Information-theoretic analysis of generalization in ML using lossy compression and finite blocklength analysis to derive lower bounds on sample complexity.
Proposes rationality measures and theory for RL agents, defining perfect rationality as maximizing hidden value functions and introducing value discrepancy metrics over deployment trajectories.
Proposes end-to-end compression techniques for tabular foundation models using in-context learning to improve efficiency while maintaining performance.
End-to-end compression techniques for tabular foundation models using transformer architecture. Compresses large models for in-context learning on tabular data.
Research on achieving Bayes-optimal binary classification while satisfying fairness constraints (statistical parity, equalized odds). ML fairness theory.
arXiv paper introducing Kinetic Path Energy diagnostic for analyzing trajectories in flow-based generative models.
arXiv paper on transformation inversion via diffusion sampling for recovering unknown group transformations in data.
arXiv paper addressing expression simplification bottleneck in amortized neural symbolic regression using normalization.
arXiv paper proposing framework for evaluating goal-directedness in LLM agents combining behavioral and representational analyses.
arXiv paper on automated detection of unverbalized biases in LLM reasoning using black-box pipeline.
arXiv paper analyzing what preference learning recovers from pairwise comparison data beyond Bradley-Terry model assumptions.
arXiv paper on token-efficient black-box detection of LLM API changes using statistical methods on output tokens.
arXiv paper studying weight decay effects on language model plasticity and downstream task adaptability.
arXiv paper on MeSP: memory-efficient structured backpropagation for on-device LLM fine-tuning on mobile with memory constraints.
arXiv paper on information geometry of softmax distributions for probing and steering AI model representations.
arXiv paper on hierarchical reinforcement learning for LLM agents in long-horizon sparse-reward tasks with explicit credit assignment.
arXiv paper analyzing induction bias limitations in transformers through in-distribution evaluation of state tracking capabilities.
arXiv paper on spectral graph neural networks using Cauchy factorizations for efficient local-to-global modeling.
arXiv paper on offline goal-conditioned reinforcement learning using physics-informed value estimation with Hamilton-Jacobi-Bellman constraints.
arXiv paper on prompt optimization for multi-agent systems using bandit algorithms and graph neural networks to improve LLM-based workflows.
arXiv paper on efficient super-resolution transformers using rank-factorized positional bias compatible with FlashAttention.
arXiv paper introducing flexible mirror descent optimization framework using group entropies for machine learning.
arXiv paper formulating LLM reasoning as optimal control problem, integrating planning into model architecture via test-time state optimization.
arXiv paper on Bayesian mixture-of-experts transformers using variational routing for uncertainty quantification in foundation models.
arXiv paper analyzing generative drifting for one-step image generation, showing equivalence to score matching under Gaussian kernels.
arXiv paper proposing H-EARS, physics-guided reward shaping for deep reinforcement learning to improve efficiency and generalization in continuous control.
arXiv paper introducing AxonAD for detecting multivariate time series anomalies by identifying shifts in cross-channel dependencies using attention mechanisms.
arXiv paper on regression-aware RL for LLM-as-a-Judge, improving numeric score prediction with ordinal structure.
arXiv paper on energy-based models for graph generation using transport-aligned energy matching.
arXiv paper modeling temporal uncertainty dynamics in probabilistic time series forecasting.
arXiv paper on perturbation approach to unconstrained linear bandits and online optimization.
arXiv paper on world models for robustness using forward-inverse asymmetry for policy learning.
arXiv paper improving neural operators for dynamical systems using mean-flow enhancement.
arXiv paper analyzing transformer in-context classification through symmetry and equivariance constraints.
arXiv paper on multi-objective Bayesian optimization via adaptive epsilon-constraint decomposition.
arXiv paper proposing survival value learning for goal-conditioned reinforcement learning with improved stability.
arXiv paper on conditional attribution framework for explaining time-series anomalies in complex systems.
arXiv paper using compiler outputs to improve LLM-based formal theorem proving by exploiting proof structure.
arXiv paper proposing Cost-Aware SGD algorithm for finite-sum objectives with non-uniform sampling costs.
Polaris: Polar hyperspherical embedding framework for learning hierarchical concept representations in knowledge structures.
Bayesian rain field reconstruction using commercial microwave links with diffusion model priors for weather sensing.
Theoretical analysis comparing DDPM and DDIM diffusion samplers, explaining why DDIM hallucination occurs in reverse dynamics.
Theoretical analysis of spurious correlation learning in DPO and preference optimization with mitigation strategies for LLMs.
Njord: Probabilistic graph neural network for ensemble ocean forecasting with probabilistic sampling capability.
Theoretical analysis of layer-specific learning rates in neural networks using Stackelberg optimization framework.
Differentiable Mixture-of-Agents framework enabling self-evolving multi-agent LLM systems with adaptive communication topologies.
FML-bench benchmark for evaluating AI research agent strategies by separating search topology from execution infrastructure.
IBAL framework for robust multi-agent RL defending against interaction-breaking adversarial attacks using information theory.
PROWL method for world model learning using KL-constrained adversarial approach to identify and train on rare critical transitions.
Block-based double decoders architecture combining benefits of encoder-decoder and decoder-only models with full loss supervision.