Exact risk ratios for data selection in linear regression
Proves 1+1/d at 2d-1, 5/3 for (3,4), 2 for (4,5); gives lower bound.
· “Exact Risk Ratios for Weighted Data Selection in Linear Regression”
cs.LG
Machine Learning
Papers on all aspects of machine learning research (supervised, unsupervised, reinforcement learning, bandit problems, and so on) including also robustness, explanation, fairness, and methodology. cs.LG is also an appropriate primary category for applications of machine learning methods.
sort pith recommended most recent
Reward-based model tuning becomes provable Newton steps
Fine-tuning and sampling share one iteration with global descent and quadratic convergence.
· “Newton Matching for Generative Modeling: A Unified Framework for Fine-Tuning and Sampling”
Fisher-information speed limit governs how fast SGD learns
Bound splits into drift and noise budgets; correctly predicts learning order in linear regression.
· “Speed Limit for Information Acquisition in Stochastic Learning Dynamics”
One sum decides which SGD anytime rates are possible
An overhead h works exactly when the sum of 1/h(2^j)² converges; the √(log n/n) rate stays just out of reach.
· “The Exact Time-Uniform Rate Frontier for Stochastic Gradient Descent on Smooth Convex Objectives”
Silver stepsize is optimal for gradient descent acceleration
New lower bounds match known upper bounds, pinning the fastest polynomial convergence exponent
· “Silver Rate Is (Almost) Optimal for Gradient Descent Acceleration”
Exact copy count for quantum tomography
Measuring k copies together buys a √k speedup, and no adaptive protocol can beat the rate.
· “Tight Lower Bounds for State Tomography with Limited Entanglement”
Model agreement zeroes noise in private text generation
Each generated token leaks at most bT of mutual information about the private secret; unanimous votes add no noise.
· “PAC-Private Autoregressive Generation: Calibrating Noise to Ensemble Disagreement”
Two feedback tokens per answer can match or beat full training
Training a student model on one or two tokens per response matches or beats dense supervision in nine configurations.
· “Extremely Sparse Supervision Incentivizes Reasoning Ability”
Sink Strength predicts text loss after vision-language tuning
A single scalar measured in seconds forecasts which LLMs keep instruction-following after VLM training
· “Text Capability Loss in Vision-Language Adaptation: An Attention-Sink Diagnosis”
OOD score instability has a closed form and a wrong sign
Distance-based scores flag the most reproducible verdicts as most uncertain, making abstention worse than random.
· “Verdict Instability of OOD Scores under Reference Resampling”
Pauli principle gives one full order of mixed Barron regularity
Index-set-dependent weights detect smoothness invisible to the isotropic Barron scale; the result is uniformly sharp.
· “Sharp Mixed Spectral Barron Regularity of Coulombic Many-Electron Wave Functions”
Graph theory sets the cost of learning tensor-network states
Cutwidth and tree-cutwidth govern both the compression to MPS or TTN and the sample complexity of tomography, in realisable and agnostic…
Matrix LASSO recovers at the sharpest possible threshold
For rank-r∗ targets, every global minimizer attains the optimal error at any search rank r ≥ r∗.
· “Sharp Restricted Isometry Thresholds for Global Minima of Rank-Restricted Matrix LASSO”
QSP yields a solvable model of quantum representation learning
For sparse data the full nonlinear dynamics reduces to independent scalar flows with explicit convergence times.
Independent learning in discounted games hits a quasi-polynomial barrier
Under ETH for PPAD, polynomial CCE computation is impossible; first radically uncoupled algorithm matches the hardness bound.
· “Independent Reinforcement Learning in Discounted Markov Games”
GNN beats traditional beamforming in wideband 6G
Bipartite message-passing GNNs outperform AMO and ICD, cutting inference time tenfold and resisting beam squint across 64 subcarriers.
· “Efficient Graph Neural Networks for Multicarrier Wideband Hybrid Beamforming Optimization”
Prune high-kernel-complexity edges to block attacks
Training-free preprocessing removes edges that inflate a generalization bound, improving GNN robustness across datasets.
· “Kernel-Complexity Edge Sanitization for Training-Free Defense against Structural Graph Attacks”
DP blocking cuts ATE estimation error by 75 percent
By adapting a classic propensity-score stratification to the private setting, the new DPBlocking algorithm outperforms earlier methods…
· “Differentially Private Average Treatment Effect Estimation by Propensity Score Blocking”
One generative model learns the full patient journey
A novel transformer with bidirectional time awareness forecasts events, outcomes, and counterfactual effects from lifelong multimodal…
ERPO boosts code generation at test time without answers
By constructing probe inputs and masking low-consensus updates, zero-shot transfer improves on multiple benchmarks.
