LLM-based peer review system that verifies claims by checking related work and executing code, improving review quality beyond manuscript-only analysis.
Theoretical framework for evaluating cyclic non-transitive interactions between LLM-based agents using equilibrium concepts instead of linear rankings.
Framework proposing that ambient AI systems transition from modeling to constituting users' cognitive functions through sustained causal coupling.
Molecular discovery framework combining LLMs with diffusion models to improve generation of chemically valid molecules by relaxing autoregressive constraints.
Memory system for deep research agents that improves trajectory retrieval and memory evolution to enhance LLM reasoning and autonomous learning.
Unsupervised fine-tuning method to improve adversarial robustness and semantic quality of vision-language models through siamese contrastive learning.
LLM-based code translation agent using execution alignment to improve cross-language code generation without parallel training data.
Multimodal LLM fine-tuned for image forgery detection and localization with interpretable visual reasoning capabilities.
Divide-and-conquer proof synthesis approach using LLMs to automate formal verification in proof assistants like Coq, improving software quality verification.
Systematic analysis of challenges in transitioning foundation model systems from demos to production, covering reliability, cost, scalability, and compliance issues.
Edge-cloud collaborative VQA system using aligned vector quantization to split vision-language model computation between edge and cloud devices, reducing bandwidth and utilizing edge resources.
Retrieval-augmented generation applied to time-series foundation models for zero-shot forecasting across domains.
VarDrop reduces computational cost in multivariate time series forecasting by eliminating variate token redundancy.
ENTER system uses event graphs for interpretable Video QA with code generation and contextual reasoning.
Entropy-based framework with Transformer for next activity prediction in business process monitoring.
LongSpec enables efficient speculative decoding for long-context LLM inference with lossless acceleration for agent applications.
Framework measuring hedging and non-affirmation behaviors in LLM responses on human rights topics across identity groups.
NativQA framework extends to multimodality for culturally-grounded LLM/VLM evaluation across languages and regions.
LLM-aided tool automates Universal Verification Methodology testbench generation for RTL IC verification.
CMP-RT diagnostic probe reveals tokenization vulnerabilities in safety-aligned LLMs through phonetic perturbations.
Polar decomposition and matrix sign methods optimized for GPU-friendly deep learning training via Muon optimizer.
Multimodal diffusion models synthesize quantum circuits for efficient compilation with reduced hardware calls and runtimes.
HeartcareGPT suite with 400K ECG dataset enables multimodal medical LLMs for dual signal-image ECG understanding.
BulletGen reconstructs 4D dynamic scenes from monocular video using generative models to complete unseen regions.
Survey of continual reinforcement learning covering sequential decision-making, generalization, and adaptation across dynamic tasks.
LaSM defends GUI agents against pop-up injection attacks using layer-wise scaling on multimodal LLMs for safer screen interaction.
Framework for detecting LLM hallucinations in black-box generators by leveraging future context patterns.
Flow matching approach for quantifying aleatoric uncertainty in medical image segmentation, modeling expert annotation variability.
ShadowNPU enables efficient on-device LLM inference by redesigning attention operator for NPU execution, improving privacy and performance.
MedShift addresses domain gap between synthetic and real X-ray images using conditional transport for improved generalization to clinical settings.
LifeAlign framework for lifelong LLM alignment across sequential tasks using memory-augmented preference optimization without catastrophic forgetting.
Training-free prompt engineering strategy using state reconstruction and history reminders for efficient multi-turn LLM dialogue.
Chiplet-based RISC-V SoC architecture with modular AI acceleration for edge AI devices with improved yield and efficiency.
StateX post-training method improves recall ability in RNNs and state-space models for long-context information retrieval.
Component-level energy assessment framework analyzing transformer efficiency to enable green AI development.
Explainable bias-aware generative framework combining multimodal attention, attribution methods, and iterative feedback for fair generation.
Template Infilling conditioning strategy enables diffusion language models to handle flexible structural prompting beyond prefix-based generation.
Knowledge Reasoning Language Model unifies language models with knowledge graphs for inductive reasoning over unknown entities and relations.
RLAIF-SPA uses structured AI feedback to improve emotional expressiveness and semantic-prosodic alignment in text-to-speech synthesis.
Methods for evaluating and mitigating fairness issues in LLMs at inference time to reduce harmful behaviors and drift.
Eigen-Value method for efficient data valuation using eigenvalue-based approach, focusing on out-of-distribution robustness.
Data-efficient approach for adapting humanoid robot whole-body motion control from single motion examples using walking priors.
Routing-based architecture for multimodal LLMs enabling continual learning across sequential tasks while preventing catastrophic forgetting.
Snowflake's Cortex AISQL production engine integrates semantic operations into SQL for querying structured and unstructured data.
LLM-based automated feedback system for physics problem solving using evidence-centered design methodology.
CB-APM applies deep learning with interpretability-by-design to stock market prediction using analyst consensus data.
MedMistake pipeline automatically extracts and replicates LLM errors in medical conversations to create evaluation benchmarks.
PhyAVBench benchmark evaluates physics-plausibility of audio in text-to-audio-video generation models.
Framework using sparse autoencoders to identify and steer high-order semantic features in LLMs for reliable control of language generation behaviors.
IBISAgent improves pixel-level visual reasoning in medical multimodal LLMs for biomedical object segmentation through enhanced training strategies.