New Machine Learning (cs.LG) papers, Wednesday 21 October 2026
249 papers announced on arXiv. Get them ranked for you, free →
OPLoRA: Orthogonal Projection LoRA Prevents Catastrophic Forgetting during Parameter-Efficient Fine-Tuning
AI summaryThe paper introduces OPLoRA, which constrains LoRA updates to the orthogonal complement of dominant pre-trained weight subspaces using double-sided projections. This theoretically grounded approach prevents catastrophic forgetting while maintaining competitive task performance across various benchmarks.
Half-Truths Break Similarity-Based Retrieval
AI summaryThis paper shows that CLIP-style models incorrectly assign higher similarity scores to captions containing plausible but false details. The authors propose Component-Supervised CLIP (CS-CLIP), which fine-tunes models on decomposed caption units and foils to significantly improve compositional understanding.
An AI Security Agent for University ACMIS: Multi-Vector Threat Detection and Automated Response
AI summaryThis paper introduces an AI-based security agent for university management systems that combines anomaly detection and behavioral analytics. It achieves a 0.966 macro-average F1 score in threat detection while maintaining sub-millisecond response latency for critical risks.
H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning
AI summaryH2DF introduces a calibrated target‑band loss for language‑model unlearning, applying gradient ascent only below a data‑dependent lower bound and reversing it above the upper bound. It achieves forget/retain NLLs close to GradDiff with far less variability and minimal replay.
High-dimensional layer-state trajectories reveal the internal dynamics of transformer inference
AI summaryThis paper analyzes transformer layer-state trajectories to show that inference unfolds as an ordered, measurable computation. The results demonstrate that specific hidden-state sequences and local transition actuators can selectively control internal model dynamics and output candidates.
LatticeMind: A Validation-Gated Agentic Framework for Density Functional Theory Workflows
AI summaryLatticeMind is a validation-gated agentic framework that uses workflow contracts and deterministic pre-execution checks to automate multi-stage density functional theory calculations. It successfully reproduces reference benchmarks across different solvers with high first-try reliability and zero execution-time syntax failures.
Steerable Pluralistic Policy Alignment from Anonymized Crowd Feedback
AI summaryP3L learns trajectory embeddings and a shared preference basis via PCA from anonymized crowd feedback, enabling fast, personalized reward learning without annotator IDs. It achieves competitive RL performance and out‑of‑distribution generalization in embodied control tasks while reducing compute.
CaMeWM: A Causal Medical World Model for Generative Tumor Evolution and Treatment Planning
AI summaryCaMeWM predicts counterfactual tumor evolution and survival after TACE by generating intervention-conditioned follow-up worlds and mapping them to restricted mean survival time. It achieves ~0.62 SSIM on CT transition and 0.615 C‑index, estimating a 9.6‑week RMST gain for the frozen policy.
ShlokaDhara: A Multi-Speaker Sanskrit Speech Corpus for Learning Meter-Aware Prosody in Text-to-Recitation
AI summaryShlokaDhara is a 43‑hour, multi‑speaker Sanskrit speech corpus that records verse with accurate meter, style, and melody, enabling text‑to‑recite synthesis. Benchmarks show XTTS‑v2 achieves the best perceptual quality, validated by expert ratings and objective Laghu‑Guru duration metrics.
Expert-Guided Rubrics as Dense Reward Signals for Reinforcement Learning in Non-Verifiable Domains
AI summaryUses expert‑rated rubrics to turn sparse, graded SME feedback into dense rewards for RL agents in non‑verifiable retrieval tasks; the hybrid signal boosts Recall@3 by 6% and MRR@3 by 6.7% over sparse‑reward baselines.
Quadratic Sample Complexity for Strictly Replicable Realizable Learning with Unknown Marginals
AI summaryThis paper establishes quadratic sample complexity bounds for strictly replicable learning under unknown marginal distributions. The main result proves a tight $\Theta(d^2/(\varepsilon\rho^2))$ sample complexity for learning the Boolean class $d$ under strict replicability requirements.
Beyond Policy Discrepancy: Capability-Aware Multi-Teacher On-Policy Distillation
AI summaryCA‑MOPD merges specialist RL models into a strong student and then adjusts teacher supervision based on verifier‑estimated capability gaps, reducing guidance when the student matches or exceeds a teacher and increasing it when regressing. It outperforms baseline distillation by 0.74‑0.86 points across six benchmarks.
Dynamic System Inspired Continuous Large Language Model for AI4Science Tasks
AI summaryContinuum-LM reformulates pretrained residual transformations in language models as continuous dynamical systems governed by ordinary differential equations. By enabling variable computational depth through parameter reuse, it improves exact-match accuracy by up to 10.96 percentage points across eight scientific tasks.
FoundPDE: A generative foundation model for data-driven symbolic discovery of partial differential equations
AI summaryFoundPDE is a pretrained transformer-based foundation model that directly maps spatio-temporal trajectories to interpretable symbolic partial differential equations. It achieves over 90% type recognition across diverse PDE families and robustly generalizes to noisy data and unseen compositions via few-shot fine-tuning.
When Weak Submodularity Meets Matroids: How Does Standard Greedy Behave?
AI summaryThis paper analyzes the performance of the standard greedy algorithm for weakly submodular maximization under general matroid constraints. The authors resolve a long-standing open question by establishing matching lower and upper bounds that prove a tight rate.
SZED at ArGuard Shared Tasks: Multi-Task Label-Attention and Cross-Modal Fusion for Arabic Hateful Meme Detection
AI summaryFusion of MARBERTv2 and CLIP ViT-L/14 embeddings yields 0.7597 macro‑F1 on Arabic hateful meme binary classification, while a multi‑task label‑attention network with cascade calibration achieves 0.3142 macro‑F1 on fine‑grained multi‑label tagging.
Transformers Learn Two Competing Sorting Algorithms, and Only One Length-Generalizes
AI summaryTransformers trained on sorting converge to two distinct solutions: one uses only the first attention layer and fails to generalize to longer sequences; the other employs both layers, generalizing perfectly. Weight initialization can steer the model toward either class.