Dimension-free complexity bounds for sparse neural nets
Upper and lower bounds match up to log factors, removing the earlier sqrt(n) factor and showing input dimension does not control sample…
· “Nearly Tight Rademacher Bounds for Sparsely Activated Neural Networks”
Bayesian duality follows from a convex dual of Bayes' rule
Solving the dual recovers the likelihood from the posterior, with Amari's coordinate swap as a special case.
One scalar decides when scale-invariant training destabilizes
Exact law: one scalar sets effective step size; accuracy peaks at the instability boundary from MNIST to GPT-2.
· “When Does Scale-Invariant Optimization Become Unstable? An Exact Schedule Law with Weight Decay”
AI agents near experts at finding features
Ten agent configs near expert selectivity but score half as well on steering — causal validation is the missing skill.
· “SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?”
Wasserstein paths isolate why curricula help from when they help
A controlled synthetic study across 12 tasks shows that easy-to-hard ordering, pacing, and smoothness each matter in different contexts.
· “Curriculum Learning as Transport: Understanding Curricula with Wasserstein Geodesics”
Zero-rollout prior halves wasted prompts in RLVR training
Using a cheap offline model to estimate difficulty before training, ThinkPrior cuts silent groups by 55% with no accuracy loss.
· “ThinkPrior: Zero-Rollout Difficulty Priors for Cold-Start Prompt Selection in RLVR”
Two seasons of data rival all seasons for grape cold-hardiness
Multi-task learning across 21 grape cultivars consistently beats single-task models and the current scientific standard.
· “Multi-Task Learning for Sparsely-Labeled Time Series: A Case Study on Cold-Hardiness Modeling”
LLM-made JavaScript games train RL agents at over 1M steps per second
A browser-free backend runs the same game file humans play, beating classic benchmark environments on speed.
Training-Free Task Vectors enable persistent LLM edits without fine-tuning
The method converts activation steering directions into rank-one weight updates, allowing addition, subtraction, and composition of…
Simple CoT beats structured formats for human LLM evaluation
Controlled study shows users prefer planning-based traces but simple chain-of-thought yields better error detection and trust calibration.
· “Do Reasoning Representations Help Humans Evaluate LLM Outputs?”
Trajectories expose hidden reasoning dynamics in LLMs
Novel representation tracks how probability shifts among answers, revealing differences that accuracy and entropy miss.
· “Answer-Distribution Trajectories: A Stochastic-Dynamics View of LLM Reasoning”
Gaia data yield closed-form equations for Milky Way's local potential
Method avoids degeneracy that plagues joint fits, recovers self-gravitating disk shape and first analytical distribution function.
· “Closed-Form of the Local Galactic Potential and Stellar Distribution Function from Gaia DR3”
One inference tweak lifts Sudoku validity to 95%
Continuous diffusion models learn constraints, but standard DDPM sampling locks in early errors.
· “Let It Go or Learn to Self-Correct: Continuous Diffusion for Constrained Discrete Tasks”
S3KG achieves top F1 on 7 of 9 datasets for context understanding
A hybrid structural-semantic KG metric reveals domain complexity as a key bottleneck in LLM QA.
· “Evaluation of Contextual Understanding in Large Language Models”
Conservation identity for LLM reasoning paths ties keyword filter
A three-channel scattering layer guarantees T+R+A=1 to machine precision but shows no accuracy gain over a six-word rule.
Physics-informed model cuts false VT alarm rate by 5 points
Requiring plausible blood-pressure waveforms removes ICU artifacts that mimic ventricular tachycardia.
· “Physics-Informed Deep Learning for False Ventricular Tachycardia Alarm Reduction in the ICU”
Transformers can sample from in-context examples
Proving that frozen attention layers simulate closed-form diffusion and particle-based generation from prompts alone.
· “Transformers as In-Context Samplers: From Closed-Form Diffusion to Estimation-Free Sampling”
Best turn-taking beats GPT-Realtime
A 9B omni interaction agent with Cerebellum-Brain architecture sets new interaction records while handling realime agentic tasks.
