INDEX

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

54,049 reviewed papers in cs.LG · newest first · page 1

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.

cs.AI arXiv submitted 2026-08-27

Learned sepsis score separates survivors within every severity stratum
Kevin Zhu +20 · “Learning a Continuous Sepsis Severity Score Without Hour-by-Hour Supervision: A Two-Site Retrospective Study”

2608.27421
cs.IR arXiv submitted 2026-08-27

Cut GNN ID-table memory 98% and lift friend adds 16%
Maksim Utushkin +2 · “Scaling Graph Neural Networks for Friend Recommendation: Multi-Hash User Embeddings and Temporal Neighbor Sampling”

2608.27413
cs.AI arXiv submitted 2026-08-27

CorporateBench: LLM accuracy drops as document sets grow past 230,000
Sil Hamilton +6 · “CorporateBench: Large-Scale Q&A Benchmarking with Temporal Knowledge Bases”

2608.27391
cs.GT arXiv submitted 2026-08-27

Token-by-token ad bidding is manipulation-proof and near-optimal
Hanbing Liu +4 · “Token-Level Advertising”

2608.27382
math.PR arXiv submitted 2026-08-27

At n = d²/4, ellipsoid fitting flips from always to never
Frederic Koehler +1 · “Universality and sharp thresholds for ellipsoid fitting”

2608.27372
cs.CL arXiv submitted 2026-08-27

Under $6,900 of consumer-GPU time yields a 2B model near Qwen2.5-1.5B
Kairong Luo +10 · “Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090”

2608.27370
cs.LG arXiv submitted 2026-08-27

ES gives LLMs broader reasoning coverage than GRPO
Yunpeng Ba +9 · “Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO”

2608.27351
cs.LG arXiv submitted 2026-08-27

Block drafters hit a 71% acceptance ceiling
Xinwei Qiang +4 · “Beyond Parallel Blindness: Information Floors and Model Gaps in Block Drafting”

2608.27339
stat.ML arXiv submitted 2026-08-27

Quantile TD error loses quantile-count dependence after burn-in
Zijie Cheng +3 · “A Finite Sample Analysis for Quantile Temporal Difference Learning in Distributional Reinforcement Learning”

2608.27313
cs.LG arXiv submitted 2026-08-27

Two-phase quantum training tops classical models on echo view ID
Mihai Udrescu-Milosav +3 · “QuantumBoostNet: Hybrid Classical-Quantum Cardiac View Identification”

2608.27302
cs.AI arXiv submitted 2026-08-27

One LLM query designs algorithms that beat specialist OR methods
Jackie Baek · “LLMs Can Design Near-Optimal OR Algorithms”

2608.27296
stat.ML arXiv submitted 2026-08-27

From audio alone, a model finds critic-linked artists at AUC 0.767
Elena Badillo-Goicoechea +1 · “Recovering Expert Critic-Sourced Network Adjacency between Musical Artists from Acoustic Distributions: A Construct-Validity Approach”

2608.27291
cs.LG arXiv submitted 2026-08-27

NMR+IR+MS accuracy jumps from 44% to 76% via expert routing
Hai-Tao Yu +6 · “MM-Spectrum: Multimodal Multi-spectral Molecular Structural Elucidation with a Stable MoE Framework”

2608.27286
cond-mat.dis-nn arXiv submitted 2026-08-27

A looped transformer that repeatedly applies the same network can solve problems using…
Gunn Kim · “Dynamical phase selection controls compute scaling in looped transformers”

2608.26556
cs.CV arXiv submitted 2026-08-26

300 generated tasks make visual reasoning trainable and verifiable
Junxiang Xu +51 · “VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning”

2608.26105
cs.CV arXiv submitted 2026-08-26

Muscle signals lift exercise-form scoring past vision-only AI
Hao Yin +8 · “MyoMechanix: Biomechanically-Grounded Compositional Skilled Activity Understanding and Coaching”

2608.26094
cs.LG arXiv submitted 2026-08-26

AI agent designs cell-edge power control at 99.5% of best solver
Ahmad Khan +3 · “Agentic Autoresearch for Cell-Edge Power Control: Radically Redefining the Researcher's Role”

