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·Apr 2
Reinforcement Learning Enhances Multimodal LLMs for Clarifying Human Instructions
Description
This research introduces a novel approach for embodied agents to interpret ambiguous human instructions by fine-tuning multimodal large language models (MLLMs) as vision-language-action policies using reinforcement learning, significantly improving task execution.
Key Points
- The study presents the Ask-to-Act task, where agents must clarify ambiguous instructions to fetch specific objects in a home environment, enhancing user intent understanding.
- A unique method is proposed that utilizes LLM-generated rewards for online reinforcement learning, eliminating the need for extensive human demonstrations in training.
- Results show that the RL-finetuned MLLM outperforms strong zero-shot baselines and supervised models, achieving a 19.1-40.3% improvement in performance across various tasks and scenes.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
LLMs Struggle to Diagnose Cognitive Skills in Math Problem-Solving
Description
This research investigates the ability of large language models (LLMs) to diagnose students' cognitive skills in math problem-solving, revealing significant challenges and limitations in their performance.
Key Points
- The study introduces MathCog, a benchmark dataset with 639 student responses to 110 math problems, annotated for cognitive skill assessment.
- Evaluation of 16 LLMs shows all models scored below 0.5 in F1 scores, indicating poor performance in diagnosing cognitive skills despite advancements in AI.
- Findings highlight a correlation between model size and performance, with larger models showing better diagnostic capabilities, but also exhibiting overconfidence in incorrect assessments.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Innovative Multi-Model Strategy Enhances LLM Performance and Reduces Inference Costs
Description
The paper introduces a cost-efficient strategy for enhancing Large Language Model (LLM) performance by utilizing a repeated-sampling-then-voting framework with multiple models, optimizing test-time compute.
Key Points
- The proposed strategy leverages the strengths of multiple models, including weaker ones, to improve performance through diverse training data and paradigms.
- Theoretical analysis and extensive experiments show that this approach outperforms existing methods while significantly reducing inference costs across six datasets.
- ModelSwitch requires only a few comparable LLMs for optimal performance and can be extended with verification methods, showcasing its versatility in the generation-verification paradigm.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Integrating Personality Traits in LLM Agents Enhances Decision-Making for Cyber Defense
Description
This research explores the integration of personality traits into LLM-based autonomous agents, demonstrating how these traits influence decision-making and task selection processes in cyber defense applications.
Key Points
- The study builds on previous work with SANDMAN, a Deceptive Agent architecture that utilizes the OCEAN personality model to enhance task planning.
- A novel method for evaluating the impact of induced personality traits on planning, scheduling, and decision-making in LLM agents is presented, revealing distinct task-selection patterns.
- The findings suggest that personality-driven design can lead to more effective and plausible Deceptive Agents, improving proactive strategies in cyber defense scenarios.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
LLM Innovations Enhance Search for Deletion-Correcting Codes
Description
The paper presents a novel approach using Large Language Models (LLMs) to enhance the search for deletion-correcting codes, integrating AI techniques with information theory principles.
Key Points
- The research explores how LLMs can be utilized to improve the efficiency of searching for deletion-correcting codes, which are crucial in error correction.
- It combines concepts from artificial intelligence and information theory, showcasing the interdisciplinary nature of the study.
- The authors, Franziska Weindel and Reinhard Heckel, contribute to the field by proposing innovative methodologies that leverage LLM capabilities for complex coding problems.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Innovative Framework Enhances Local Model Numerical Reasoning with Data Protection
Description
This paper presents a collaborative framework for numerical reasoning that enhances local model performance while ensuring data protection, addressing the challenges of computation-constrained devices.
Key Points
- The proposed framework includes a context-aware synthesis strategy that maintains logical consistency while shifting query domains, improving reasoning accuracy.
- A tool-based answer reconstruction approach allows local models to leverage remote problem-solving patterns without exposing sensitive data, reducing data leakage significantly.
- Experimental results show an accuracy improvement of 16.2% to 43.6% and a data leakage reduction of 2.3% to 44.6% compared to existing methods, demonstrating the framework's effectiveness.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Study Reveals Limitations in Large Language Models' Numerical Reasoning Abilities
Description
This research paper evaluates the numerical reasoning abilities of Large Language Models (LLMs) through a test called "Numberland," revealing significant limitations in their basic mathematical skills despite their advanced capabilities.
