Every paper Pith has read and judged, in one searchable index.

56,949 reviewed papers in cs.LG · newest first · page 2

Pith rank means recent reader upvotes with each upvote's ranking weight halved after 24 hours. The archive itself is sorted by arXiv submission date.

51. 51
    cs.LG arXiv submitted 2026-09-23
    A spectral fix lifts attention accuracy in 12 of 12 runs

    Xiaohe Jiang (1) +3 · “NS-ATTENTION: Newton-Schulz Transformations of Attention Outputs in Vision Transformers”

    2609.27735
52. 52
    cs.LG arXiv submitted 2026-09-23
    Risk-gated adversaries win on all four MuJoCo control tasks

    Jiaxi Wu +4 · “Robust Adversarial Reinforcement Learning with Risk Sensitivity and Critic Consistency Regularization”

    2609.27667
53. 53
    cs.LG arXiv submitted 2026-09-23
    Recursive gradients shrink privacy noise in decentralized learning

    Yizhao Fan +2 · “Private Decentralized Optimization with Noise Reduction and Bias Correction”

    2609.27658
54. 54
    cs.LG arXiv submitted 2026-09-23
    Memory of failed tries turns greedy decoding into search

    Oleksii Streltsov +1 · “FLEET: From Logits Entropy to Enhanced Trajectories in Text Generation”

    2609.27657
55. 55
    stat.ML arXiv submitted 2026-09-23
    Federated income scoring trails pooled data by just 1.8%

    Sultan Amed +2 · “FedIncome: Federated Learning for Income Estimation in Digital Lending Under Data Sovereignty Constraints”

    2609.27654
56. 56
    cs.LG arXiv submitted 2026-09-23
    LoReST cuts large-scale traffic forecast error on every tested network

    Qi Feng +6 · “Learning Local Heterogeneity and Cross-Region Context for Large-Scale Traffic Forecasting”

    2609.27637
57. 57
    cs.LG arXiv submitted 2026-09-23
    Clustering makes bandit sketching exact

    Hantao Yang +2 · “Efficient Linear Bandits via Cluster-Aware Sketching”

    2609.27594
58. 58
    cs.LG arXiv submitted 2026-09-23
    Concept erasure hides its target

    Akash Samanta +2 · “Hidden not Deleted: How Networks Suppress Entangled Features”

    2609.27593
59. 59
    cs.LG arXiv submitted 2026-09-23
    Below 59M parameters, Step Law's form survives

    Egor Romanyukov +5 · “Does Step Law Transfer to Small-Scale Language Models? An Empirical Recalibration Below 59M Parameters”

    2609.27581
60. 60
    cs.LG arXiv submitted 2026-09-23
    Gradient noise sets a per-modality momentum for joint training

    Zhongjing Gu +5 · “VCMM: Variance-Calibrated Momentum for Multimodal Learning”

    2609.27577
61. 61
    cs.LG arXiv submitted 2026-09-23
    A 4B model approaches 235B once the reward judge co-evolves

    Henan Sun +6 · “DCRL: Decoupling and Coupling Reinforcement Learning via Policy-Reward Manifold Alignment”

    2609.27572
62. 62
    cs.LG arXiv submitted 2026-09-23
    Data-fit neuron formulas beat generic ones in every test

    Meng Wang +7 · “TNLearn: An Open Source Python Package for Task-based Neurons”

    2609.27564
63. 63
    cs.LG arXiv submitted 2026-09-23
    Treating physics data as PDE fields beats tokenizing it

    Henan Sun +6 · “PhyMo: A Physical-Field Modality for Multimodal AI4Physics”

    2609.27554
64. 64
    cs.LG arXiv submitted 2026-09-23
    Idle GPUs reclaimed for 3.5x embodied RL training speed

    Liang Mi +14 · “EBRL: Asynchronous Embodied RL by Multi-Grained Resource Management”

    2609.27547
65. 65
    math.ST arXiv submitted 2026-09-23
    Distribution shift hits diffusion scores as radius squared

    Wei Luo +3 · “Robustness of Diffusion Models under Distribution Shift”

    2609.27546
66. 66
    cs.LG arXiv submitted 2026-09-23
    Verified progress beats outcome rewards when credited per turn

    Ming Ma +9 · “ProCredit: From Outcome Rewards to Progress Credit in Agentic Reinforcement Learning”