Latent Projection for Latent Planning
AI summaryLatent Projection for Latent Planning proposes SIGReg, a normal‑matrix constraint that aligns the world model’s latent dynamics to a stable eigenbasis, allowing the planner to restrict its ℓ₂ cost to the low‑uncertainty subspace and boost success rates by up to 54.7 % in 2‑D navigation.
Dynamic Dual-Feedback Conformal Inference for Time Series Forecasting
AI summaryThis paper introduces Dynamic Dual-Feedback Conformal Inference, a method for time series uncertainty quantification that combines actual miscoverage feedback with estimated conformal thresholds. The approach achieves more stable and narrower prediction intervals while maintaining target coverage.
CoASFL: Coupled Consistency Optimization for Asynchronous Split Federated Learning
AI summaryCoASFL reduces training lag in asynchronous split federated learning by merging client activations and adjusting aggregation weights to handle horizontal and vertical inconsistency, boosting accuracy by up to 3.26 % and speeding up training 1.6–3.7× on heterogeneous edge devices.
CorteXforge: Scenario-Based RF Dataset Generation on Real Testbeds
AI summaryCorteXforge automatically turns high‑level RF scenarios into synchronized experiments on SDR testbeds, generating realistic, reproducible SIGMF datasets that include hardware impairments and propagation effects. On an OTA dataset it achieves strong modulation classification, confirming synthetic pretraining benefits for domain adaptation.
ORBIT: Autoregressive Time Series Generation via Next Frequency-Band Prediction
AI summaryORBIT learns hierarchical frequency‑band representations and generates time series autoregressively from low to high frequencies, enabling better disentanglement of temporal patterns. It achieves state‑of‑the‑art generation quality and faster inference on real‑world benchmarks.
HierDTA: Hierarchical Structure-Aware Representation Learning for Drug--Target Affinity Prediction
AI summaryHierDTA models drugs as atom‑motif graphs with a hierarchical augmentation that propagates motif semantics to atoms, and converts protein kinase domains into residue interaction graphs; cross‑attention fuses these with fingerprints and ESM‑2 embeddings, achieving state‑of‑the‑art DTA performance (r² = 0.791 on KIBA) and strong cold‑start generalization.
Robust Latent Variable Graphical Models: Exact Recovery and Fundamental Limits
AI summaryConvex estimator recovers a Gaussian precision matrix decomposed into a sparse minus low‑rank part using measure‑transformed covariance and a nuclear‑norm penalty; it achieves exact edge, sign, rank, and inertia recovery with optimal sample complexity and robustness to Huber outliers.
Exploration via Activation-Space Steering for Reinforcement Learning with Verifiable Rewards
AI summaryExploration via Activation-Space Steering (ASE) diversifies reinforcement‑learning rollouts by clustering activation differences between correct and incorrect responses, then steering future rollouts with latent vectors. ASE boosts Qwen3 model performance on math benchmarks by 4.87 and 1.87 percentage points over GRPO.
MultiCard: Multiple Attribute Cardinality Estimation via Deep Learning
AI summaryMultiCard is a deep learning estimator that unifies string LIKE predicates and numerical filters for database query optimization. It reduces PostgreSQL's mean Q-error by up to 16 times on mixed workloads while keeping inference latency below 5 ms.
Reliable Estimation of Differential Privacy under Selection Bias
AI summaryThis paper identifies and corrects a selection bias in membership inference attacks used for differential privacy audits. The proposed statistical corrections yield valid privacy lower bounds that respect theoretical upper bounds at the expected confidence levels.
Gaussian Approximation of Besov Functions for Logistic Classification on Manifolds
AI summaryGaussian kernel logistic classifiers on a compact manifold achieve near‑optimal excess risk without assuming a bounded Bayes function. The authors construct bias‑canceling Gaussian approximants that respect intrinsic dimension, yielding logistic excess‑risk rates matching minimax bounds for bounded densities.
EGiME: Entropy-Guided Model Evaluation in Few-Shot Learning
AI summaryEGiME evaluates few‑shot models by clustering features and computing inter‑cluster and intra‑cluster entropy, without training a classifier. It achieves a 0.966 correlation with full‑pool accuracy and yields far more stable rankings than cross‑validated baselines.
Hilbert-Scale Stability of Posterior Diffusion Priors for Ill-Posed Signal Recovery
AI summaryThis paper establishes a Hilbert-scale stability theory for posterior diffusion priors in ill-posed inverse problems. The main result provides non-asymptotic recovery bounds that define an explicit regularity threshold where score errors remain controlled under smoothing forward operators.
ReuF2F: Benchmarking Coding Agents on Formalized Undergraduate Research Problems
AI summaryReuF2F presents a benchmark of 244 formalized open problems from REU projects, evaluated in Lean 4 with coding agents. Even the best model solves only about 40 % of the tasks, highlighting significant gaps in current formal problem‑solving abilities.
Discrete Neural Interlingua for Ontology-Agnostic EHR Modeling
AI summaryThe paper introduces MEDIATOR, a discrete ontology-agnostic tokenizer that maps heterogeneous clinical concepts into a shared vector-quantized codebook using text descriptions and graph supervision. When used in EHR Transformers, MEDIATOR improves zero-shot transfer performance across health systems by 8.6 AUPRC points.
Graph Neural Networks for Astrophysics and Cosmology
AI summaryThe paper reviews how Graph Neural Networks naturally model relational data in astrophysics and cosmology, outperforming other DL methods by respecting geometry, symmetries, and prior knowledge. It surveys diverse applications, notes a data‑usage gap, and offers best‑practice guidelines.
LEAF: LUT-aware Efficient Attention on FPGAs
AI summaryLEAF introduces a LUT-aware training method for FPGA-based transformers that tabulates constant matrix-vector multiplications while keeping activations in arithmetic form. This approach achieves sub-microsecond inference, improving jet tagging accuracy by 1.8 percentage points while using 2.9 times fewer LUTs than previous methods.