Key Points
- The study introduces "Numberland," a 100-problem test assessing LLMs on basic operations, advanced calculations, and problem-solving integration, highlighting their numerical reasoning skills.
- Five LLMs, including OpenAI's o1 and Google Gemini, scored well on deterministic tasks but struggled significantly in trial-and-error scenarios, particularly in the 24 game.
- The findings indicate that LLMs exhibit fragile numerical reasoning, with performance dropping sharply on more complex problems, emphasizing the need for targeted evaluations of their mathematical capabilities.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Exploring LLMs: A Pathway to Enhanced Explainable AI
Description
This survey explores the role of Large Language Models (LLMs) in enhancing Explainable AI (XAI) by transforming complex outputs into understandable narratives, addressing transparency issues in AI models.
Key Points
- LLMs can bridge the gap between complex AI model behavior and human interpretability, improving user trust in AI decisions.
- The survey reviews existing approaches to LLMs for XAI, highlighting challenges, limitations, and evaluation techniques for generated explanations.
- Future directions emphasize the need for more interpretable, automated, and user-centric approaches to XAI using LLMs, fostering multidisciplinary collaboration.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
SC Outperforms GenRM in Compute Efficiency for LLM Reasoning Tasks
Description
This paper evaluates the efficiency of Self-Consistency (SC) versus Generative Reward Models (GenRM) in enhancing reasoning capabilities of large language models (LLMs) under fixed inference budgets, revealing SC's superior compute efficiency.
Key Points
- The study compares SC, which generates multiple solutions and selects the most common, with GenRM, which reframes verification as a next-token prediction task.
- Findings indicate that SC is more compute-efficient than GenRM for most practical inference budgets, requiring significantly less compute to outperform GenRM.
- The research derives inference scaling laws for GenRM, suggesting that optimal inference favors aggressive scaling of solution generation over verification, providing practical guidance for optimizing test-time scaling.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Token Embeddings in LLMs: Challenging the Manifold Hypothesis
Description
This paper investigates the structure of token embeddings in large language models (LLMs), revealing that they often violate the manifold hypothesis, which can lead to flawed interpretations of LLM behavior.
Key Points
- The study introduces a generalized model based on fiber bundles to analyze token embeddings, highlighting the distinction between signal and noise dimensions in the input space.
- Empirical tests on various open-source LLMs show frequent rejection of the fiber bundle null hypothesis, indicating that token subspaces do not conform to manifold structures.
- The findings suggest that prompts containing statistically significant tokens lead to greater output variability, emphasizing the importance of local signal dimensions in LLM responses.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
Zero-shot Benchmarking Framework Enhances Evaluation of Language Models Across Tasks
Description
The paper introduces Zero-shot Benchmarking (ZSB), a framework for creating scalable and flexible benchmarks for evaluating language models across various tasks and languages, enhancing automatic evaluation methods.
Key Points
- ZSB automates the creation of test data and evaluation prompts, making it adaptable for any task without relying on pre-existing datasets.
- The framework has been tested on five text-only tasks and one multi-modal task, showing strong correlation with human rankings and outperforming standard benchmarks.
- Key factors influencing benchmark performance include the size of the judge model and the variety of datasets used, emphasizing the importance of these elements in evaluation accuracy.
linkCopy linkshare_windowsShare - arXiv - Artificial Intelligence·Apr 2
MedReason: Enhancing Medical Reasoning in LLMs with Knowledge Graphs
Description
MedReason introduces a large-scale dataset aimed at enhancing medical reasoning in LLMs through structured knowledge graphs, providing detailed, verifiable reasoning paths for medical problem-solving.
Key Points
- The dataset consists of 32,682 question-answer pairs with step-by-step explanations, improving the reliability of AI in medical tasks.
- Fine-tuning LLMs with MedReason leads to significant performance improvements, with the MedReason-8B model outperforming existing state-of-the-art models on clinical benchmarks.
- The dataset's quality is validated by medical professionals, ensuring that it meets the standards of accuracy and coherence in medical reasoning.
linkCopy linkshare_windowsShare
