Continual learning framework using hierarchical exploration-exploitation to acquire knowledge from task streams without catastrophic forgetting.
Combines MCMC correction with score-based diffusion models using Metropolis-Hastings steps for improved sampling in model composition.
Novel k-means clustering approach incorporating causal inference to identify heterogeneous treatment effects across unknown subgroups.
Method for estimating intrinsic dimensionality of datasets accounting for scale-dependent effects and measurement noise in unsupervised learning.
LLM-based approach for unsupervised code correctness evaluation that separates code comprehension from auditing to improve accuracy without reference implementations.
Method for learning stochastic differential equations from temporal snapshots without observable trajectories, applied to gene networks and financial markets.
Project management framework using GenAI agents to optimize team composition by matching personality roles.
Trust-region stochastic SQP algorithm for nonlinear optimization with complexity bounds under heavy-tailed noise.
ScienceT2I dataset and evaluation of image generation models for physical plausibility across 16 scientific domains.
Theoretical framework using stochastic optimal control to adaptively determine guidance weights in diffusion models.
Framework for robots to learn hidden state representations online in unstructured environments via situational awareness.
Addresses negative transfer in fine-tuning by selectively forgetting unhelpful pre-trained knowledge in language models.
Variance-based pruning method for compressing trained networks including Vision Transformers with minimal retraining.
NES framework for low-latency code edit suggestions without explicit instructions, using learned editing trajectories.
Federated learning framework for person re-identification addressing non-IID data and communication efficiency.
Network embeddings on Dutch population-scale data reveal educational divides correlated with right-wing populist voting.
Open source CayleyPy library for efficient Cayley and Schreier graph computations, with 200+ new conjectures in group theory.
Retrieval-of-Thought (RoT) system reuses prior reasoning steps organized in thought graphs to improve LLM inference efficiency.
Evaluates self-replication risks in LLM agents through realistic testing of autonomous agent behaviors and safety concerns.
Watermarking method (MOLM) for detecting and attributing synthetically generated images using LoRA markers.
Framework integrating data-driven learning with knowledge bases for genetic perturbation prediction in cellular systems.
TempoControl method for fine-grained temporal control in text-to-video generation models.
Learning-theoretic framework quantifying trade-offs between synthetic and real data using algorithmic stability bounds.
Proposes flow matching method for Bayesian posterior inference without likelihood evaluation, using block-triangular velocity fields.
RAG system for exhaustive multi-document question answering that checks all relevant documents without clear stopping conditions.
Multi-Stream VAE architecture combining discrete and continuous latent spaces for disentangled representation learning.
Multi-agent reasoning framework using AI agents for interpreting gene clusters in antimicrobial resistance transcriptomic data.
Framework using conformal prediction to assess correctness of LLM outputs and construct confidence sets for generative model responses.
Data-free quantization techniques for CLIP vision-language models enabling model compression without real data access for privacy-sensitive scenarios.
Study showing structured prompts significantly improve language model evaluation accuracy compared to single static prompt configurations in benchmarking.
LLM-based framework bridging cross-domain data sources for stablecoin transparency in circulation, reserves, and disclosure records.
Co-design framework for learning task-specific robot hand morphology and dexterous control policies simultaneously.
RoboNeuron middleware layer connecting Vision-Language-Action models and LLM agents to robot middleware, standardizing tool API integration for embodied AI.
SPDMark framework for robust in-generation video watermarking balancing imperceptibility, robustness, and computational efficiency.
EvalBlocks modular framework for efficient evaluation of foundation models in medical imaging, reducing manual experiment tracking workflows.
Graph learning via integer programming for causal discovery and inferring dependence structures in complex systems without restrictive assumptions.
Survey of meta-learning and meta-reinforcement learning methods enabling rapid task adaptation with minimal data, tracing DeepMind's adaptive agent research.
OPERA data pruning framework for efficient dense retriever adaptation, balancing quality-coverage tradeoff in domain-specific finetuning.
Self-attention CycleGAN for multi-site MRI harmonization addressing scanner-induced covariate shifts in neuroimaging data.
AI agents autonomously perform high energy physics analysis pipeline stages including event selection, background estimation, and statistical inference using LLMs.
LLM Router uses internal prefill activations for query-specific model selection, outperforming semantic routing by capturing model-specific failures.
Closed-form conditional diffusion models for data assimilation using analytical score functions instead of neural network approximations.
Theoretical proof that No-Clash Teaching Dimension is bounded by VC Dimension in machine learning theory.
CarbonEdge framework for carbon-aware deep learning inference on edge devices, optimizing for environmental impact alongside latency and throughput.
Theoretical analysis of man-in-the-middle attack detection in model-free reinforcement learning systems using Bellman Deviation Detection framework.
Physics-Informed Neural Networks framework (Deflation-PINNs) for solving nonlinear PDEs by identifying multiple distinct solutions using deflation loss.
Agent2 open-source runtime for production AI agents with schema-to-API capabilities, auth, and provider routing.
Database optimizers become critical infrastructure when AI agents autonomously generate SQL queries.
3D semantic atlas of 188 constitutions using embeddings and UMAP for conceptual law search.
YC startup Delve faces whistleblower allegations of misusing open source code without attribution.