archive
Every paper Pith has read. Search by title, abstract, or pith.
14903 papers in cs.LG · page 13
-
RoPeSLR cuts DiT FLOPs 10x at 90% sparsity
RoPeSLR: 3D RoPE-driven Sparse-LowRank Attention for Efficient Diffusion Transformers
-
Reflector embeds reflection to block indirect jailbreaks
REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak
-
Early entropy drop signals when CoT reasoning helps LLMs
When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions
-
Attention model doubles perfect multi-user Wi-Fi activity predictions
AMAR: Lightweight Attention-Based Multi-User Activity Recognition from Wi-Fi CSI
-
RL method produces ready-to-bend pipes for aeroengines
Design for Manufacturing: A Manufacturability Knowledge-Integrated Reinforcement Learning Framework for Free-Form Pipe Routing in Aeroengines
-
Self-distillation balances consensus across views to cut noise from privileged signals
AVSD: Adaptive-View Self-Distillation by Balancing Consensus and Teacher-Specific Privileged Signals
-
Hard labels beat soft labels with sparse annotator votes
Same Target, Different Basins: Hard vs. Soft Labels for Annotator Distributions
-
LLM compilation creates hidden backdoor attack surface
Trusted Weights, Treacherous Optimizations? Optimization-Triggered Backdoor Attacks on LLMs
-
Weak-form latent models cut PDE optimization time by five orders
Time-Dependent PDE-Constrained Optimization via Weak-Form Latent Dynamics
-
Localization method builds Transformers from local kernels
The General Theory of Localization Methods
-
Autoregressive diffusion cuts video restoration latency to seconds
Accelerating Video Inverse Problem Solvers with Autoregressive Diffusion Models
-
Local updates cut Shapley recompute cost by 1000 times
Dynamic Shapley Computation
-
CDF inversion fixes uneven Pareto front sampling
SURF: Steering the Scalarization Weight to Uniformly Traverse the Pareto Front
-
Nested concept models reduce intervention costs to O(log K)
Matryoshka Concept Bottleneck Models
-
Latent analogies compose optimal plans for unseen goals in offline RL
Compositional Transduction with Latent Analogies for Offline Goal-Conditioned Reinforcement Learning
-
Vision model separates content from style to assure landing safety
Mechanistic Interpretability for Learning Assurance of a Vision-Based Landing System
-
Self-training amplifies surface markers while deep syntax dies
Self-Training Doesn't Flatten Language -- It Restructures It: Surface Markers Amplify While Deep Syntax Dies
-
Failure notes lift diagnostic AI accuracy up to 7%
MedExpMem: Adapting Experience Memory for Differential Diagnosis
-
Unlearning by shifting erased points to retained semantic neighbors
Approximate Machine Unlearning through Manifold Representation Forgetting Guided by Self Mode Connectivity
-
Adaptive kernels and LOOCV improve RBF KAN models
Adaptive RBF-KAN: A Comparative Evaluation of Dynamic Shape Parameters in Kolmogorov-Arnold Networks
-
Five features and six moves classify upper-limb EMG for prosthetics
Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven
-
Reversed updates raise Q-learning rewards from 9% to 79% in hard MDPs
ReversedQ: Opportunities for Faster Q-Learning in Episodic Online Reinforcement Learning
-
Three-stream GNN cuts MLIP energy errors by 57% at 20K samples
TriForces: Augmenting Atomistic GNNs for Transferable Representations
-
AI surrogate emulates ocean tipping 465 times faster
Deep Learning Surrogates for Emulating Stochastic Climate Tipping Dynamics
-
JAX simulator runs Mahjong at 2 million steps per second
Mahjax: A GPU-Accelerated Mahjong Simulator for Reinforcement Learning in JAX
-
Small models copy last CoT number for 89-92% of arithmetic accuracy
The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models
-
State management beats workspace isolation in multi-agent tasks
Multi-agent Collaboration with State Management
-
Overlapping nuclear norms recover subgroup low-rank geometry
Group-Aware Matrix Estimation and Latent Subspace Recovery
-
Logit averaging in GRPO matches KL-regularized accuracy
Complementing reinforcement learning with SFT through logit averaging in the post training of LLMs
-
Bandits learn smooth graph payoffs scaling only with effective dimension
Spectral bandits for smooth graph functions with applications in recommender systems
-
Learn image-space generators matching latent-process marginals
Latent Process Generator Matching
-
Transfer learning reaches O(m^(-(α+1)/d)) rate for d>3
Sample Complexity of Transfer Learning: An Optimal Transport Approach
-
Open seismic dataset trains generative models for inversion
OpenSeisML: Open Large-Scale Real Seismic and well-log Dataset for Generative AI
-
Geometric axioms explain neural network mechanisms
Axiomatizing Neural Networks via Pursuit of Subspaces
-
SVD preconditioning beats full fine-tuning at equal parameter count
FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning
-
Optimizer mixes local and global moments to blend AdamW and SGD
Ada2MS: A Hybrid Optimization Algorithm Based on Exponential Mixing of Elementwise and Global Second-Moment Estimates
-
Proofs verified by checking natural language modules separately
Pseudo-Formalization for Automatic Proof Verification
-
LLM agent accuracy drops to 0.54-0.62 without labels
AgentAtlas: Beyond Outcome Leaderboards for LLM Agents
-
Tri-stage training cuts multimodal edge energy by 33x
FusionSense: Tri-Stage Near-Sensor Learning for Runtime-Adaptive Multimodal Edge Intelligence
-
AI models lag behind text-only on 3D brain MRI benchmark
NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding
5 Piths -
Compact neural net edges FIB-4 on advanced MASLD fibrosis detection
Machine-Learning-Enhanced Non-Invasive Testing for MASLD Fibrosis: Shallow-Deep Neural Networks Versus FIB-4, Tabular Foundation Models, and Large Language Models
-
Quadratic approx yields private fine-tuning via exact normal sampling
An exponential mechanism based on quadratic approximations for fine-tuning machine learning models with privacy guarantees
-
Online conformal prediction can keep its calibration guarantees when feedback about past…
Online Conformal Prediction with Corrupted Feedback
-
Neurons encode exact Maxwell solutions for fast sparse field reconstruction
Fast Reconstruction of Exact Maxwell Dynamics from Sparse Data
-
Verbal feedback in RL makes LLM simulations more human-like
Reinforcing Human Behavior Simulation via Verbal Feedback
-
Min-gate fuses diffusion models to catch all four OOD shifts
Tippett-minimum Fusion of Representation-space Diffusion Models for Multi-Encoder Out-of-Distribution Detection
-
10,000-year cyclone catalog reproduces observed track densities
A 10,000-Year Global Stochastic Tropical Cyclone Catalog with Wind-Dependent Track Transitions (WHITS)
-
New metrics score uncertainty-augmented systems as one proper rule
ECUAS$_n$: A family of metrics for principled evaluation of uncertainty-augmented systems
-
ECUAS_n metrics score uncertainty-augmented systems with one tunable rule
ECUAS$_n$: A family of metrics for principled evaluation of uncertainty-augmented systems
-
ZEBRA keeps 94% of quality on half an LLM budget
ZEBRA: Zero-shot Budgeted Resource Allocation for LLM Orchestration