ClockMark: Adaptive and Host-Dependent Watermarking for Large Language Models
AI summaryClockMark embeds watermarks in large language model outputs by adapting to the evolving token distribution and the current generation state, ensuring distribution preservation. It achieves up to 16.08 pp higher bit accuracy than prior baselines while maintaining generation quality.
LOB-Impact Benchmark: Emergent Market Impact in Generative Limit Order Book Models
AI summaryNeural generative LOB models trained only for next-token prediction naturally develop market impact; injecting metaorders into five autoregressive models shows impact exponents 0.63–1.21 and no decay, providing an open-source benchmark for evaluating realistic impact in simulated markets.
This is SParTA: A modular, efficient and scalable transformer architecture for jet-tagging
AI summarySParTA is a transformer that splits jet constituents into a hard core of the N hardest particles for full quadratic attention and a soft tail handled by linear cross‑attention, reducing parameters to 300K while matching ParT’s performance and improving background rejection.
RGH-GNN : Risk-Gated Heterogeneous Graph Neural Network for Financial Fraud Detection
AI summaryRGH‑GNN adds risk‑gated aggregation to a heterogeneous graph, weighting transaction edges by LightGBM risk scores so high‑risk nodes influence message passing more. On IEEE‑CIS data it beats baselines, achieving an AUC‑ROC of 0.914.
Reconstruction-Optimized Expert Pruning for Mixture-of-Experts Language Models
AI summaryThe paper introduces OptEP, a continuous relaxation method for post-training expert pruning in Mixture-of-Experts models that optimizes expert subset selection via output reconstruction. OptEP significantly outperforms existing baselines, yielding absolute accuracy gains of up to 16.2% under aggressive pruning across multiple large language models.
Glass-Box Deep Reinforcement Learning for High-Speed Flow Control
AI summaryThis perspective examines the challenges and opportunities of applying deep reinforcement learning to high-speed flow control. It concludes that success requires bridging data-driven optimization with physics-aware learning to ensure safety and reliability in extreme aerodynamic environments.
Ammonia-Hydrogen Phase Equilibria in Planetary Envelopes from First-Principles-Based Machine-Learning Molecular Dynamics
AI summaryUsing machine-learning molecular dynamics, researchers simulated ammonia-hydrogen phase equilibria under extreme planetary conditions. They found that ammonia likely rains out in Uranus and Neptune, but remains fully miscible in warm sub-Neptunes.
Ontic state recovery through prediction
AI summaryThis paper formalizes a system's state as a complete response profile over queries and analyzes how predictive models can recover these underlying states. The authors show that the fidelity of state recovery depends on query ambiguity, state-space geometry, and the choice of prediction targets.
Representing short-term movement to reveal animal identity and learning progress
AI summaryThis paper introduces a self-supervised framework that learns compact representations from short-term animal movement trajectories using temporal proximity. The resulting embeddings accurately predict individual identity and learning progress, revealing that egocentric pose and allocentric motion encode distinct behavioral signals.
Towards Interaction State Recognition in Sensorized Gecko-Inspired Adhesive Interfaces for Orbital Servicing
AI summaryThe paper develops a real‑time state‑recognition system for sensorized gecko‑inspired adhesives used in orbital servicing. Using embedded capacitive sensors and a Random‑Forest model combined with a rule‑based state machine, it achieves 93.1% accuracy with 53 ms inference.
ICLHerd: Herding-Based Data Selection for LLMs with In-Context Learning as a Proxy
AI summaryICLHerd balances mode-seeking and mode-averaging for LLM data selection using a von Mises-Fisher kernel and in-context learning as a proxy. Results show it outperforms existing data selection algorithms.
Cross-disorder Learning of Imaging-derived Joint Mophometric Markers in Alzheimer's and Parkinson's Diseases
AI summaryThe paper introduces a spatially adaptive transfer learning model that learns joint morphometric biomarkers from ADNI and PPMI neuroimaging data. The model achieves higher diagnostic accuracy than disease‑specific models, showing cross‑disorder transfer improves performance on limited samples.
A Hybrid TFT-NBEATS Framework for Robust Multi-Horizon Photovoltaic Power Forecasting
AI summaryThis paper introduces a hybrid deep learning framework combining Temporal Fusion Transformers and N-BEATS for multi-horizon photovoltaic power forecasting. The proposed model achieves a 35 to 40 percent reduction in RMSE compared to baseline methods while maintaining high interpretability across multiple time horizons.
Understanding the Benefits of Scaffolds through Computational Separations
Inference-time harnesses, i.e., external procedures that call a model many times and keep state across calls, can substantially improve model capabilities. In this paper, we investigate the benefits of using such harnesses and the extent to which these benefits can be…
Meta-iLaD: Identifiable Latent Dynamics via Meta-Learning of Dynamics Environments
AI summaryThis paper introduces Meta-iLaD, a meta-learning framework that jointly achieves identifiability for latent states, environment variables, and dynamics functions without requiring predefined environment labels. Experiments show it successfully generalizes to unseen environments and improves trajectory forecasting.
On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models
AI summaryLanguage models err more on composite prompts than on isolated tasks because they implicitly budget tokens, using fewer per task. Grouping tasks by fragility lets the budget be reallocated, cutting token usage by 43 % while keeping accuracy.
Low Residuals, Wrong Value: Learning the Viscosity Solution of Hamilton-Jacobi-Isaacs Equations with PINNs
AI summaryThe paper investigates why physics-informed neural networks minimize residuals to wrong solutions for Hamilton-Jacobi-Isaacs equations and shows that incorporating distance features or structural diagnostics is essential to recover the correct viscosity solution.
Diffusion Score Estimation with Shallow Neural Networks: Regularity, Approximation, and Finite-Sample Rates
AI summaryThe paper shows that shallow ReLU networks trained by denoising score matching can estimate Ornstein–Uhlenbeck diffusion scores with high-probability risk $\widetilde O(n^{-(d+3)/(2d+3)})$, yielding a total-variation error $\widetilde O(n^{-(d+3)/(4d+6)})$ for reverse-process sampling.
