefficiencies in existing LLM serving frameworks by allowing applications to specify precise latency requirements, improving resource management during traffic surges.
- The system employs a dynamic chunking mechanism to enhance throughput while ensuring strict QoS guarantees, significantly increasing serving capacity by 32% over siloed deployments.
- Niyama's hybrid prioritization policy balances fairness and efficiency, reducing Service Level Objective (SLO) violations by an order of magnitude under extreme load conditions, demonstrating its effectiveness in real-world scenarios.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Enhancing LLMs: New Method Unlocks Multi-Modal Generative Capabilities
Description
This research paper presents a method to enhance pre-trained text-only large language models (LLMs) with multi-modal generation capabilities while maintaining original performance and efficiency through parameter redundancy exploitation.
Key Points
- The proposed method utilizes Mixture-of-Experts (MoEs) to leverage parameter redundancy, allowing for the addition of new modalities without significantly increasing the model size.
- A novel low-rank adaptation technique is applied exclusively to new modality tokens, preserving the original language generation capabilities while ensuring scalability.
- The introduction of a Gromov-Wasserstein distance-based parameter initialization scheme enhances convergence and training stability, facilitating the transition from uni-modal to multi-modal architectures.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
Survey on Evaluating LLM Agents in Multi-Turn Conversations Establishes Comprehensive Framework
Description
This survey evaluates methods for assessing LLM-based agents in multi-turn conversations, establishing a comprehensive framework that categorizes evaluation components and methodologies for effective analysis.
Key Points
- The study reviews nearly 250 scholarly sources, creating a structured approach to evaluate LLM-based agents in conversational settings using a PRISMA-inspired framework.
- Two taxonomy systems are developed: one for identifying key evaluation components like task completion and user experience, and another for categorizing evaluation methodologies.
- The evaluation methodologies include annotation-based evaluations, automated metrics, hybrid strategies, and self-judging methods, capturing both traditional and advanced metrics relevant to multi-turn dialogues.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
EGSW: Enhancing RL-Based LLM Fine-Tuning Through Dynamic Sequence Weighting
Description
The paper presents Entropy-Guided Sequence Weighting (EGSW), a method that improves exploration-exploitation in Reinforcement Learning for fine-tuning Large Language Models by dynamically weighting outputs based on their advantage and entropy.
Key Points
- EGSW combines entropy regularization with advantage-based weighting to enhance policy updates, facilitating efficient exploration in complex state spaces.
- The method employs temperature-scaled softmax weighting to prioritize high-reward, high-uncertainty outputs while ensuring training stability during fine-tuning.
- Empirical results show that EGSW significantly improves sample efficiency and reasoning ability in Group Relative Policy Optimization, with potential applications in other RL algorithms.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
CoSIL: A Breakthrough LLM Method for Enhanced Software Issue Localization
Description
CoSIL is a novel LLM-driven method for software issue localization that enhances the accuracy of patch generation by dynamically constructing function call graphs without pre-parsing, achieving significant performance improvements over existing methods.
Key Points
- CoSIL utilizes module call graphs to iteratively search for relevant contexts, effectively reducing the search space for issue localization in software engineering.
- The method achieves a Top-1 localization success rate of 43% and 44.6% on SWE bench Lite and Verified, outperforming previous techniques by up to 98.2%.
- When integrated into the patch generation process, CoSIL improves the resolved rate by 9.3% to 31.5%, demonstrating its effectiveness in enhancing software repair tasks.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 31
LLMs Enhance Typosquatting Detection, Achieving 98% Accuracy in Cybersecurity
Description
This research presents a novel approach to typosquatting detection using large language models (LLMs), enhancing cybersecurity by addressing sophisticated domain-based deception tactics.
Key Points
- Typosquatting exploits human typing errors, leading to significant cybersecurity threats, which traditional detection methods struggle to address effectively.
- The study introduces a character-level transformation training method for LLMs, resulting in a more adaptable detection mechanism for complex typosquatting attacks.
- Experimental results show that the Phi-4 14B model achieved a 98% accuracy rate with minimal training samples, demonstrating the effectiveness of LLMs in cybersecurity applications.
linkCopy linkshare_windowsShare