2608.26093
astro-ph.HE arXiv submitted 2026-08-26

Sparse AI features cut neutrino angular error from 20° to 3.2°
Rapha\"el Bonnet-Guerrini +3 · “Finding and using interpretable latents in a neutrino foundation model with sparse autoencoders”

2608.26090
cs.AI arXiv submitted 2026-08-26

Autonomous engine beats expert geospatial models
Evelyn Ma +27 · “Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings”

2608.26088
cs.LG arXiv submitted 2026-08-26

TraceML finds AI agents loop in place
Jiarui Yan +4 · “TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development”

2608.26086
cs.LG arXiv submitted 2026-08-26

ICON distinguishes used concepts from correlated decoys
Roshan Prakash Rane +9 · “ICON Decomposition: Auditing Deep Neural Networks with Multivariate Variance-based Concept-level Explanations”

2608.26083
cs.CL arXiv submitted 2026-08-26

Dropping mid-reasoning tokens makes models 3x faster
Niklas Muennighoff +17 · “Prefix Sliding for efficient test-time scaling”

2608.26070
cs.LG arXiv submitted 2026-08-26

Compress large-kernel CNN weights 81% via shared low-rank factors
Hao Luo +12 · “Group-Shared Low-Rank Approximation for Mobile-Efficient Pointwise Convolutions in Large-Kernel CNNs”

2608.26069
cs.LG arXiv submitted 2026-08-26

LoRA rank needs follow the task-weighted spectral tail
Gerard Conangla Planes · “How Much Rank Does LoRA Need? Rank-Error Bounds for Transformer Attention”

2608.26052
cs.LG arXiv submitted 2026-08-26

Curve-built experts lift Union accuracy by 2.4 points
Xu Zhang +1 · “Robust CurveMoE: Multi-Norm Adversarial Defense for Mixture-of-Experts Models via Mode Connectivity”

2608.26043
cs.AI arXiv submitted 2026-08-26

Learned proof policies beat leanCoP by up to 46%
Fredrik R{\o}mming +3 · “Imitation Learning for Connection-Tableau Construction”

2608.26009
eess.SP arXiv submitted 2026-08-26

Fusion gates spot lost sensors
Navaneetha Krishnan Kamalakannan +1 · “CardioFusion-AI: Robust ECG--PPG Fusion for Multimodal Physiological Monitoring Under Signal Degradation”

2608.26000
cs.LG arXiv submitted 2026-08-26

One number predicts how pruning degrades SAE interpretability
Suchit Gupte +2 · “When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs”

2608.25941
cs.LG arXiv submitted 2026-08-26

Three levers govern AI self-distillation collapse
Justin Robert +1 · “One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation”

2608.25936
cs.AI arXiv submitted 2026-08-26

Best fact-checking model drops from 0.70 to 0.31 across datasets
Aida Usmanova +3 · “How Robust Are Automated Fact-Checking Systems? A Cross-Benchmark Evaluation”

2608.25934
physics.geo-ph arXiv submitted 2026-08-26

Neural surrogate inverts 2M EM soundings in seconds
Jaehong Chung +2 · “Continually learning neural-operator surrogate for three-dimensional airborne electromagnetic Bayesian inversion”

2608.25932
stat.ME arXiv submitted 2026-08-26

Shortcut learning is omitted variable bias—last-layer refit removes it
Manuel Pfeuffer +3 · “Controlling for Omitted Variable Bias in Deep Neural Networks”

2608.25930
cs.CL arXiv submitted 2026-08-26

ML models become native SystemC-AMS simulation blocks
Andrei Mihai Albu +1 · “SAMpLE: A SystemC-AMS Machine LEarning-based Framework for Virtual Prototyping”

2608.25910
cs.LG arXiv submitted 2026-08-26

Unsupervised model splits drivers into three traffic regimes
Mohammad Elayan +2 · “Quantum-Inspired Modeling of Driving Behavior”