    2609.27532
67. 67
    cs.CV arXiv submitted 2026-09-23
    Mammography encoder posts 97.78% validation accuracy

    Zheng Yu +4 · “M3D-Net: Hierarchical Coordination of Spatial Context, Feature Reuse, and Differential Attention for Mammography Classification”

    2609.27523
68. 68
    cs.AI arXiv submitted 2026-09-23
    Agents pick the best config while understanding almost nothing

    Jingjie Ning +3 · “WhatWorkedBench: Benchmarking Experimental Understanding in AI Agents”

    2609.27490
69. 69
    cs.CV arXiv submitted 2026-09-23
    State drift beats cache signals for streaming video memory

    Taeyoun Kwon +3 · “DeltaS: Reading the Gated Linear Attention State for KV Cache Eviction in Streaming Video”

    2609.27470
70. 70
    cs.LG arXiv submitted 2026-09-23
    Small quantum circuit learns cost-delay-aware cloud scheduling

    An N. H. Phan +3 · “Quantum Reinforcement Learning for Cost and Delay Tradeoffs in Quantum Cloud Orchestration”

    2609.27446
71. 71
    cs.LG arXiv submitted 2026-09-23
    A 1.5B student beats its 7B teacher on constraint following

    Yanzhao Zheng +8 · “Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following”

    2609.27421
72. 72
    cs.LG arXiv submitted 2026-09-23
    Small oscillatory model tops 8 rivals on scarce vibration labels

    Mainak Mallick +1 · “When Labels Are Scarce: An Oscillatory State Space Model for Vibration Diagnosis”

    2609.27411
73. 73
    cs.LG arXiv submitted 2026-09-23
    Active learning's label savings barely survive validation

    Ben McEwen +2 · “Active Learning for Biodiversity Monitoring: From Label Efficiency to Reliable Ecological Inference”

    2609.27409
74. 74
    cs.SD arXiv submitted 2026-09-23
    Self-evolving audio models gain up to 6.3 points

    Yuxiang Wang +8 · “EvoAudio: Recursive Self-Improvement for Audio Understanding”

    2609.27389
75. 75
    cs.LG arXiv submitted 2026-09-23
    An agent beat fixed forecast policies with 2.5% of the budget

    Shunya Nagashima · “Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools”

    2609.27385
76. 76
    cs.CL arXiv submitted 2026-09-23
    Recurrent language models keep recomputing global attention at every recurrent step

    Ke Wan +1 · “Attention Routing Stabilizes Early: Working-Set Inference for Recurrent Language Models”

    2609.27373
77. 77
    cs.LG arXiv submitted 2026-09-23
    AUC bound tunes anomaly detectors with zero anomaly labels

    Kevin Wilkinghoff +1 · “Anomaly-Free Self-Optimization via AUC Bounds”

    2609.27362
78. 78
    cs.LG arXiv submitted 2026-09-23
    Curvature ratio finds the weights that kill forgetting under 4-bit

    Jialu Wang +8 · “Quantization-Robust Unlearning through the Lens of Retain-Forget Loss Landscapes Interaction”

    2609.27355
79. 79
    eess.SY arXiv submitted 2026-09-23
    LLM-evolved code beats static 5G slicing by 44.5%

    Faezeh Dehghan Tarzjani +1 · “Evolving Inspectable O-RAN Slicing xApps with LLMs”

    2609.27337
80. 80
    cs.RO arXiv submitted 2026-09-23
    Decoupling safety from tactics yields unexploitable robots

    Ruihan Wu +4 · “Turning Safety into Competence: Minimally Exploitable Robot Policies via Safety-Filtered Reinforcement Learning”

    2609.27312
81. 81
    cs.SE arXiv submitted 2026-09-23
    Fairness test finds up to 412% more unfair MARL episodes

    Xiaotong Wang +1 · “FairTest: Search-Based Fairness Testing for Multi-Agent Reinforcement Learning Systems”

    2609.27309
82. 82
    cs.LG arXiv submitted 2026-09-23
    Normalized diffusion reaches 3D spin lattices

    Kewen Pan +1 · “Discrete Diffusion Models via Evolving Variational Autoregressive Networks”

    2609.27306
83. 83
    cs.LG arXiv submitted 2026-09-23
    One policy solves whether, when, whom, and what to say live