You Are the Trigger: Speaker-Activated Backdoors
AI summarySpeaker-activated backdoors use a target speaker’s natural voice as a trigger introduced during fine‑tuning, without inference‑time access. The two‑phase method attains >90% success on Whisper‑Medium and Parakeet with <2% false positives, enabling content‑dependent attacks on downstream LLM agents.
A Theoretical Analysis of Environment-Aware Out-of-Distribution Detection for Vision-Language Models
AI summaryThe paper develops a theoretical framework for environment‑aware OOD detection in continual vision‑language models, introducing a cross‑environment MCM score that uses pairwise cosine expectations of image‑text features. It proves the score’s false‑positive bound and characterizes when detection succeeds or fails across environments.
Deep catastrophic-risk-aware reinforcement learning
AI summaryExtends extreme value theory to deep RL by deriving a differentiable CVaR objective via peaks‑over‑threshold estimation, enabling backpropagation of tail risk in neural policies. In deep‑hedging tasks, EVT‑trained agents match empirical‑CVaR performance normally but incur substantially lower losses in rare extreme events.
Neighborhood-Aware Unlearning for Generative Recommendation
AI summaryThis paper identifies that unlearning in generative recommenders causes unintended utility degradation in the semantic neighbors of forgotten items. To fix this, the authors propose Neighborhood-Aware Unlearning, which explicitly controls probability redistribution and successfully restores local recommendation performance.
TxnQA: Benchmarking Analytical Reasoning and Evidence Grounding over Transaction Event Sequences
Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark…
Guided Diffusion for Trustworthy Synthetic Text Generation
AI summaryThis paper introduces Guided Diffusion for private synthetic text generation, using novel sampling and re-masking procedures in diffusion language models. The method achieves superior distributional fidelity and downstream utility compared to existing differentially private baselines while maintaining strong privacy guarantees.
Solving the Ho--Nguyen Conjecture for Gaussian Mixtures: Applications to Mixtures of Experts
AI summaryShows that for any over‑fitted Gaussian mixture of size ≥ 2 the cancellation polynomial has exact order 2M, proving Ho–Nguyen’s conjecture. This yields the sharp learning exponent and local minimax rates, and clarifies symmetry‑based rate limits in mixtures of experts.
Auditing Safety Preference Supervision with Recorded Derivations
AI summaryThe paper introduces Symbolic Preference Labeling, which records attribute classifications, rule-based verdicts, and a ranking to recompute preference labels without re‑running a language model, enabling auditability. On 200 prompts it achieves 96% structured output, with policy models refusing 40% benign and 85% toxic prompts.
What Would Removing This Corruption Change? Failure Attribution for Vision-Language-Action Models
AI summaryThis paper introduces a factorized risk model to attribute vision-language-action model failures to specific visual corruptions. The approach estimates the benefits of corruption removal through latent intervention without requiring paired clean data, achieving roughly three times lower attribution error than frontier vision-language models.
A latent control model for realistic rodent motion
AI summarySCAMPER trains a virtual rat using two variational priors—one for motion control and one for exploratory behavior—producing realistic movements during both goal-directed tasks and random exploration, even with sparse rewards.
VPFlow: Value-Predictive Planning with Imperfect World Models for Offline Goal Reaching
AI summaryThe paper introduces VPFlow, an offline goal-conditioned reinforcement learning method that trains world models to predict values rather than future states. This value-predictive planning approach outperforms existing latent planners and model-free baselines across diverse benchmarks.
TEX: Test-Time Scaling Testing Agents via Execution-based Cross-Validation
AI summaryTEX parallelizes agents to generate test cases and patches, cross‑executing each pair to build a $K\times K$ feedback matrix that guides refinement without trace sharing. It attains 86.4% on SWT‑Bench Verified, a +9.0% boost over single‑agent baselines.
Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource Languages
AI summaryThey test ten large language models in 79 languages for translation‑bypass attacks, finding that safe English prompts still lead to unsafe outputs 22–44% of the time, especially in certain language families, and that model rankings are unstable across conditions.
What Not to Erase: Saliency-Constrained Augmentation for Low-Resource Handwritten Character Recognition
Optical Character Recognition (OCR) and handwritten character recognition (HCR) in low-resource historical material collections often rely on data augmentation, yet augmentation strategies are typically evaluated almost exclusively through predictive accuracy. This leaves open…
Physics-Inspired Optimizers beyond the Edge of Stability: Period-Two Orbits and Delayed Instability
AI summaryThis paper analyzes physics-inspired optimizers beyond the edge of stability, showing how kinetic maps create stable period-two orbits and exhibit delayed instability governed by slowly swept bifurcation theory.
VampPrior Unleashed
Deep latent variable models (DLVMs) underpin many modern analysis pipelines, for example in single-cell genomics, yet their prior distributions have received markedly little attention. The VampPrior approximates the optimal prior (the aggregate posterior) and often attains high…
EMBER: Predicting TCR--pMHC Binding Without a Docked Complex
AI summaryEMBER predicts TCR-pMHC binding by independently encoding monomer structures into SE(3)-invariant graphs without requiring a docked complex. This approach avoids placement error and outperforms complex-based methods across all evaluation metrics and folding tools.
Automated Brick Schema Knowledge Graph Completion Across Buildings
AI summaryWe treat Brick model completion as a knowledge‑graph completion task and benchmark three approaches. CompGCN achieves the best macro‑F1 and AUPRC on held‑out buildings, but no method consistently generalizes across all datasets or relations.
Robust Lightweight IoT Device Fingerprinting Under Network Traffic Anomalies
AI summaryThis paper proposes a lightweight IoT device fingerprinting method using packet headers and locality-sensitive hashing for resource-constrained environments. The approach achieves up to 99 percent accuracy and maintains robustness against various network traffic anomalies using short observation windows.