Autmated graph features cut engineering time 10× in fraud detection
A distibuted system produces interpretable structural statistics that beat expert crafting and GNNs on Alipay transaction netwks with…
One lattice settles the learnability of Boolean formulas
Where your connectives sit in Post's lattice decides what you can fit, compress, and learn.
· “Fitting and Learning Basis-Restricted Propositional Formulas”
600-step cylinder flow prediction on unstructured meshes
Level-set encoding and divergence regularization stabilize long-horizon predictions across laminar, transitional, and turbulent regimes.
· “ONE CYLinder: A Benchmark for Graph-Based Surrogate Modeling of Unsteady Bluff-Body Flows”
One pass reverses apparent forgetting by 78 points
A weight-preserving BatchNorm update unmasks a measurement artifact that skews unlearning evaluation
· “The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation”
Pinned magnetization makes cold spin glasses tractable
Fixing the spin count removes the low-temperature barrier and cuts sparse Bayesian sampling to k^1.5 measurements.
World model turns Earth emulator into interactive what-if tool
Transition-action pretraining lets the model respond to user edits without manual action labels.
· “Earth System World Model for What-If Simulations: A Case Study for Terrestrial Ecosystems”
Transformer sample bound drops from double to single exponential
Run-length-compressed training strings make the required training data single-exponential in size.
Closed-form estimator replaces clean-anchor assumption
Under a single-common-factor model, anchor–judge error correlation is identified in closed form with a diagnostic battery for violations.
Takes 50x larger steps through stiff contact
Keeps gradients reliable at 0.1 s steps: 8,192 worlds on one GPU, control tasks MJX cannot solve.
· “Ostrich: Taking Large Strides Through Stiff Contact in Differentiable Dynamics”
Weak teacher accelerates strong student beyond itself
OPRD amplifies the aligned component of a student's own reward gradient along a teacher's learned shift, beating the teacher with fewer…
· “Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation”
Splitting attention along time and electrodes lifts EEG accuracy
An axis-factorized transformer gains 3–5 balanced-accuracy points on six EEG tasks, and the same split carries to audio.
· “Adaptive Anisotropic Attention for Axis-Structured Signals”
ZK-Trace: 712/720 collusion mixtures traced
A certificate-gated adjudication rule bounds innocent-naming probability to 10⁻³ per investigation, verified on federated GNSS and…
EDEN scalings match CDEF coefficients per rotation
Pointwise +1 identity holds exactly for each Haar rotation; classical relation only in large-d limit
PAC-Bayesian bounds for LTI systems with sub-Gaussian noise
Theorem 3.1 gives O((ln N)^r/√N) rates for bounded, Lipschitz, and quadratic loss, plus parameter estimation.
Two-parameter filter beats both extremes in ad A/B tests
Dropping only logs where control and treatment disagree keeps the model closer to an unbiased reference.
· “BAFF: Bid-Aware Filter Family for Mitigating Training Data Interference in RTB A/B Tests”
Mixture of graph experts matches task-trained models in most tests
One architecture handles node, link, and graph classification on unseen datasets with few labeled examples.
· “Chimaera: A Mixture-of-Graph-Experts Architecture for Cross-Task and Cross-Dataset Graph Learning”
2.9B-parameter TTS streams and matches ElevenLabs on prosody
A four-stage transformer model built on a semantic-first codec runs on a single RTX 5090 and rivals top commercial voices in audiobook…
· “TontaubeV1: Streaming Text-to-Speech with Hierarchical Codec Modeling and Bounded Context”
Sparser MoE models need lower learning rates and larger batches
Two power laws add the activation ratio as a third dimension, letting hyperparameters transfer across sparsity levels.
HOPE: Open-set node classification succeeds where graphs defy homophily
A plug-in framework uses pseudo-extrapolation along cross-class edges to reject unknowns on heterophilic graphs.
· “HOPE: Heterophily-Aware Open-Set Node Classification with Pseudo-Extrapolation”
Without adversarial training or purification
A method combining oscillatory neurons and predictive self-supervision achieves strong adversarial robustness without ever generating…
Noisy discrete functional regression hits minimax rate
Estimator matches optimal prediction error of continuous setting when grid is fine enough.
· “Optimal estimation for Functional Linear Regression with Noisy Discretized Data”
Hybrid RL lets robots build spanning arches adaptively
A novel extension of soft actor-critic handles mixed actions and graph states, achieving real-robot closed-loop assembly.
· “Learning to build covering structures with continuous adjustments”
MoE scaling beats reasoning tokens for multimodal embedding
MOEMB achieves 70.4% on MMEB-V2 with 3.1B parameters, surpassing TTE models using over 12B parameters and less compute.
