Proposes four-stage post-training workflow for LLM reasoning combining sparse-reward RL, forward-KL warmup, and on-policy distillation. Evaluated on math reasoning tasks.
ODRPO method for robust LLM alignment using ordinal decomposition of discrete rewards to handle stochastic multi-tier rating schemes.
Protocol-Driven Development framework using machine-executable protocols with invariants and continuous evidence to govern AI-generated software.
Reframing pairwise ranking prompting reranking as active learning problem to improve top-K ranking under noisy and intransitive LLM judgments.
Concept-level machine unlearning for vision-language models using interpretable concept decomposition to remove target knowledge precisely.
Method reframing GUI agent critique from binary classification to continuous semantic alignment for improved action ranking in test-time scaling.
MARS framework for memory-augmented LLM agents in recommendation, using hierarchical belief-state memory instead of flat representations.
Context-Driven Decomposition probe diagnosing when RAG systems prioritize retrieved context over parametric knowledge under conflicting information.
Study evaluating whether coding agents understand least-privilege authorization principles, introducing permission-boundary inference task.
Framework combining closed-loop reasoning and verification for complex text-to-image generation, addressing hallucinations and optimization instabilities.
Architectural solution enabling continual pre-training of large language models on heterogeneous domains without catastrophic forgetting, replay buffers, or task labels.
Method for early termination of local LLM agents to reduce energy consumption on consumer devices while preserving privacy and eliminating API costs.
Research on multi-agent LLM coordination identifying and addressing the compounding occupancy shift problem in sequential fine-tuning of shared-context teams.
Study showing post-training quantization of LLMs causes bias emergence and alignment degradation across model families and precision levels.
On-policy self-distillation approach to reduce performance degradation (safety tax) in aligned LLMs while maintaining safety.
Privacy evaluation framework for generative models (GANs, VAEs, diffusion) trained on trajectory data.
GQLA: hardware-adaptive attention mechanism for LLM decoding supporting multi-token prediction on commodity hardware.
Curriculum learning strategy for Physics-Informed Neural Networks based on spatial correlation to improve PINN training.
Study showing LLMs can obfuscate chain-of-thought reasoning when trained on monitoring documents, posing detection evasion risks.
Tadpole: foundation model for 3D PDEs using autoencoders with online data generation for improved transferability and scalability.
Maximal update parameterization (μP) extension for grouped query attention enabling hyperparameter transfer across LLM architectures.
GESD framework for measuring explanation stability disparities in ML fairness, extending beyond outcome-oriented fairness metrics.
Discovery Agent: LLM-based system for IO2Code program synthesis from input-output examples, advancing beyond NL2Code.
LEAP evaluates LLMs in iterative scientific design by measuring learning trajectory efficiency and convergence rates, not just final outcomes.
Controllable foundation model for molecular graph generation using RL post-training to optimize chemistry-specific design tasks.
SeqMem-Eval: diagnostic framework for evaluating LLM memory over sequential tasks, measuring forgetting and negative transfer beyond aggregate metrics.
LPDS evaluates LLM robustness by scaling problem difficulty while preserving logic, testing entity-invariant reasoning capabilities.
Tests Joint-Embedding Predictive Architecture principles for LLM fine-tuning on Llama-3.2, evaluating hidden-state geometry improvements.
φ-Balancing: principled framework for balanced expert utilization in Mixture-of-Experts models targeting population-level objectives.
Comprehensive empirical analysis of 118 transformer models revealing fundamental scalability performance walls and deployment constraints.
Proposes Margin-Adaptive Confidence Ranking to improve reliability of LLM judgments through learned confidence estimation and hypothesis testing.
Introduces f-Trajectory Balance loss for training GFlowNets and LLMs with mixed on/off-policy data, generalizing MSE approaches.
Proposes Lagrangian Flow Matching framework for designing probability paths in generative models beyond straight-line transport.
DualKV optimizes Flash Attention for RL post-training by sharing prompt tokens across rollouts, reducing compute/memory redundancy in GRPO and DAPO.
Studies structural plasticity during training via network growth and pruning. Shows growth is not inverse of pruning in deep learning.
Sampling method for loss minima using Riemannian mechanics. Characterizes reparameterization invariant solutions in neural networks.
DrugSAGE: Self-evolving agent that accumulates experience for efficient drug discovery. Reuses past search knowledge to find SOTA models across tasks.
Layer-wise Derivative Controlled Networks balance accuracy, efficiency, and stability. Addresses spiky/unpredictable behavior in complex models.
GRLO: Framework for generalizable reinforcement learning in open-ended environments. Integrates RLHF and RLVR paradigms for LLM post-training.
NormWear-2 world model for physiological signal prediction. Encodes multivariate health signals with chaos-theoretic approach for long-horizon forecasting.
Njord: Probabilistic graph neural network for ocean forecasting. Combines latent variable framework with GNN for ensemble weather prediction.
Energy-based models for inverse problems in imaging with explicit prior density. Addresses diffusion model limitations for ill-posed problems.
SurvivalPFN applies prior-data fitted networks for survival analysis. Amortizes Bayesian inference for censored observations via in-context learning.
Ghosted Layers: Training-free method to recover performance of layer-pruned LLMs by solving boundary activation alignment. Closed-form solution for layer mismatch.
STS: Sparse attention mechanism for LLM inference without retraining. Enables efficient processing of multi-million token sequences for agentic applications.
Framework unifying tensor-parallel RL environments, control barrier functions, and safety-constrained training for UAVs. Open-source safety filter for reinforcement learning.
Benchmark framework for Graph Neural Networks on biological omics data. Addresses n<<p regime overlooked by existing GNN benchmarks.
Reveals vulnerabilities in data attribution methods under distributed training. Shows single participants can inflate attribution values while preserving model performance.
Introduces differential forms for point cloud learning to capture higher-order geometric information. Proposes neural point-forms architecture.
Analyzes non-uniform layer-specific learning rates using Stackelberg game theory. Provides principled understanding of when different layers benefit from different learning rates.