LiBRA: Logit Fingerprints for Byzantine-Robust Client Selection in Federated Learning
Byzantine-robust aggregation shields federated learning from malicious updates, but the rules that filter out suspect updates do so by comparing clients' parameter vectors. That comparison requires access to every client's update, which secure aggregation exists to withhold,…
Predictive Closure: A Set-Level Characterization of JEPA's Feature Selection
In joint-embedding predictive architectures (JEPA), an encoder and a predictor are trained together to produce a representation that minimizes the error in predicting the representation of the next observation. Because the representations of both the current and the next…
Trail-Weighted Influence Scores for Approximate Outcome Prediction in Bayesian Networks
We study trail-weighted pruning of evidence for Bayesian-network prediction. A traversal enumerates the active trails from each evidence node to the outcome, scores each trail by the product of its edge strengths, and drops evidence whose score falls below a threshold. For…
Shaped SAEs: Expanding the Toolkit for Mechanistic Interpretability
AI summaryThis paper introduces Shaped SAEs and Matched Filter SAEs, which apply linear and learned affine transformations to LLM activations before dictionary learning. This approach improves semantic selectivity and outperforms standard sparse autoencoders on prominent interpretability benchmarks.
DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding
Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same masked block, without…
Sample-efficient learning with Dense Associative Memories
We analytically study the evolution of the energy function of Dense Associative Memories (DenseAMs) as they train on randomly sampled codewords of random expander low-density parity-check codes (a.k.a. solutions of random XORSAT problems). We identify two distinct regimes and a…
Training Controllable World Models via Action Influence Maximization
AI summaryThis paper identifies that action-conditioned video models fail because expert training data lacks action influence, limiting controllability. To fix this, the authors introduce Action Influence Maximization (AIM), an active data collection algorithm that achieves an 85x reduction in expert data needed for controllable world models.
Fixed-Size Active Statistical Inference
Variance reduction techniques improve the efficiency of drawing conclusions from data. Recent proposals for variance-reduction use two complementary techniques to reduce variance: incorporating machine learning predictions into estimators and weighted sampling. A popular…
Communication Complexity of Universal Features
AI summaryThis paper studies the communication complexity of universal data representations for arbitrary function classes. The authors prove that an encoder can compress data into $O(d^\log(1/\varepsilon))$ bits—where $d^$ is the dual VC dimension—achieving worst-case optimality across all representation tasks.
Steering intonational tunes in text-to-speech synthesis
AI summaryThis paper investigates how text-to-speech models encode prosodic intonation and shows that steering the transformer backbone's hidden states with imitation-derived vectors successfully controls sentence-level rising and falling melodies with minimal loss in naturalness.
Learning the Geometry of Materials Space with Trained Neural Tangent Kernels
AI summaryThis paper uses trained neural tangent kernels to analyze the internal geometry of networks predicting superconducting transition temperatures. The resulting kernel successfully reproduces network performance and reveals that the learned materials space is organized into recognizable chemical and superconducting families.
The Router Already Knows
Mixture-of-experts (MoE) models already make a sparse choice for every token: which experts should process it. We ask whether a sparse dictionary should reuse that choice. NR-SAE gives each routed expert a small dictionary and opens only those the frozen router selects. At block…
Baseline and Incremental: User, Item, and Request Debiasing in Instagram Feed Ranking
AI summaryThe paper proposes keeping raw engagement predictions and adding a separate incremental term in the value model for Instagram feed ranking. A four‑arm ablation shows the two‑term design outperforms single‑term versions in most metrics, and the operator is now widely used across Meta apps.
MLIPS: Multi-Layer Indexer Pool Sharing
Sparse attention limits the keys used by each query, but finding those keys can still require a full-prefix scan at every layer. We introduce \textbf{MLIPS (Multi-Layer Indexer Pool Sharing)}, a training-free method that shares candidate search while retaining each consumer's…
Scaling Graph Reasoning in Language Models with Learnable Graph Positional Encodings
Solving reasoning tasks supported on graphs typically requires providing a Language Reasoning Model (LRM) with a serialized description of the graph. This approach results in the input length growing with the number of edges, which limits the size of the problems that can be…
Learning-Theoretic Foundation for General Coded Computing: The Adversarial Setting
Adversarial workers pose a severe challenge in distributed computing systems by returning corrupted outputs that can substantially distort the recovered computations. Coded computing addresses this threat by introducing redundancy across worker assignments. Existing exact and…
Using Set Learning to Improve Set-Valued Classifiers on Long-tailed Data
Set-valued classification (SVC) offers a practical approach to communicating uncertainty while maintaining accuracy by predicting variable-sized sets per input rather than fixed-sized sets (e.g. top-1 or top-5). However, set-valued classifiers (SVCs) exhibit significantly…
LoRASpace: A Pool-Wide Shared Substrate for Static and Dynamic Multi-LoRA Composition
Low-rank adaptation (LoRA) attaches small low-rank matrices to a frozen backbone, and public repositories host many LoRAs per backbone, each for a different concept or task. Combining several LoRAs in one generation forces a trade-off. Merging their weights into one update needs…
Decision Sufficiency of Learned Representation Interfaces for Safe Generative Control
A controller can guide a language model toward safe and useful responses by choosing when to assist, ask for clarification, or refuse. When the controller uses selected internal features, poor performance alone cannot distinguish missing information from limitations of its…
Beyond Scalar Leaderboards: Adaptive Sampling for Pareto Frontier Identification in Multi-Objective LLM Arenas
The evaluation of large language models (LLMs) is inherently multi-objective, as different models may offer different tradeoffs across performance dimensions such as math, code, and writing. Because such evaluations are often based on costly and time-consuming human feedback, it…
Projected Distillation: On Learning to Solve Hard Problems from Close, Correct Targets
On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilities. On hard problems, however, on-policy RL provides no reward-based learning signal when all attempts fail. On-policy…
Training-Free Continual Learning for Multimodal Forecasting via Prompt Evolution
Adapting a forecasting system to a changing world classically means retraining. We ask whether the task of continually learning regimes for time series forecasting can be represented by a piece of text: a guidance block, evolved by outcome-driven reflection and used in a frozen…
Beyond the Prompt: Recovering Hidden Concepts in Text-to-Image Models
AI summaryThis paper introduces EmbGA, an embedding-space genetic algorithm that efficiently recovers hidden visual concepts unmentioned in the prompts of text-to-image models. EmbGA outperforms existing optimizers by achieving higher concept alignment and faster execution speeds.