    Shujian Gao +8 · “Live Assistant: Learning Whether, When, and Whom to Assist in Real-World Live Social Streams”

    2609.27303
84. 84
    stat.ME arXiv submitted 2026-09-23
    More data can't cure IV exclusion bias

    Spandan Ghose Chowdhury · “Beyond the Illusion of Power: Calibrating Quasi-Experiments in Observational IS”

    2609.27299
85. 85
    cs.LG arXiv submitted 2026-09-23
    Cheaper prefill lets a 67B model beat 47B and 63B at equal compute

    Zhiheng Hu +10 · “KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling”

    2609.27294
86. 86
    cs.LG arXiv submitted 2026-09-23
    Network size becomes one differentiable count

    Lixing Li · “NGN: Learning Neural Network Size as a Differentiable Count”

    2609.27291
87. 87
    cs.LG arXiv submitted 2026-09-23
    Frozen LLM plus verified rules beats retrained fraud detectors

    Xuwei Tan +2 · “SR-Fraud: An Outcome-Supervised Reflective LLM Agent Framework for Non-Stationary Payment Fraud Detection”

    2609.27287
88. 88
    stat.ML arXiv submitted 2026-09-23
    Pairwise fusion hits the optimal multitask rate

    Xiaodong Li +1 · “Multitask Regression with Pairwise Fusion”

    2609.27280
89. 89
    cs.LG arXiv submitted 2026-09-23
    Spectral priors keep graph lasso convex and lift recovery

    Mingxiao Liu (1) +8 · “Graph Learning with Spectral Connectivity Priors for Scarce Data”

    2609.27278
90. 90
    cs.AI arXiv submitted 2026-09-23
    Self-grown tool library lifts accuracy on all 30 task runs

    Jie Yang +10 · “TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent”

    2609.27277
91. 91
    cs.LG arXiv submitted 2026-09-23
    Models converge on relations

    Junwon You +3 · “What Converges in the Platonic Representation Hypothesis? Structure over Geometry”

    2609.27252
92. 92
    cs.LG arXiv submitted 2026-09-23
    Attention masks turn an LLM into a near-perfect text autoencoder

    Arkanath Pathak +2 · “Repurposing Pre-trained LLMs as High Fidelity Continuous Text Autoencoders”

    2609.27248
93. 93
    cs.LG arXiv submitted 2026-09-23
    First full-covariance smoother for Bayesian neural nets

    Oren Wright +5 · “Full-Covariance Smoothing of Bayesian Neural Networks for Online Adaptation”

    2609.27244
94. 94
    stat.ML arXiv submitted 2026-09-23
    Membership queries make hard classes exponentially easy

    Ganghua Wang +1 · “On the Sample Complexity of Active Learning with Membership Queries”

    2609.27241
95. 95
    cs.LG arXiv submitted 2026-09-23
    Top scorer never uses its documented compound input

    Mengran Li +5 · “Discover, Falsify, Revise: Auditing Input-Use Claims from Source Code to Predictive Contribution in Agent-Discovered Cell Models”

    2609.27234
96. 96
    cs.LG arXiv submitted 2026-09-23
    Compute alone does not predict fMRI model performance

    Wenhao Ye +5 · “A Scaling Study for fMRI Foundation Models”

    2609.27232
97. 97
    cs.SD arXiv submitted 2026-09-23
    Lung-sound AI spots pneumonia in nursing homes

    Nicholas Rasmussen +7 · “Physiologically Informed Digital Auscultation for Pneumonia Detection in Long-term Care Residents”

    2609.27222
98. 98
    cs.LG arXiv submitted 2026-09-23
    Learning ellipsoid shape from the worst tail cuts severe misses

    Xiang Zhang · “Tail-Aware Geometry Learning for Conformal Ellipsoids”

    2609.27221
99. 99
    cs.CE arXiv submitted 2026-09-23
    Neural surrogate cuts topology-optimization time 15 to 110x

    Shengyu Yan +1 · “KATOsuper: Surrogate-accelerated neural topology optimization with sensitivity-consistent Fourier neural operators”

    2609.27216
100. 100
     cs.LG arXiv submitted 2026-09-23
     Curvature-guided sampling wins on six of seven GNN sets

     Chaoqun Fei +3 · “Scalable Subgraph Sampling via Resistance Curvature”

     2609.27209