Open-source Python simulation package for modeling antibiotic prescribing and antimicrobial resistance dynamics in RL-compatible environment.
Zero-shot learning approach for automatically detecting semantic column types in relational tables without labeled training data.
Unsupervised neural combinatorial optimization method (UniHetCO) for multi-problem learning on graph subset-selection problems without ground-truth solutions.
Dual-path approach combining discrete mark prediction and continuous dynamics for marked temporal point processes using neural ODEs.
Proximal relaxation method for improving nonlinear probabilistic latent variable models in soft sensor applications, addressing training inefficiencies.
Deep learning network-temporal models for multivariate time series traffic prediction, addressing topological interdependency and complex temporal patterns.
Sorometry pipeline for automated phytolith analysis using AI to digitize and classify microscope images, replacing manual labor-intensive analysis.
Neuro-symbolic VLM agents for time series event detection using natural language descriptions, addressing semantic event classification with scarce labeled data.
Theoretical analysis proving attention sinks are necessary in softmax transformers for trigger-conditional tasks, formalizing why attention collapses to content-agnostic positions.
KEPo analyzes poisoning attacks on Graph-RAG systems where attackers inject malicious texts to manipulate LLM outputs.
Sharpness-aware minimization for stable item embedding learning in federated recommendation systems preserving privacy.
LongFlow compresses KV cache for reasoning models like o1 and R1, reducing memory and bandwidth during long output generation.
Conditional feature disentanglement approach for user-controllable privacy in wearable sensor-based human activity recognition.
Multi-Task Anti-Causal learning framework exploiting cross-task invariances to infer latent causes from observed urban event reports.
CAETC method using adversarial autoencoding for counterfactual estimation with time-dependent confounding in observational data.
Integrates survival analysis with classification for early chronic disease risk prediction using EMR data.
H-EARS combines potential-based reward shaping with energy-aware regularization for efficient deep reinforcement learning control.
AutoScout automates ML system configuration via structured optimization over model parallelism, communication, and runtime parameters.
Investigates partial RoPE rotations in transformers, reducing memory at long contexts while maintaining performance.
Personalized federated learning using Gaussian generative modeling to handle data heterogeneity across distributed clients.
Studies continual RL for Vision-Language-Action models, finding sequential fine-tuning avoids catastrophic forgetting without complex strategies.
Neuromodulated constrained autoencoders for context-dependent dimensionality reduction in varying environments.
Policy gradient methods for LLM reasoning naturally reduce trajectory diversity; proposes entropy-preserving training approach.
EvoFlows model for protein engineering using edit-based flow-matching to predict mutations on template sequences.
Examines calibration's role in reducing predictive multiplicity and improving stability in high-stakes ML classifier deployments.
Social bandit learning framework combining individual and collective learning in populations of reinforcement learning agents.
Theoretical study of Follow-the-Perturbed-Leader algorithm optimality in semi-bandit problems with best-of-both-worlds guarantees.
Analyzes model collapse when LLM-generated text re-enters training data as data consumption grows, proposing replay-based solutions.
Method for agents to autonomously discover symmetry groups for disentangled representation learning without prior structural knowledge.
Unifies membership inference attacks (LiRA, RMIA, BASE) as instances of exponential-family statistical framework for model privacy auditing.
Two-stage framework using Conditional Fourier Neural Operators to recover hidden ODE parameters from sparse observations.
Analyzes the role of reversible instance normalization in time series forecasting, addressing distribution shifts in temporal and spatial data.
Uses reinforcement learning to adapt LLM-based recommender systems for dynamic, need-specific objectives and complex recommendation goals.
EnTransformer generative architecture for multivariate probabilistic forecasting with reliable uncertainty quantification.
Chem4DLLM multimodal LLM interprets 4D molecular trajectories to explain chemical dynamics and reactions.
MobileKernelBench evaluates LLM capabilities for generating efficient compute kernels optimized for mobile devices.
Circuit mapping and mechanistic interpretability of Geneformer foundation model reveals redundancy and layer-dependent control.
Formalizes statistical and structural identifiability as distinct properties explaining representation learning model stability.
Flowcean framework automates data-driven model generation for cyber-physical systems with modular architecture.
Analyzes frequentist consistency of prior-data fitted networks for causal inference compared to classical estimators.
Training-free decoding acceleration for LLMs exploiting stable attention patterns within semantic spans during generation.
Addresses cross-domain reinforcement learning challenges when source and target domains have different state or action spaces.
Uses LLMs with feedback memory to automate neural architecture search for CNNs on consumer GPUs without fine-tuning.
Cornserve distributed serving system for any-to-any multimodal models with different input/output modalities and scaling characteristics.
Automated generation of high-performance RL environments using prompt templates, verification, and agent-assisted repair for <$10 compute cost.
IsoCompute scaling laws for optimal allocation of sampling compute across rollouts, problems, and update steps in LLM RL post-training.
Theoretical analysis of catastrophic forgetting in continual post-training of generative models under two-mode mixture abstraction.
Neural Thickets shows task-specific expert solutions exist in pretrained weight distributions, enabling discovery through structured optimization.
Perplexity's analysis of security considerations for frontier AI agents based on operating agentic systems at scale.
STAMP framework for text privatization using task-aware token-level privacy budget allocation balancing privacy sensitivity and task utility.