When Parameter Gradient Flows Close in Representation Space
The represented gradient flow of a parameterized model depends on a mobility that can retain hidden parameter information. We study when one entire represented mobility map works for every autonomous loss from one fixed initialization. For two-factor matrix representations wider…
Orthogonal Boundary Unlearning for Dual-Use Concepts in Large Language Models
Mitigating catastrophic risks from large language models requires excising hazardous operational capabilities, such as cyberwarfare exploitation and pathogen enhancement, without degrading benign scientific utility. In these dual-use domains, legitimate defensive science and…
MiCLA: Mixture-to-Constituent Latent Alignment for Polyphonic Audio Representation Learning
AI summaryMiCLA is a self-supervised pre-training framework for polyphonic audio that aligns mixture representations with their constituent clips using bidirectional latent alignment. It achieves state-of-the-art performance on AudioSet fine-tuning and probing evaluations, effectively handling concurrent sound events.
Self-Evolving Agents through Privileged Retrospective Distillation
Self-evolving agents aim to improve from their own experience rather than remain fixed after deployment, but naturally collected trajectories often lack ground-truth solutions or verifier-derived rewards. We introduce Retrospective Self-Distillation (RSD), a multi-round…
Soliton Profiles in 1D: Enhanced Neural-Network Solvers and Constrained Deep Ritz Methods
AI summaryThis paper develops and compares enhanced neural-network solvers, including novel constrained Deep Ritz methods, for finding one-dimensional soliton profiles. The Weinstein Deep Ritz formulation achieves the best error-runtime balance among neural approaches, offering a promising mesh-free foundation for higher-dimensional problems.
What Moved the Alarm Score? Causal Attribution with Exact Computation and Scarce-Incident Estimation
Alarms in industrial systems and scientific instruments often demand immediate diagnosis despite scarce incident data. To guide intervention, an explanation must trace how changes in the data-generating process propagate through the causal structure to alter the alarm score,…
BEACON: Bridging Multimodal Semantics and Asynchronous Propagation for Cloud System Anomaly Detection
AI summaryBEACON fuses multimodal cloud observability data using a Dynamic Interpretable Multimodal Aggregator that aligns metrics, logs, and system graphs in a shared prototype space. Its Asynchronous Spatio‑Temporal Propagator captures multi‑scale temporal attention and topology‑aware spatial relations, achieving state‑of‑the‑art anomaly detection on three real datasets.
Timing-Aware and Interpretable Self-Attention for Event-Based Wearable Task Transition Detection
Task transition detection aims to identify when a person shifts from one task to another, providing a foundation for continuous health monitoring and timely context-aware interventions. While statistical and deep feature representations have demonstrated strong classification…
Lower Bounds for Smooth Nonconvex Zeroth-Order Optimization: The Price of Determinism under Low Curvature Trace
Low curvature trace, an effective rank condition, is a common explanation for why randomized zeroth-order methods scale to high-dimensional tasks such as language model fine-tuning. Can deterministic methods achieve the same savings? We show that they cannot. We study finding…
Can Native Safety Travel? Reusing Safety Pretraining Across Models via On-Policy Distillation
Safety pretraining has emerged as a promising paradigm for embedding safety deeply into a model during pretraining, yielding robustness to subsequent fine-tuning that conventional safety post-training often lacks. We refer to this safety acquired during model pretraining as…
AnchorRM: Parallel Intra-Batch State Reconstruction with Anchor-Based Residual Memory for Temporal Graph Neural Networks
Memory-based temporal graph neural networks (TGN) rely on recurrent node memories to capture evolving interaction histories. However, temporal mini-batching introduces a trade-off between computational efficiency and temporal fidelity: when a node appears multiple times within…
HypoNet Nankai PS: rapid hypocenter determination for the Nankai Trough using physics-informed neural networks with three-dimensional P- and S-wave velocity models
AI summaryHypoNet Nankai PS adds S‑wave travel‑time inference to the existing P‑wave physics‑informed neural network, improving RMS error by revising collocation points, architecture, and optimizer. Applied to 223 DONET events, it reproduces JMA hypocenters near stations and finds shallower, oceanic‑crust locations farther out.
Learned Readout Tokens for Extracting Large Language Model Confidence
AI summaryLearned readout tokens replace existing-token probes for LLM confidence estimation, yielding higher AUROC across datasets and models. A multi‑token extension further improves performance, and the method can be applied to any probe, boosting accuracy both in‑domain and out‑of‑domain.