2608.25907
cs.DB arXiv submitted 2026-08-26

SQL metapath pruning cuts GNN training time by up to 10x
Fahim Shahriar Khan +1 · “MetaSieve: Faster Relational Deep Learning through SQL-Based Metapath Selection”

2608.25903
cs.LG arXiv submitted 2026-08-26

Score-trained beliefs drop the loss-weight search
Pavel Prochazka · “Forecasting Multiple Observables with SCROLL: Score-Trained Uncertainty for Stochastic Dynamics”

2608.25898
cs.LG arXiv submitted 2026-08-26

Unified loss gives faithful attributions and stable counterfactuals
Xu Zheng +9 · “Towards A Unified Information Bottleneck Framework for Time Series Explanations”

2608.25897
cs.LG arXiv submitted 2026-08-26

Fine-tuned 3D molecular model transfers across four smell tasks
Yikun Han +4 · “A General-Purpose Molecular Foundation Model Transfers Across Diverse Olfactory Tasks”

2608.25893
cs.DC arXiv submitted 2026-08-26

RNN load balancer cuts multi-GPU tissue-sim imbalance to 3.5%
Matvey Moisseyev +4 · “Scalable Multi-GPU Simulation of 3D Multicellular Growth with RNN-Based Workload Balancing”

2608.25890
stat.ML arXiv submitted 2026-08-26

Nearest-neighbour matrix estimates the Density Information Matrix
David P. Hofmeyr · “Efficient Estimation of High Information Projections using Nearest Neighbours”

2608.25887
cs.LG arXiv submitted 2026-08-26

SCAFFOLD loses to FedAvg because of edge-of-stability sharpness
Anant Khandelwal +2 · “How Edge of Stability Hinders SCAFFOLD in Federated Optimization”

2608.25873
cs.LG arXiv submitted 2026-08-26

Budget-aware forecaster beats passive baselines
Junjie Meng +8 · “CEDAR: Controlled and Event-Driven Demand Forecasting via Residual Decomposition”

2608.25871
cs.CV arXiv submitted 2026-08-26

Cross-attention beats concatenation for extreme rainfall downscaling
Victor Nascimento Ribeiro +9 · “Precipitation Downscaling Using Foundation Model-Conditioned Diffusion”

2608.25858
cs.CL arXiv submitted 2026-08-26

Re-annotated key points beat original in all 15 human evals
Zhiqiang Shi +1 · “Key Point Analysis Needs Structure Recovery: Task Definition, Dataset Diagnosis, and a Structure-Aware Benchmark”

2608.25854
eess.AS arXiv submitted 2026-08-26

Cough-based TB models don't transfer across datasets
Wensi Zhang +3 · “Why ML-based cough models do not generalize: a systematic cross-dataset evaluation for tuberculosis screening”

2608.25846
cs.LG arXiv submitted 2026-08-26

Closed-loop validation lifts drug-synergy region recall to 0.826
Fan-Sheng Chuang +3 · “VINCENT: Validated Interaction Network for Cross-drug Explanation of Therapeutics”

2608.25841
cs.CL arXiv submitted 2026-08-26

Same model, same rules, different skill by language
Bobby Cheng +6 · “Skill Issue: Are Skills Language-Invariant in LLMs?”

2608.25832
cs.CV arXiv submitted 2026-08-26

FlowMoDL beats all baselines on 4D flow MRI from 10x to 50x
Tristan Gottwald +8 · “FlowMoDL: Model-Based Deep Learning with Conjugate-Gradient Data Consistency for Highly Accelerated 4D Flow MRI Reconstruction”

2608.25828
cs.CL arXiv submitted 2026-08-26

Unfolding papers into writing plans beats training on plain text
Qiankai Xu +6 · “Unfolding Scientific Papers into Multi-Turn Generation Trajectories for Continued Pre-Training”

2608.25826
cs.LG arXiv submitted 2026-08-26

One neural field answers any lead time and grid for temperature forecasts
Chunlei Shi +6 · “Learning Continuous Regional Temperature Fields with Lead-Time and Resolution Queries”

2608.25823