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 7
Identifying Reasoning Limitations in Large Language Models with Math-RoB Benchmark
Description
This paper identifies critical limitations in the reasoning robustness of large language models (LLMs), revealing performance drops on novel data and introducing a benchmark, Math-RoB, to assess these challenges.
Key Points
- The study uncovers four main limitations affecting LLMs: positional bias, instruction sensitivity, numerical fragility, and memory dependence, leading to reliance on memorized patterns.
- Performance metrics show significant accuracy drops in models like GPT-4o and Qwen2.5 Series when faced with incomplete data or specific instructions, highlighting reasoning gaps.
- The Math-RoB benchmark is introduced to evaluate reasoning robustness by generating diverse datasets that mimic training distributions, aiming to improve LLM reasoning frameworks.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
Transferring Safety Interventions Between Large Language Models: A New Approach
Description
This research explores the transferability of safety interventions between Large Language Models (LLMs) by utilizing shared activation spaces, demonstrating effective methods for backdoor removal and harmful prompt refusal.
Key Points
- The study shows that safety interventions can be transferred between models, enhancing their alignment and safety through learned mappings of activation spaces.
- A new task, 'corrupted capabilities,' is introduced to assess models' ability to distinguish useful skills from backdoors, reflecting real-world challenges.
- Experiments across Llama, Qwen, and Gemma models reveal that smaller models can effectively align larger ones, with autoencoder mappings acting as dynamic safety switches for model behavior.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
Innovative VLM-Based System Enhances Collaborative CAD Design Processes
Description
This paper presents a Vision Language Model (VLM)-based Multi-Agent System designed to automate the CAD modeling process, enhancing collaborative design in industrial and hobbyist contexts.
Key Points
- The system mimics a team structure by integrating agents for requirements engineering, CAD modeling, and quality assurance, streamlining the design process.
- It allows for automatic model generation from sketches or textual descriptions, facilitating iterative refinement through user collaboration.
- The architecture's effectiveness is demonstrated through various design tasks, showcasing the advantages of its individual components in enhancing design efficiency.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
AgentSafe Framework Enhances Security for Large Language Model-based Multi-Agent Systems
Description
AgentSafe is a framework designed to enhance the security of Large Language Model-based multi-agent systems by implementing hierarchical data management and memory protection strategies to mitigate security threats.
Key Points
- AgentSafe classifies information by security levels, ensuring that sensitive data is only accessible to authorized agents, thus preventing unauthorized access.
- The framework includes ThreatSieve for secure communication, verifying information authority to prevent impersonation, and HierarCache for adaptive memory management against unauthorized access.
- Experiments show AgentSafe achieves over 80% defense success rates under adversarial conditions, demonstrating its effectiveness and scalability in securing multi-agent systems as complexity increases.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
MathMistake Checker: Revolutionizing Math Error Detection with LLMs
Description
The MathMistake Checker is an innovative system that automates the identification of mistakes in step-by-step mathematical problem-solving, leveraging advanced LLM capabilities to enhance grading and learning experiences.
Key Points
- The system employs a two-stage process integrating computer vision and LLMs to efficiently identify errors in lengthy mathematical answers, improving grading accuracy.
- It supports open-ended grading without needing reference answers, allowing for personalized feedback tailored to individual learning needs.
- Demonstrated effectiveness across various math problem types, including calculations and word problems, showcasing its versatility in educational settings.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
VirtualXAI Framework Enhances Explainability in AI with User-Centric Assessments
Description
The paper presents VirtualXAI, a framework that enhances explainability in AI by integrating quantitative and qualitative assessments through GPT-generated personas, addressing challenges in evaluating XAI methods.
Key Points
- VirtualXAI combines quantitative metrics like fidelity with qualitative user satisfaction, improving the evaluation of explainable AI methods.
- The framework utilizes virtual personas based on LLM backstories to provide context for user assessments, enhancing interpretability and trust in AI models.
