Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback
Multi-agent reinforcement learning framework addressing robustness to data corruption in preference-based learning from human feedback.
Multi-agent reinforcement learning framework addressing robustness to data corruption in preference-based learning from human feedback.
Analysis of AI scaling requiring repeated efficiency doublings, distinguishing logical compute from physical resource implementation efficiency.
Biomimetic physics-informed neural networks for modeling microstructure-forming phase transitions in cellular matrices.
Physics-informed label-free pretraining method for coupled multiphysics simulation surrogates using operator-split latent prediction.
Bayesian information-theoretic approach to training data attribution that traces model predictions to influential training examples for interpretability.
STQuant framework for adaptive spatio-temporal quantization of optimizer states during large multimodal model training to reduce memory costs.
Quantum computing approach for option pricing using tensor networks to prepare quantum states encoding asset price distributions.
Adaptive neural networks for autonomous micro-drones with computational constraints via dynamic slimmable network architecture.
Theoretical analysis of offset noise in diffusion models to address brightness value generation challenges in large-scale models.
DMin framework for scalable training data influence estimation in diffusion models, enabling identification of influential training samples on generated outputs.
VAE-diffusion framework for generating high-quality SVG graphics from text with structural understanding.
Comprehensive privacy attack analysis on image autoregressive models, identifying membership inference and extraction vulnerabilities.
Method for enforcing syntactic and semantic constraints in LLM decoding through MCTS-guided token-level control.
Large-scale corpus of 324,843 Python classes from open-source projects for training and evaluating LLMs on code generation.
RAG-based LLM workflow using domain-specific knowledge graph for automated single-cell type annotation in biology.
Study evaluating sparse autoencoders for detecting bugs in Java code, addressing software vulnerability detection.
Technique for improving BEV-based 3D object detection in autonomous driving by reparametrizing regression targets.
Task and dataset for detecting when users are reading in egocentric smart glasses video using multimodal models.
Benchmark dataset (SealQA) for evaluating search-augmented LLMs on fact-seeking questions with conflicting or noisy search results.
Deployment case study of LLM-based platform for automated assessment of Romanian Bacalaureat exam questions using Gemini Flash.
Framework for 3D reconstruction of articulated objects using part-aware Gaussian splatting representation.
Survey of AI applications in marine robotics for ecosystem monitoring and conservation using underwater perception.
Analysis of clinician bias in emergency psychiatry using NLP to detect negative language linked to diagnostic disparities.
Adversarial training framework for neural networks that mitigates inter-class feature overlap to improve robustness.
Inference method for VideoLLMs that processes multiple frame subsets in parallel to improve temporal detail without increasing context window.
Technique to improve CLIP few-shot classification by addressing modality gap through semantic bridging between image and text embeddings.
Benchmark for evaluating LLMs on detecting demographic-targeted social biases across diverse content types and demographics.
Method to improve LLM performance in multi-turn conversations by reinforcing long-term planning and goal tracking through prompting.
Protein structure alignment using optimal transport for identifying and comparing local structural motifs.
Lightweight Disentangled Concept Bottleneck Model addressing bias in input-to-concept mapping for interpretable multimedia recognition.
Framework enabling diffusion models to adapt generation quality based on real-time network bandwidth constraints in cloud-to-device scenarios.
Minimax optimal algorithm for best arm identification under fixed sampling budget with applications to A/B testing.
Convex optimization framework for robust scheduling of aggregated EV battery storage under uncertainty.
Study of task transfer in Vision-Language Models examining how finetuning on one perception task affects performance on others.
Philosophical analysis arguing static value alignment approaches cannot ensure robust AI alignment under capability scaling and distribution shift.
PINNs applied to source inversion in advection-diffusion equations with sparse measurements for scientific computing.
OxEnsemble: Fair classification approach for low-data, imbalanced settings with demographic group constraints.
Method for determining singular value thresholds in DNN weight compression using random matrix theory.
Parameter-efficient transfer learning with neural operators for microseismic phase picking across varying signal conditions.
Study on selecting minimal training data subsets for example-based explanations of language model predictions using influence estimation.
Wireless ML inference via programmable metasurfaces for over-the-air extreme learning machines in MIMO systems.
Accordion-Thinking: Framework enabling LLMs to self-regulate reasoning step granularity through dynamic summarization for efficient inference.
Neuro-symbolic framework using differentiable logic programming to design and optimize quantum circuits.
Complexity analysis of accelerated proximal-gradient methods for ℓ1-regularized PageRank computation.
Theoretical analysis of flow matching generative models' adaptation to data manifold structures.
ZipMap: Stateful 3D reconstruction model achieving linear-time complexity for large image collections via test-time training.
Evaluation of 17 LLMs showing diagnostic reasoning degrades across multi-turn conversations compared to single-turn benchmarks.
HiCI: Hierarchical attention module for long-context language modeling, organizing information from local to global levels.
tBayes-MICE: Bayesian approach to multiple imputation for time-series data with missing values via MCMC sampling.
CodecSight optimizes streaming vision-language model inference by leveraging video codec signals for end-to-end efficiency.