· “MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models”
Overlapping chart elements expose a common AI blind spot
Vision-language models identify chart roles and links but cannot resolve which component is in front, a new benchmark shows.
· “Charts Are Beyond Pixels: Probing for Layer-Wise Chart Understanding and Editing”
Difficulty-adaptive tree search expands reasoning pass@k by up to 3 points
DATPO combines adaptive rollout and sentence-level diversity to improve test-time scaling in mathematical reasoning.
· “Difficulty-Adaptive Tree-Structured Policy Optimization for Expanding Reasoning Coverage in RLVR”
SUN: multiplicative goal score beats additive baselines in exploration
A product of successor value and pseudocount jointly scores novelty and reachability, outperforming tuned additive methods with fewer…
Suan matches DPO safety without DPO's over-refusal
A margin-only preference loss keeps helpfulness and compliance at SFT level while matching DPO's harmlessness.
· “Suan: Rectifying Direct Preference Safety Alignment in Large Language Models”
Synthetic CT generator matches real anatomy within 1.44 mm
A conditional VAE learns patient-specific deformation modes and generates realistic synthetic repeat CTs for adaptive proton therapy.
· “SynthRCT: Scalable Conditional Deformation Synthesis for Synthetic Repeat CT Generation”
Separate graph for concurrent cases boosts next-activity prediction
Four encoding strategies tested; best model uses a distinct context graph with full prefix histories, outperforming prior graph-based…
· “Leveraging contextual events on structure-aware next activity prediction”
Neural GP learns its own tuning
Auto-hetSNGP achieves 96.8% accuracy on dynamic aperture while converging 3× faster than standard DDU and producing uncertainty maps…
· “Flexible Spectral-Normalized Neural Gaussian Process for Dynamic Aperture Prediction”
Four quick tests predict how a model will respond to training
Standardized pulses of micro-training reveal a checkpoint's hidden response state, cutting prediction error by up to 78% over benchmark…
Outcome-indexed attention stabilizes multi-outcome learning
Replacing one salience per feature with a per-outcome matrix prevents both bound collapse and sign cancellation.
· “Why shared attention vectors fail: a case for outcome-indexed tuning”
Neural level sets cut 3-D scattering error by 89%
Multi-level-set physics-driven solver yields clear boundaries and uniform material regions without manual tuning.
· “Multi-Level-Set-Based Physics-Driven Neural Network to Solve 3-D Inverse Scattering Problems”
GAN-based semantic comms beats non-IID and fading by 58%
FedGenSC uses local discriminators and a prototype bank to improve text reconstruction across all SNRs.
· “FedGenSC: Federated Generative Semantic Communication with Channel-Aware Adaptation”
Shortcut and leakage undermine automated simulatability
When LLMs evaluate explanations, they often just solve the task instead, and anonymization merely rewards label leakage.
· “Limitations of Automated Simulatability: LLM Simulators Can Bypass Explanations”
Imaging-supervised ECG improves Chagas detection by 5 points
Contrastive pre-training on CMR data transfers structural heart knowledge to ECG, boosting performance even for a disease never seen…
Persistent memory and stochastic returns improve stock signal discovery
AlphaRJM wins 10 of 12 metric comparisons by retaining evaluation history across formula episodes.
· “AlphaRJM: Reward-Jump Memory for Stochastic Return-Guided Alpha Discovery”
Non-adaptive 1-bit mean matches adaptive rate
For all moment regimes, no query interaction needed; full sample-interval tradeoff characterized.
· “Non-Adaptive 1-Bit Mean Estimation: Minimax Rates and the Sample-Interval Tradeoff”
Triple class entanglement trails its strongest pair 97% of the time
Certified topological scans of 111 networks show pairs carry nearly all class-overlap structure
Gradient surgery across variables cuts forecasting error 3.6 percent
A single-backward-pass method reconstructs per-variable gradients and aligns them, improving forecasts on seven datasets.
Spectral balance fixes video glitches without retraining
A signed diagnostic reveals when temporal attention is too sharp or too diffuse; a query-temperature regulator restores coherence at…
· “Temporal State Transport in Video Generation: Diagnosing and Correcting Spectral Imbalance”
This paper shows that secure aggregation in decentralized federated learning does not…
Lattice reduction attacks can reconstruct private model updates from secure neighborhood aggregates in decentralized federated learning…
