archive
Every paper Pith has read. Search by title, abstract, or pith.
2684 papers in stat.ML · page 7
-
Reference pool of failures controls AI release errors on hard tasks
When Should an AI Workflow Release? Always-Valid Inference for Black-Box Generate-Verify Systems
-
Strong model learns task by eliciting pre-trained features from weak outputs
The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledge
-
-
Digital twins build synthetic controls for single-arm trials
Digital Twins as Synthetic Controls in Single-Arm Trials
-
Bias from thresholding predicted by residual score variance
When to Trust Confidence Thresholding: Calibration Diagnostics for Pseudo-Labelled Regression
-
Supply chain digital twin tests foundation models on logistics data
ISOMORPH: A Supply Chain Digital Twin for Simulation, Dataset Generation, and Forecasting Benchmarks
-
Neural SDE cuts yield curve forecast error to 6.58 bps
Yield Curves Dynamics Using Variational Autoencoders Under No-arbitrage
-
Symmetries in token targets force circulant logit matrices in LLMs
Uncovering Symmetry Transfer in Large Language Models via Layer-Peeled Optimization
-
Task structures and latents identifiable without supervision
From Generalist to Specialist Representation
-
Optimal tests aggregate into log-optimal e-processes
Optimal sequential tests yield log-optimal e-processes
-
Gap counting sets critical scale for attention softmax
A Unified Framework for Critical Scaling of Inverse Temperature in Self-Attention
4 Piths -
Online optimization produces nested conformal sets at all levels
Online Conformal Prediction: Enforcing monotonicity via Online Optimization
-
Inference-time optimization lifts RL trading returns without retraining
Plan Before You Trade: Inference-Time Optimization for RL Trading Agents
-
KANs gain first risk bounds for mini-batch DP-SGD with correlated noise
Population Risk Bounds for Kolmogorov-Arnold Networks Trained by DP-SGD with Correlated Noise
-
Orthogonal transformations keep weight singular values fixed
Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation
-
Sampler matches smooth-case rate for composite log-concave densities
A proximal gradient algorithm for composite log-concave sampling
-
Bootstrap yields valid CIs for offline RL value functions
Model-based Bootstrap of Controlled Markov Chains
-
Conformal prediction optimizes sets without data splits
Multi-Variable Conformal Prediction: Optimizing Prediction Sets without Data Splitting
-
Online deferral to varying experts yields T to the two-thirds regret
Online Learning-to-Defer with Varying Experts
-
Online deferral algorithm manages varying experts with sublinear regret
Online Learning-to-Defer with Varying Experts
-
Budget-coverage policy learning reduces to affine threshold rule
Optimal Policy Learning under Budget and Coverage Constraints
-
Self-predicted data calibrates Bayesian regression better than Laplace
Self-Supervised Laplace Approximation for Bayesian Uncertainty Quantification
-
Sequential CMI bounds adaptive generalization gaps
Information-Theoretic Generalization Bounds for Sequential Decision Making
-
Score gradients cut simulation needs for neural surrogates
Keeping Score: Efficiency Improvements in Neural Likelihood Surrogate Training via Score-Augmented Loss Functions
-
Laplacian neural operators learn PDE maps with polynomial cost
Approximation Theory of Laplacian-Based Neural Operators for Reaction-Diffusion System
-
GNNs output random sets over classes to quantify epistemic uncertainty
Random-Set Graph Neural Networks
-
Anchor quantization stabilizes Schrödinger bridge couplings
QDSB: Quantized Diffusion Schr\"odinger Bridges
-
Random sphere points give strong quantization for moderate n
Non-asymptotic quantisation of spherically symmetric distributions
-
LOFT improves orthogonal fine-tuning via task-aware support selection
LOFT: Low-Rank Orthogonal Fine-Tuning via Task-Aware Support Selection
-
Two anchors make reward variance identifiable from preferences
Variance-aware Reward Modeling with Anchor Guidance
-
Kernel eigenvalue decay determines random forest rates
Minimax Rates and Spectral Distillation for Tree Ensembles
-
W-Flow reaches 1.29 FID in one ImageNet generation step
One-Step Generative Modeling via Wasserstein Gradient Flows
-
Sparse Bayesian KANs achieve near-minimax contraction
Posterior Contraction Rates for Sparse Kolmogorov-Arnold Networks in Anisotropic Besov Spaces
-
Active label queries cut U-statistic variance with fixed budget
Learning U-Statistics with Active Inference
-
Noise-subspace estimator matches minimax rate for probabilistic PLS
Exact Stiefel Optimization for Probabilistic PLS: Closed-Form Updates, Error Bounds, and Calibrated Uncertainty
-
Composite function stabilizes training of binary-activation networks
A Composite Activation Function for Learning Stable Binary Representations
-
Post-ADC inference restores valid stats after adaptive sampling
Post-ADC Inference: Valid Inference After Active Data Collection
-
Calibration algorithms adapt error bounds to unknown non-stationarity
Adaptive Calibration in Non-Stationary Environments
-
Algorithms adapt calibration error to unknown non-stationarity
Adaptive Calibration in Non-Stationary Environments
-
Vector codebook cuts KV cache to 34x compression at 0.95 similarity
FibQuant: Universal Vector Quantization for Random-Access KV-Cache Compression
-
Barrier smoothing yields O(K^{-2/3}) stationarity for constrained bilevel opt
A Barrier-Metric First-Order Method for Linearly Constrained Bilevel Optimization
-
PPO reformulated to beat SAC in multi-task RL
TOPPO: Rethinking PPO for Multi-Task Reinforcement Learning with Critic Balancing
-
Adapter adds closed-form spatial covariance to frozen predictors
Spatial Adapter: Structured Spatial Decomposition and Closed-Form Covariance for Frozen Predictors
-
Causal model recovers recourse effects from observational data
Causal Algorithmic Recourse: Foundations and Methods
-
Decompositions split generative AI bias into pathways and mechanism shifts
Causal Bias Detection in Generative Artificial Intelligence
-
Decompositions isolate bias pathways in generative models
Causal Bias Detection in Generative Artificial Intelligence
-
Causal paths break down survival disparities over time
Causal Fairness for Survival Analysis
-
Algorithm identifies ε-good subtrees without knowing ε
$\varepsilon$-Good Action Identification in Fixed-Budget Monte Carlo Tree Search
-
Coupled noises lift diversity in diffusion batches at zero added cost
Couple to Control: Joint Initial Noise Design in Diffusion Models
-
Dual form computes influence functions from data size not parameters
Extending Kernel Trick to Influence Functions