Deep learning approach for clinical mortality prediction from incomplete multimodal Electronic Health Records using point cloud paradigm.
Method to mitigate reward hacking in Best-of-N sampling for language models using pessimism, addressing inference-time compute scaling challenges.
Novel Anticipatory Reinforcement Learning framework for non-Markovian decision processes with jump-diffusions and structural breaks, designed for single trajectory learning.
Batch Loss Score metric for dynamic data pruning using exponential moving averages, accelerating deep learning training.
Quantization method for LLMs combining mixed-precision and low-rank decomposition for efficient INT computation on NPU devices.
Sampling parallelism approach for efficient Bayesian neural networks and uncertainty quantification in risk-sensitive applications.
Mechanistic analysis decomposing GPT-2 Small's final MLP into legible exception handler with 27 named neurons routing decisions.
Method using task reformulation to enable LLMs to learn from difficult problems via reinforcement learning from verifiable rewards.
Complete pipeline for federated unlearning with evaluation framework, enabling models to forget deleted data in distributed learning.
Algorithms for automatic concept selection in interpretable reinforcement learning policies without manual domain expertise.
Research on how generator access constraints affect autoregressive post-training and learning from rollouts vs prefix queries.
Python toolkit for intersectional fairness analysis in clinical ML models, addressing compounded disparities beyond single-axis comparisons.
Empirical robustness analysis of TabPFN's attention mechanisms for in-context learning on tabular data, examining noise immunity without retraining.
DSPy framework for optimizing LLM prompt engineering through declarative learning instead of manual trial-and-error, improving scalability and reproducibility.
Formalizes data attribution methods for adaptive learning settings where training data is generated by models themselves, addressing feedback loop in online/RL systems.
Investigation into interpretability challenges of latent reasoning models that operate without explicit natural language reasoning, examining two approaches.
Hierarchical Instance-Conditioned Mixture-of-Experts architecture for object detection using sparse routing at instance level rather than image/patch level.
Hierarchical instance-conditioned mixture-of-experts architecture for object detection with sparse parameter activation.
Graph neural networks with contrastive learning for predicting power outages from extreme weather events.
Novel stratification-based semantics for Signal Temporal Logic with applications to reinforcement learning.
CNN-attention hybrid model for decoding hand kinematics from EEG in brain-computer interfaces.
Training method enabling Code LLMs to simulate program execution step-by-step, improving competitive programming performance.
Multi-agent research showing emergence of compositional communication protocols for representing latent physical properties without explicit supervision.
IPSL-AID: generative diffusion model for climate downscaling from global to regional resolutions.
SpikeVPR: neuromorphic approach using event-based cameras and spiking neural networks for energy-efficient visual place recognition.
Cross-Stage Attention Residuals mechanism for medical image segmentation using selective aggregation of encoder-decoder outputs.
Lossless compression method for LLMs enabling fast inference on Ascend NPUs, addressing weight data transfer bottleneck.
Generative molecular language models pre-trained on chemical data and fine-tuned for energetic materials discovery.
Proposes automated discovery approach for computer architecture design using AI, addressing post-Moore's Law era challenges.
TensorBoard plugin for interactive multi-metric visualization and fairness analysis during ML model training.
Multimodal violence detection model combining VideoMamba and AudioMamba with conditional LoRA steering.
Addresses iterative image quality degradation in multi-turn editing with agentic systems using multi-modal models.
Diffusion policy approach with Bayesian expert selection for active multi-target tracking balancing exploration and exploitation.
Zero-shot quantization technique using weight-space arithmetic to transfer quantization robustness between models without training data.
Inference optimization method for frozen vision transformers through circuit duplication for marine species classification.
Study comparing RAG-based approach with traditional methods for Agile story point estimation in sprint planning.
Lightweight query routing classifier for selecting optimal retrieval strategies in RAG pipelines based on query characteristics.
Wearable AI agent on smart glasses enabling continuous perception and speech-driven task execution with OpenClaw agentic framework.
Method using sparse autoencoders to discover language-specific features from monolingual data for controlling LLM output language without parallel data.
Discrete diffusion language model using tree-structured token prediction to reduce parameters and memory in language generation.
Method for LLMs to dynamically compress intermediate reasoning thoughts into compact representations while maintaining reasoning quality.
Deep reinforcement learning framework for optimizing land-use allocation in Lake Malawi Basin to maximize ecosystem service value.
Method using Rényi attention entropy for patch pruning in transformers to reduce quadratic self-attention cost.
Study on adversarial attacks against transformer-based malware detectors using control flow graphs, examining robustness of RoBERTa models.
SecureAFL: Asynchronous federated learning framework addressing straggler problem while maintaining security.
Study comparing LLM probed representations with performance on narrative analogical reasoning tasks.
Secure-by-design GenAI framework integrating PromptShield for LLM-based cloud security and forensic analysis.
Cross-Modal Graphical Lasso for learning interpretable multimodal representations by disentangling shared and specific topologies.
Value-based safety forecasting for streaming LLM outputs, improving response moderation on partial generations.
Fixed-confidence best arm identification in semiparametric bandits with instance-optimal sample complexity bounds.