TRCM: Topology-Aware Route-Consistency Modeling for Robust Neural Vehicle Routing
Neural combinatorial optimization has enabled fast constructive solvers for the Capacitated Vehicle Routing Problem (CVRP), yet most autoregressive models still construct solutions through sequential node selection and represent route organization only implicitly. We investigate…
Two Is Enough: Binary Classification Priors Suffice for Multi-Class In-Context Tabular Prediction
Tabular foundation models (TFMs) pretrain a transformer on large-scale synthetic priors so that a single forward pass yields zero-shot tabular prediction. TabICLv2 samples the class count of its classification prior during pretraining, and the prevailing configuration caps it at…
Bayesian Neural Networks via Dirichlet Process Priors on the Data-Generating Distribution for Randomized Value Functions in Deep Reinforcement Learning
Bayesian neural networks (BNNs) offer a principled framework for quantifying epistemic uncertainty. However, standard weight-space priors yield complex, high-dimensional neural network posteriors that are computationally prohibitive to sample from, and popular approximate…
When LLMs Manipulate: Measuring Behavioral, Social, and Temporal Effects
Large language models (LLMs) can influence user trust and decision-making through subtle linguistic cues, yet the behavioral effects of manipulative outputs remain poorly quantified. We introduce a behavioral computing framework for measuring \textit{behavioral vulnerability},…
From POMDP Theory to Deep RL with Particle Filters
Model-based reinforcement learning in partially observable environments requires inferring hidden states, learning latent dynamics, and planning under the resulting belief. Practical algorithms often solve one aspect of the above pipeline while relying on approximations for the…
Factual Access States in Diffusion Language Models: From Mechanism to Knowledge Editing and Retrieval
Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR)…
Bridging Digital Deception and Physical Fidelity: A Spatio-Temporal Attack on Cognitive Radars
As cognitive radars evolve towards the end-to-end processing of one-dimensional complex baseband sequences, deep networks face severe adversarial threats. However, existing attacks encounter dual limitations in practical electronic warfare: low cross-architecture transferability…
Sensitivity-Orthogonal Residual Learning for Adaptive Sideslip and Cornering Stiffness Estimation
Online adaptation of the cornering stiffness keeps a model-based sideslip estimator valid as tires change, but competes with learned model correction for the same filter innovation. A corrector with enough parameters can absorb the innovation pattern that a stiffness change…
Ask to Learn: Query-Based Collaborative Reinforcement Learning
Offline reinforcement learning is fundamentally limited by the coverage of the available data, while collecting additional trajectories through online exploration may be costly, unsafe, or infeasible. We study whether this limitation can be overcome without directly interacting…
Auditing the Product, Not the Recursion: The Evaluation Evidence Behind Recursive Self-Improvement Claims
AI summaryThe paper audits 114 studies claiming recursive self‑improvement, finding that while 90% report product‑level evaluation, only 22% assess repeatability with full loops, and only 4 compare identical procedures under matched budgets. It proposes a claim‑indexed reporting standard.
Learning Inter-Graph Interactions Between Heterogeneous Substructures of Chemical Systems
Complex chemical systems containing heterogeneous substructures are common in real-world applications. Various physical phenomena of the complex chemical systems are derived from the interactions between the heterogeneous substructures. However, existing graph representation…
A Radiation-Hardened, Fault-Aware Quantized Neural Network That Self-Heals
AI summaryThis paper introduces Projection Healing, a runtime method that repairs quantized neural networks corrupted by radiation-induced bit flips on commodity hardware without storing replicas. The approach reduces error accumulation and extends the operational scrub interval by a median of over twofold compared to classical error correction.
From Outcomes to Steps: Reliable Success-Aware Critic Learning for Long-Horizon Language Agents
Reinforcement learning for long-horizon language agents requires identifying whether intermediate actions make eventual task success more or less likely when supervision is available only at trajectory level. The main difficulty is that a critic with aggregate return-prediction…
Ensemble-Aware Structural Encoding (EASE): Conditioning Protein Encoders on Peptide Conformational Ensembles
AI summaryEASE extracts peptide representations by selecting central conformers from conformational ensembles and fusing their embeddings. On six property-prediction tasks, it improves performance by about 3 % MAE over single‑structure baselines, with gains increasing as more conformers are incorporated.
BCT: Bayes-Consistent Tempering for Flow Language Models
AI summaryBayes-Consistent Tempering (BCT) integrates temperature into the flow distribution and denoiser dynamics, aligning training and inference. BCT boosts GSM8K and Sudoku performance across temperatures, outperforming post‑hoc scaling and training‑free variants on hyperspherical and von Mises–Fisher flows.
CUA-Recovery: Benchmarking Error Recovery in Long-Horizon Computer-Use Agents
AI summaryCUA-Recovery benchmarks error detection and recovery for long-horizon desktop‑agent tasks, providing 2,250 verified failure states and 8.6K recovery trajectories. Fine‑tuning Qwen3.5 on this data boosts error awareness from 20 % to 35 % and task success scores markedly.
Inferring continuous cognitive traits from limited neuroimaging data using pairwise convolutional neural networks
AI summaryPairwise convolutional ranking learns from whole‑brain MRI by comparing subjects, yielding accurate predictions of spatial ability in small samples (N = 108) and outperforming linear and pointwise neural baselines. Backward elimination highlights shared and task‑specific cortical regions linked to performance.
Pairwise Learning Increases Data Efficiency in Deep Neural Networks
AI summaryPairwise learning reframes regression as aggregated ranking comparisons, yielding massive implicit data. It improves data efficiency, enabling a lightweight ResNet10 to achieve 3.64‑year MAE on UTKFace and boosts STL10 classification by up to 5.2% with limited labels.
Neural Operator Splitting for Fluid Dynamics in Complex Geometry
AI summaryNeural Operator Splitting decomposes each fluid‑dynamic time step into separate physically interpretable operators learned via a conditional‑linear Fourier neural operator, using intermediate solver states. It reduces rollout error by 68 % in 2D and 40 % in 3D while being ~1000× faster than Monte Carlo solvers.
Hybrid Momentum Methods for Optimizing Deep Learning Models
AI summaryHybrid Momentum (HyM) blends a linearly weighted gradient average with a fast exponential moving average, retaining long‑term history while emphasizing recent changes. HyMUON, which adds Newton–Schulz orthogonalization, consistently outperforms Muon on language modeling and large‑scale GPT‑2 and LLaMA pretraining tasks.
CyberHull: Deterministic Malicious URL Isolation via Convex Hull
AI summaryCyberHull uses convex hulls over structural, graph, and semantic URL features to deterministically isolate malicious URLs. It achieves 92 % detection accuracy with O(n log n) training and lightweight point‑in‑polyhedron classification, outperforming PCA (32 %) and deep‑learning baselines.
Influence Is the Target-Response Jacobian Times Residuals
AI summaryThe paper defines the target‑response Jacobian \(J=[\partial \hat y_i/\partial y_j]\) to measure how predictions change with training targets and shows \(J\) is equivalent to influence up to a residual multiplier. On several datasets \(J\) predicts target perturbations with \(R^2>0.99\), while influence performs poorly.
