A probabilistic framework for online test-time adaptation
Probabilistic framework for adapting trained models to unlabeled test data under distribution shift.
Probabilistic framework for adapting trained models to unlabeled test data under distribution shift.
Defense mechanisms for tool-using LLM agents against indirect prompt injection via out-of-band policies.
Study of performance gaps in deep learning fault diagnosis techniques across seen vs. unseen programs.
Mean-field reinforcement learning framework for continuous-time control with discrete-time training data.
RAG system addressing temporal validity problem where stale facts cause AI agents to serve outdated information.
Transformer architecture with patch-mean decoupling for long-term time series forecasting across multiple domains.
Explainable ensemble ML models for cirrhosis detection in hepatitis C patients using medical data.
Latent-space diffusion model for solving PDE-constrained inverse problems with Bayesian posterior sampling.
Event camera simulation framework using fractional-relaxation pixel dynamics for asynchronous vision tasks.
DroidBreaker: practical problem-space adversarial attacks on ML-based Android malware detectors with functional APK modifications.
Develops interpretable CNN models for stress/affect recognition from respiratory signals with feature analysis and autocorrelation lags.
Formalizes scientific discovery as meta-optimization where LLMs simultaneously modify evaluation criteria and optimize within expanded theory spaces.
ProtoKV: constant-footprint memory system for streaming video understanding with asynchronous delayed queries using prototype-based summarization.
Systematic analysis of evaluation pitfalls in multimedia event extraction across text and image modalities.
Proposes learnable feature caching calibration for Diffusion Transformers to reduce error accumulation and accelerate inference.
Presents MIRROR, a memory-guided MCTS framework for systematic red-teaming of multimodal agentic RAG systems across multiple attack surfaces.
Studies memory depth and parametric consolidation in long-running language agents, introducing loop-drift protocol to test durable behavior adaptation.
Develops scalable quantum graph neural networks using message-passing within Weisfeiler-Leman hierarchy for chemistry and biology data.
Analyzes frozen language model representations as neural predictors during naturalistic language comprehension using brain imaging data.
Introduces SamAdams adaptive timestepping and position-adaptive Langevin dynamics for accelerated sampling in stiff phase spaces.
Proposes synthetic data generation for fiber bundle segmentation in tracer histology using dMRI tractography validation.
Examines automated jailbreak selection using bandit algorithms for non-expert malicious actors to craft effective LLM attacks.
Proposes geometric gradient rectification for semi-supervised learning with out-of-distribution outliers in unlabeled data.
Develops XMSE-aware adaptive empirical Bayes estimator interpolating between ML and kernel-based EB to address second-order alignment issues.
Proposes self-supervised learned primal-dual method for X-ray CT reconstruction in low-dose settings without ground-truth data.
Describes RolloutPipe, a system for overlapping pipelined rollout and training in disaggregated RL architectures for LLM post-training with verifiable rewards.
Proposes semantic early-stopping for multi-agent LLM loops using embedding similarity to halt when output meaning stops improving, reducing token waste.
Introduces parametric open-source games, a continuous model where players choose parameters converted to actions, with equilibrium existence results.
Introduces Mass Index and regularized extended KL divergence for local-mass analysis in Bayesian inference beyond global divergence objectives.
Proposes DMuon, distributed training method for matrix-orthogonalization optimizers reducing communication overhead compared to element-wise optimization.
Describes prizewinning bimanual garment folding solution combining vision-language-action policy with reinforcement learning loop for robotic manipulation.
Uses sparse autoencoders to inspect LLM internal states for forecasting tasks, identifying time-specific knowledge versus generalizable patterns.
Introduces Hierarchical Muon (HiMuon), tiled Newton-Schulz optimization for efficient dense neural network training with reduced computational overhead.
Proposes CARVE, memory-aware recurrent architecture with content-aware gating for efficient chunk-parallel linear attention in sequence models.
Introduces Ribbon, scalable approximation to Dirichlet-reweighted bootstrap for efficient uncertainty quantification in high-dimensional models.
Analyzes fundamental ceiling on multi-model LLM systems (routing, voting, mixture-of-agents), showing accuracy limited by co-failure rate across 67 frontier models.
Proposes methods for implementing generative models on analog hardware with physics-determined dynamics for low-power computation.
Develops efficient algorithm for learning high-dimensional Gaussian distributions truncated to unknown halfspace with optimal sample complexity.
Introduces planning experience exploration for GUI agents using small open-source MLLMs, improving task planning and cross-website generalization.
Studies domain-aware distribution alignment in entity matching under data constraints, applying low-resource learning to data integration.
Investigates alignment between sequence probability and correctness in LLMs, quantifying when higher likelihood corresponds to correct outputs.
Theoretical analysis of frequency principle phenomenon showing DNNs learn target functions from low to high frequencies during training.
Proposes kernel distance method for ranking generative models in distributed settings based on output fidelity and diversity.
Develops gradient testing and estimation algorithms using only comparison oracle queries on smooth functions.
Studies relationship between over-parameterization in neural networks and adversarial robustness, analyzing vulnerability to adversarial examples.
Finite-sample analysis of decentralized best-response learning in two-player zero-sum games and stochastic games.
Byzantine-robust aggregation algorithms for secure decentralized federated learning without central servers.
ML approach to model how air traffic controllers build mental representations of complex air traffic situations.
Bayesian optimization method for identifying chemical reaction conditions that work across multiple substrates efficiently.
Chisme: gossip learning framework addressing heterogeneity in resource-constrained edge devices for privacy-preserving distributed learning.