FourierSpecNet hybrid framework combining Fourier spectral methods with neural networks to approximate collision operators for solving the Boltzmann equation efficiently.
State Space Neural Operator for learning solution operators of time-dependent PDEs using structured state space models with adaptive damping and learnable frequency modulation.
arXiv paper providing theoretical analysis of GRPO (Group Relative Policy Optimization) for LLM fine-tuning from human feedback.
arXiv paper analyzing EM algorithm behavior under model misspecification in mixture models with excess components.
arXiv paper analyzing GNN-based SAT solvers through graph Ricci curvature geometric perspective to explain performance degradation.
arXiv paper on efficient world models for heterogeneous multi-task planning, addressing gradient conflicts and plasticity loss.
arXiv paper on assessing performance of language model applications in healthcare, addressing evaluation methodology.
arXiv paper introducing Answer-Then-Check safety alignment method to defend LLMs against jailbreak attacks using reasoning.
arXiv paper on prompt-based federated continual learning addressing class-wise and temporal forgetting across distributed clients.
arXiv paper on training diffusion language models with planner-aware path learning to optimize generation strategies.
arXiv paper formulating diffusion model alignment as variational EM to reduce reward over-optimization and mode collapse.
arXiv paper on adapting decoder-only LLMs to partial differential equations via cross-modal learning for scientific machine learning.
arXiv paper introducing KLASS sampling method for masked diffusion models using token-level KL divergence to accelerate inference.
SQDF applies soft Q-function RL to fine-tune diffusion models with KL regularization, mitigating reward over-optimization.
Analyzes diversity loss in RL-trained LLMs caused by mode-seeking reverse KL; proposes forward KL filtering for reasoning tasks.
A-3PO accelerates asynchronous LLM RL training with staleness-aware proximal policy approximation, improving over decoupled PPO.
Analysis of optimization challenges in hyperbolic deep RL identifying gradient factors affecting training success for hierarchical state embeddings.
CARE failure-centric post-training framework uses contrastive learning on wrong rollouts to improve multimodal reasoning with verifiable rewards.
LLMTM benchmarks LLMs on temporal motif analysis in dynamic graphs for anomaly detection and structural understanding.
Spectral embedding approach for domain-invariant representations via optimal transport plans, addressing distributional shift.
EDIS analyzes token-level entropy trajectories during LLM generation to diagnose reasoning quality beyond aggregate confidence statistics.
Red-teaming framework to stress-test LLM alignment audits against strategic deception prompts in white-box and black-box settings.
LaPha trains AlphaZero-style LLM agents in hyperbolic space using Poincaré geometry for efficient tree search and dense reward shaping.
Aletheia agent iteratively generates, verifies and revises mathematical proofs using LLMs, advancing autonomous research capabilities.
FLoRG enables parameter-efficient federated fine-tuning of LLMs using low-rank adaptation with Procrustes alignment across distributed clients.
EMPO² framework combines on/off-policy RL with memory augmentation to improve exploration in LLM agents, addressing novel state discovery limitations.
Web-to-Knowledge-to-Web pipeline iteratively crawling domain-specific sources to discover SME suppliers in specialized industry sectors.
Framework establishing first-order equivalence between activation steering and weight updates for parameter-efficient LLM adaptation.
MatRIS: Foundation machine learning interatomic potentials with equivariant inductive bias for efficient material simulation.
Geometric pretraining approach for protein design combining structure learning and conformational ensembles with rigidity-aware representations.
mlx-vis: Python library implementing GPU-accelerated dimensionality reduction (UMAP, t-SNE, PaCMAP, etc.) on Apple Silicon via MLX.
Method for detecting visual hallucinations in VLM outputs on cartoon character images using pose information.
BInD: Diffusion model for multi-objective structure-based drug design balancing molecular generation with protein interaction requirements.
Philosophical investigation of LLMs' ontological status as agents, analyzing architecture, training, and extensions enabling agent-like behavior.
FALCON: Self-supervised video pretraining for UAV action recognition addressing spatial imbalance in aerial footage with object-centric learning.
Self-supervised seismic data reconstruction method using self-consistency learning for handling irregularly distributed seismic receiver data.
Survey on LLMs transforming scientific research covering literature search, hypothesis generation, experimentation, content generation, and peer review assistance.
Research paper proposing reward modeling with chain-of-thought reasoning for improved LLM alignment with human preferences via reinforcement learning.
ContextBench: benchmark for generating targeted linguistically fluent inputs that activate specific latent features in language models for safety analysis.
Sysformer: method for safeguarding frozen LLMs using adaptive system prompts to ensure safety compliance without model retraining in deployment scenarios.
SPoT: tokenization strategy for Vision Transformers enabling continuous subpixel token placement instead of grid-based constraints for sparse regime exploitation.
Cross-attention analysis in transformers for interpreting TCR-pMHC binding predictions using TULIP model for understanding immune system mechanisms.
Multi-agent RL framework with student-teacher curriculum for autonomous driving behavior generation addressing complex real-world traffic scenarios and critical situations.
Multivariate fields of experts framework for learning image priors using Moreau envelopes for inverse problems including denoising, deblurring, and MRI reconstruction.
Kernel VICReg: self-supervised learning method operating in reproducing kernel Hilbert space to capture nonlinear dependencies in representation learning.
VEGA: AI agent for electric vehicle routing combining physics-informed neural operators with reinforcement learning for energy-aware charge-conscious path planning.
Tensor Atomic Cluster Expansion (TACE): equivariant atomistic machine learning in Cartesian space unifying scalar and tensorial modeling for chemistry applications.
Taxonomy-aware dynamic motion generation for robots using hyperbolic manifolds to incorporate hierarchical biomechanical structure into movement models.
Self-speculative masked diffusions: discrete data generative models reducing function evaluations through speculative sampling without factorization approximations.
TCR-EML: explainable machine learning model layers for predicting T cell receptor-peptide MHC binding with interpretability for immunotherapy applications.