- A content-based recommender system matches new data with benchmarked datasets, offering tailored recommendations for optimal AI models and XAI methods in specific scenarios.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
KidneyTalk-open: A No-Code LLM System for Secure Kidney Disease Management
Description
KidneyTalk-open is a no-code desktop system designed for secure deployment of large language models (LLMs) in kidney disease management, integrating medical documentation for enhanced clinical decision support.
Key Points
- The system addresses data security risks of cloud-based LLMs and eliminates the need for technical expertise in local model deployment, making advanced AI accessible to clinicians.
- KidneyTalk-open features a medical document processing pipeline and an Adaptive Retrieval and Augmentation Pipeline (AddRep) that improves the accuracy of medical queries by integrating intelligent knowledge.
- Experimental results show a 29.1% accuracy in nephrology exam questions, outperforming mainstream products and establishing a new framework for privacy-sensitive clinical AI applications.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
New Scaling Law Enhances Long-Context Language Modeling in Large Language Models
Description
The paper presents a bipartite mutual information scaling law that governs long-range dependencies in language modeling, establishing a theoretical foundation for enhancing long-context language models.
Key Points
- The proposed scaling law is distinct from conventional mutual information, crucial for understanding long-context language modeling.
- The Long-context Language Modeling (L^2M) condition links a model's capacity for long context length to its latent state size for past information storage.
- Experimental validation on transformers and state space models supports the theoretical framework, guiding future developments in large language models for longer contexts.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
Advocating for Long-Output Generation in Large Language Models Research
Description
This paper emphasizes the need to shift focus in long-context LLM research from input processing to the generation of long-form outputs, addressing a significant gap in current capabilities.
Key Points
- Recent advancements in long-context LLMs have prioritized input comprehension, neglecting the generation of coherent long-form outputs essential for tasks like novel writing and complex reasoning.
- The authors advocate for a paradigm shift in NLP research to develop foundational LLMs that can produce high-quality, contextually rich, and logically consistent extended text.
- Addressing the challenges of long-output generation is crucial for real-world applications, highlighting the importance of this under-explored domain in LLM research.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
LLMs Exhibit Bayesian Reasoning Through In-Context Learning with Coin Flips
Description
This research paper examines how large language models (LLMs) utilize in-context learning (ICL) to perform structured reasoning consistent with Bayesian principles, particularly through biased coin flip scenarios.
Key Points
- LLMs often start with biased priors, leading to divergence in zero-shot settings, indicating initial misalignment in reasoning.
- In-context evidence significantly influences LLMs, often overriding explicit bias instructions, showcasing their adaptability in reasoning tasks.
- The study finds that LLMs generally adhere to Bayesian posterior updates, with deviations stemming from miscalibrated priors rather than flawed inference processes.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
Universal Scaling Laws for Hyperparameter Optimization in Large Language Models
Description
This research paper presents universal scaling laws for hyperparameter optimization in Large Language Models (LLMs), revealing optimal relationships for learning rates and batch sizes across various configurations.
Key Points
- The study identifies a power-law relationship for optimal learning rates with model parameters and data sizes, while batch sizes primarily scale with data sizes.
- A convex optimization landscape for hyperparameters is established, indicating an optimal plateau for hyperparameter settings under fixed conditions.
- The research utilized nearly one million GPU hours to train 3,700 LLMs, demonstrating the robustness of the scaling laws across different model structures and data distributions.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Mar 7
Watermarking Techniques Essential for Detecting Misuse of Open-Source LLMs
Description
This research paper addresses the need for watermarking techniques to detect misuse of open-source large language models (LLMs), focusing on intellectual property violations and usage violations.
Key Points
- The study identifies two misuse scenarios for open-source LLMs: intellectual property violation and LLM usage violation, highlighting the need for effective detection methods.
- It evaluates two watermarking techniques: backdoor watermarking, which effectively detects IP violations, and inference-time watermark distillation, which is less robust but applicable in both misuse scenarios.
- The research emphasizes the importance of developing advanced watermarking methods for open-source LLMs to enhance misuse detection and maintain model performance.
linkCopy linkshare_windowsShare
