Consistency-guided decoding approach for three-way logical QA in LLMs, addressing negation inconsistency and epistemic Unknown failures.
Textual time-series corpus of 136 GLP-1RA case reports with LLM-based timeline extraction and risk modeling for longitudinal analysis.
Framework combining LLM analysis with energy-system modeling to forecast electricity footprint of AI data centers through 2030.
Text2DistBench benchmark for evaluating LLM ability to infer distributional knowledge and population-level trends from text collections.
Theoretical framework for cross-lingual transfer and parameter-efficient adaptation in low-resource Turkic languages using LLMs.
Ethical design space exploration for sensor-fused LLM agents in health applications, addressing privacy and bias concerns.
SensorPersona system that extracts user personas from mobile sensor streams using LLM-based agents for improved personalization.
Tool-MCoT framework for content safety moderation using small language models augmented with external tools to reduce computational costs.
Analysis of latent cultural themes in training corpora by prompting six leading LLMs to identify recurring patterns about human culture.
Study comparing demonstration selection strategies for in-context learning in LLM-based next point-of-interest prediction.
Comparison of LLMs versus classical ontology methods for extracting breast cancer phenotypes from unstructured clinical notes.
DOVE benchmark for evaluating LLM cultural value alignment using open-ended generation rather than multiple-choice formats.
Study on improving faithfulness and traceability in retrieval-augmented generation through illocutionary explanation planning.
Research investigating implicit intersectional biases in LLMs under persona-driven contexts, introducing Bias Amplification framework to capture dynamic bias shifts.
Unsupervised neural network using BioClinicalBERT to automatically classify surgical urgency levels from medical transcriptions.
Research on detecting hallucinations and omissions in LLM-powered mental health chatbots, finding LLM judges achieve only 52% accuracy on safety-critical healthcare data.
ArXiv paper arguing the foundation model era (2020-2025) is ending as open-source models reach frontier performance and inference costs drop, making pre-training non-competitive.
ML and deep learning model comparison for triboelectric sensor-based sign language recognition.
Probabilistic language tries framework unifying compression, decision policies, and execution reuse for sequence models.
LLM-based code synthesis for database-native functions addressing hallucination and domain-specific challenges.
Framework for building verifiers for computer use agents with principles for constructing rubrics and training signals.
Unified defense mechanism against textual and visual jailbreaks and prompt injections in LLMs and VLMs.
Study on perception-integration gap in vision-language models: models extract visual info but fail in reasoning.
Fourier-based low-rank parameter-efficient fine-tuning for cross-lingual code generation with Code Llama 7B.
SE-Enhanced Vision Transformer and BiLSTM hybrid architecture for intrusion detection in IoT/medical environments.
Analysis of training dynamics during grokking via spectral edge directions, showing mechanistic interpretability tools miss structured functional modes of learning.
S³: stratified scaling search for test-time inference in diffusion language models using verifier-guided sampling to improve generation quality without retraining.
Dual-stream calibration method for improving contextual clinical reasoning in LLMs through dynamic representation adjustment via in-context learning and RAG.
IAMFM framework optimizing sponsorship configurations in LLM-generated ads using VCG incentives and multi-fidelity optimization under advertiser strategic behavior.
ToxReason benchmark for evaluating LLM mechanistic reasoning in chemical toxicity prediction via adverse outcome pathways beyond structural features.
Attribution-based explainability for LLM-based intrusion detection in SDN environments, addressing transparency and interpretability in security-critical applications.
MAT-Cell: neuro-symbolic multi-agent framework combining LLMs with biological priors for single-cell annotation, addressing generalization and spurious association problems.
LLM methodology for automated security profiling in Ukrainian cybersecurity compliance, integrating ISO/IEC 27001 and NIST frameworks.
Plasma GraphRAG: integrates Graph RAG with LLMs for automated physics-grounded parameter selection in gyrokinetic plasma simulations.
DosimeTron: agentic AI system automating personalized Monte Carlo radiation dosimetry in PET/CT with GPT as reasoning engine and 23 specialized tools.
ClawLess: security framework enforcing formally verified policies on autonomous LLM-based AI agents to mitigate code execution and data retrieval risks.
Hyperbolic geometry-based approach for detecting and sanitizing harmful prompts in Vision-Language Models.
TalkLoRA: communication-aware Mixture-of-Experts LoRA framework for parameter-efficient fine-tuning of LLMs with stable expert routing.
AgentOpt v0.1: client-side optimization framework for LLM-based agents composing local tools and remote APIs to reduce latency and costs.
Automatic framework for detecting speaker drift in diffusion-based text-to-speech models using binary classification.
Survey of 51 industry practitioners on preparing software engineers for generative AI tools, covering hiring and skill gaps.
Severity-aware weighted loss function for fine-tuning Arabic language models on medical complaint text generation.
Bi-level optimization framework (BiSDG) for single domain generalization without target domain access during training.
Analyzes in-context learning in speech language models, studying acoustic features, linguistic structure, and induction heads for text-to-speech tasks.
Severity-based curriculum learning strategy for Arabic medical text generation using LLMs.
WebSP-Eval benchmark for evaluating web agents on security and privacy tasks like cookie management and account settings configuration.
Empirical study of design issues in large-scale code generation by AI IDEs with agentic capabilities, beyond functional correctness.
Proposes Master Key Hypothesis and UNLOCK method for training-free capability transfer across different LLM scales via linear subspace alignment.
Research on foundation models for graph-structured data in biomedical applications, extending language/vision foundation model approaches to graph analysis.
arXiv paper on using LLMs for toxic habit extraction in Spanish clinical texts. Zero-shot and few-shot approaches for substance use classification.