H-probes method to extract hierarchical structures from LLM latent representations. Analyzes geometric representations enabling hierarchical reasoning.
Open Earth System Foundation Model extending Aurora for weather/climate forecasting. Unified framework for heterogeneous geophysical data integration.
Evaluation framework for text-to-speech synthesis quality using crest factor, spectrum balance, and cepstral metrics across six TTS models.
BRITE: benchmark for text-to-video evaluation on implausible scenarios including audio-visual alignment and QA-based interpretable assessment.
Latent space probing framework for detecting adult content in video generative models by analyzing internal representations during generation.
OceanPile: multimodal ocean corpus addressing fragmented ocean data for training foundation models on climate and marine biodiversity tasks.
X2SAM integrates MLLMs with foundation segmentation models enabling pixel-level perception from conversational instructions across images and videos.
Retrieval-guided generation approach for medical image captioning reducing hallucinations and factual inconsistency in histopathology image descriptions.
DIAGRAMS: review framework for creating reasoning-level attribution in diagram QA, providing structured evidence annotation across visual diagrams and infographics.
TRIP-Evaluate: open multimodal benchmark for assessing LLMs and MLLMs on transportation tasks including regulation QA, traffic management, and autonomous driving reasoning.
Empirical study of multi-agent LLM debate showing isolated self-correction outperforms unguided homogeneous debate across teams of 10 models on high-difficulty benchmarks.
Survey of generative AI use in qualitative research, discussing suitability for different research approaches (small-q positivist vs Big Q non-positivist) in software engineering.
StyleShield: demonstrates fragility of AIGC detectors through continuous controllable style transfer, exposing reliability paradox as language models improve.
Code World Model preparedness report: Meta's code generation and reasoning model assessed for frontier AI risks; found no additional catastrophic risks beyond current ecosystem.
Interpretable experiential learning model based on state history and global feedback for reinforcement learning in resource-constrained environments, evaluated on Atari.
E-MIA: black-box membership inference attack against RAG systems that infers whether documents are in retrieval corpus by analyzing LLM response interactions.
Certainty-aware RAG system enabling LLMs to express appropriate confidence and say 'I don't know' to improve user trust.
Ablation study isolating contributions of LLM, vision perception, and control components in human-robot interaction for object detection.
Evaluation of retrieval-augmented generation chatbots in realistic multi-turn information-seeking workflows with compliance constraints.
Physics-aligned rotary positional encoding method for wireless foundation models in channel state information tasks.
Medical audio QA dataset with diverse clinical scenarios for benchmarking language and audio reasoning models.
Visual analytics workbench for exploring weather and climate data using embedding-based representations and similarity search.
Method using perplexity differencing to reveal finetuning objectives in LLMs, enabling detection of intentionally introduced behaviors.
Framework assessing how ambiguity and uncertainty in real-world medical queries degrade LLM reasoning compared to simplified benchmarks.
Benchmark and investigation of multimodal LLM behavior for emotion recognition under modality conflict and missingness conditions.
Certified purity architecture converting governance enforcement into structural boundaries for cognitive workflow execution systems.
Multi-agent reinforcement learning approach for tactical deconfliction and separation assurance between heterogeneous unmanned aerial systems.
Unlearning method to suppress hallucinations in LLMs, specifically addressing supply-chain attacks from fictional package recommendations.
Framework for computing optimal policies and safety filters using value functions with temporal logic specifications.
Lightweight policy orchestration framework using ML to select cache eviction policies for nonstationary object caches in cloud services.
Pretraining method combining layer-aligned distillation with convergence-based early exit for efficient transformer inference.
Defense mechanism against malicious instruction injection attacks in retrieval-augmented generation and tool-integrated LLM agents.
Legal analysis of EU AI Act governance gaps for autonomous agents in critical infrastructure systems like traffic and power management.
Knowledge tracing approach for AI tutoring systems using LLM-powered dialogue to assess student performance and provide personalized support.
Component-aware self-speculative decoding accelerates hybrid language models with architectural heterogeneity without external drafters.
DUET framework enables efficient inference by decomposing tasks: capable model produces reasoning signal, lightweight model interprets for prediction.
Forager testbed for continual reinforcement learning with partial observability, addressing plasticity loss in non-stationary environments.
Approach using multi-perspective transformers with test-time training and products of experts on ARC-AGI-2 visual reasoning benchmark.
Analysis of contradictory productivity claims in AI-assisted coding: controlled studies show gains but RCTs and telemetry show slowdowns and longer review times.
Method to minimize unintended side effects (collateral damage) when using activation steering to control LLM behavior through internal representation intervention.
CNN-based multi-input multi-output model for efficient spatiotemporal prediction addressing RNN parallelization and error accumulation.
Position paper arguing LLM inference serving systems need mathematical optimization and algorithmic foundations beyond current heuristics like FIFO and LRU.
Pre-trained deep learning model for automated plant leaf disease classification to enable early detection in agriculture.
Chain of Evidence method adds pixel-level visual attribution to iterative retrieval-augmented generation for multi-hop question answering.
Framework for autonomous learning systems handling concept drift and non-stationary data streams beyond traditional temporal shifts.
GraphSculptor method for constructing efficient pre-training coresets in graph self-supervised learning by exploiting dataset redundancy.
Sequential experimental design framework for vision-language models to overcome perceptual bandwidth bottleneck via active vision.
MAD-OPD uses multi-agent debate to overcome single-teacher capability ceiling in on-policy distillation for agentic tasks.
Self-contrast decoding strategy for diffusion large language models to improve generation quality via high-information-density context modeling.
Empirical study of how developers use LLMs in software design tasks, surveying practitioners on benefits and drawbacks.