The Archive

ProtoFlow uses prototype-guided flow matching with vector quantization for efficient non-iterative multivariate time series forecasting.

Study reconciles contradictory conclusions on model collapse in diffusion models, showing fixed-budget training prevents collapse with mixed synthetic-real data.

PPNAS combines prediction-powered inference to fuse sparse performance labels with noisy zero-cost proxies for efficient neural architecture search.

Stylometric analysis of 137k French Compar:IA votes reveals formatting and lexical diversity influence LLM rankings beyond content quality.

EP-Flow predicts disordered crystal structures via Occupancy Distribution Matrices, enabling generation without site-level annotations.

DAYJOB benchmark of 130 professional tasks (healthcare, finance) with containerized environments and expert rubrics for evaluating agentic long-horizon work.

Federated LLM fine-tuning over mobile RANs requires coordinated transport for asynchronous model updates with shared deadlines.

Continual learning framework for 6-DoF grasp synthesis using memory-based adaptation in embedding space rather than model finetuning.

Label-free energy disaggregation system for smart meter data using event detection and questionnaire-guided estimation for appliance classification.

ITC-MoE compression method for MoE Diffusion Language Models using importance-guided token-aware factorization with dynamic rank allocation.

Tutorial on model validation methods from hold-out splits to nested group cross-validation with examples from biomedical and applied research.

Review of trustworthy DataOps and MLOps practices for Intelligent Transportation Systems and Logistics applications.

PickMoment method for continuous-time single-image-to-video generation by learning motion blur deblurring and blur-to-video synthesis.

SCOPE-AD framework for cost-aware Alzheimer's disease diagnosis using ordinal-belief planning and energy-based models to guide sequential test acquisition.

Analysis of token-retention failure in uniform-state diffusion models and method to improve self-correction by preventing unwarranted revisions.

REPAIR method recovers evidence from frozen encoder representations to improve personalization encoder preference states for ranking and generation tasks.

IQS-BO framework for Bayesian Optimization using in-context query selection with Prior-data Fitted Networks to amortize surrogate and acquisition function costs.

Study pilots GenAI for scalable real-time feedback in large math classes, balancing automation with human oversight to reduce grading delays.

PROMO enables quadrupedal robots to balance multiple locomotion objectives (tracking, stability, energy) as runtime preferences via multi-objective RL.

SciUtopia: closed-loop LLM-agent simulation framework models co-evolving academic research ecosystems including collaboration, peer review, and funding dynamics.

DeFA traces failure propagation in LLM agents by constructing dependency graphs to localize multi-step errors separated by execution steps.

APGEM adapts error mitigation strategies for quantum RL on NISQ devices by dynamically responding to noise and training state changes.

Dynamic task routing benchmark tests agent ability to handle revised task bindings and temporal consistency across MMLU, MATH, HumanEval variants.

OffQuery benchmark reveals multi-agent collaboration can produce correct answers while corrupting shared state, tested on healthcare and disaster response.

Audit of VLM image judges on culturally situated prompts finds 4B-parameter models barely outperform random selection and CLIP baselines.

Robustness evaluation of Japanese LLMs against IME-related and typographical errors across character transposition, replacement, and homophone perturbations.

Mixture-Trained Merging combines multi-objective expert models (math, code, reasoning) via weight-space fusion to avoid sequential training order sensitivity.

Framework for efficient SLM-LLM collaboration via learned information acquisition rather than fixed delegation, reducing context transmission overhead.

Training approach enables language agents to internalize control decisions (revision, stopping) from harness-supported trajectories, reducing runtime dependency.

ASCRIBE framework for Thai clinical SOAP note generation using significance-weighted atomic facts; introduces ThaiClinicBench benchmark dataset.