Understanding Data Augmentation under Label Noise
AI summaryThe paper provides a theory for when data augmentation helps or hurts under label noise, giving conditions for both i.i.d. and domain‑generalization settings. It shows augmentation amplifies uncertainty, reducing stability and generalization, and derives broad bounds to guide DA design.
Multi-Objective Evolutionary Semi-Supervised Learning for Resource-Aware Smart City Classification
AI summaryThe paper introduces MOEA-SSL, a multi-objective evolutionary approach for semi-supervised tabular classification in smart cities. It jointly optimizes predictive performance and feature usage, achieving competitive accuracy with significantly fewer features than existing methods.
Privacy Accounting for Empirical Denoisers
AI summaryThe paper proposes a clipped, privacy‑accountable empirical denoiser that allows exact sensitivity calculation for diffusion sampling, enabling DP guarantees for each denoising step. It shows privacy loss peaks during semantic formation, motivating hybrid samplers that switch to a public denoiser, and reports concrete (ε, δ) bounds on CIFAR‑10 and CelebA‑64.
When Does Experience Memory Help Closed-Loop LLM Driving Agents? A Controlled Evaluation with Resampled Warmup Pools, Token-Matched Fillers, and Audited Baselines
AI summaryThe study evaluates closed‑loop LLM driving agents with controlled warm‑up experience pools and token‑matched fillers on a HighwayEnv testbed, finding that memory gains depend on the specific warm‑up pool, token fillers dominate, representational depth is ineffective, and baseline quality reverses apparent advantages.
Anytime-Valid Runtime Monitoring of Learned Barrier Certificates
AI summaryThe paper presents an anytime‑valid runtime monitor that tests learned barrier certificates using closed‑loop trajectories, forming a sequential hypothesis test and a supermartingale to control false alarms under any stopping time. It demonstrates empirical alarm behavior on valid and invalid certificates.
EEG-Based Classification of Hypnotic Susceptibility Using Two-Dimensional Convolutional Neural Networks
AI summaryEEG signals are classified with 2D CNNs to predict hypnotic susceptibility, achieving up to 74% accuracy using only bandpass filtering and z‑score normalization. Anterior frontal electrodes provide the most generalizable markers across subjects.
Constant Step Sizes for Muon-Type Spectral Optimizers on Strongly Convex Quadratics
AI summaryThis paper analyzes spectral optimizers like Muon with constant step sizes on strongly convex quadratics. It shows that convergence depends entirely on the update's gain and length, categorizing practical variants into three distinct regimes ranging from standard preconditioned convergence to non-convergence.
Detection of Memorization in Diffusion Models Through Thermodynamic Singularity
AI summaryThe paper analytically characterizes reverse processes that memorize training samples in diffusion models and proposes a criterion detecting the resulting thermodynamic singularity. The criterion reliably separates properly learned models from memorizing ones across various neural network designs.
Give the Model a Finger: Developmentally Aligned Evaluation of Counting in Vision-Language Models
Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic actions: pointing at each item while reciting number words, moving counted…
AC/DC on a Budget - Alternating Sparse Phases
AI summaryARC replaces expensive dense phases with a lightweight sparse rewiring step, zero‑initializes masked weights, and shortens dense periods, achieving AC/DC‑level accuracy on CIFAR‑100, ImageNet, and GPT‑2 while cutting FLOPs and improving perplexity and zero‑shot scores.
Foundation-Style Pretraining for Scalable and Transferable Distributed PV Forecasting
AI summaryA foundation-style pretraining framework learns a shared PV forecasting model from heterogeneous multi‑site data, using mixed reconstruction objectives. It outperforms station‑specific, shared, and generalist models on 479 stations, enabling accurate zero‑shot forecasting for new sites and easy fine‑tuning.
Engram Memory Network: Brain-Inspired Prototype Explanations
AI summaryThe paper introduces the Engram Memory Network, a brain-inspired prototype model that uses a Hopfield-like memory module for robust classification and explainable retrieval. Experiments show it significantly improves visual similarity, subclass alignment, and information preservation while maintaining high classification accuracy across diverse datasets.
How Much Do $\ell_1$ Attacks Need Concentrated Perturbations?
AI summaryThis paper investigates perturbation concentration in L1 adversarial attacks by enforcing per-coordinate caps during optimization. The results show that limiting large coordinate changes retains 97 percent of attack success while significantly reducing peak perturbation magnitudes.
Foundations of the Scaling Law
AI summaryThe paper derives quantitative scaling laws linking validation loss to model size, data, and compute by geometrically approximating function error, spectrally analyzing feature visibility, and optimizing parameter-data trade‑offs. It shows how these laws predict resource allocation, training‑inference trade‑offs, and domain‑specific gains in vision and molecular tasks.
TerraNostra-Bench: Benchmarking Spatial Downscaling and Temporal Rollout for Global 1-km Land-Use Reconstruction
AI summaryThe paper presents TerraNostra‑Bench, a data‑driven 1 km land‑use reconstruction framework that separates spatial downscaling from temporal rollout and evaluates rare change pixels. Autoregressive models are most informative but lose 5.7‑7.2 accuracy over 13 years; tuning prior‑trust or loss weighting recovers up to 10 points.
Certified Adaptive List Decoding for Error-Correcting Codes
AI summaryThe paper introduces a neural decoder that uses a probability beam search on a diffusion model and two code‑specific linear‑programming bounds to certify maximum‑likelihood optimality per frame. It certifies up to 99.9% of frames and matches or exceeds neural baselines.
SliMOO: Interpretable Multi-Objective Evolutionary Search for LLM Depth Pruning
AI summarySliMOO is a multi‑objective evolutionary framework that prunes LLM depth by first mapping task‑layer importance and then searching Pareto‑optimal masks guided by that map. It outperforms greedy and scalarized baselines, e.g., improving Llama‑3.1‑70B 25% sparsity math performance from 24.33% to 34.40%.
ProVe: Proposer-Verifier Decoding
AI summaryProVe fine‑tunes a large model to propose a block of tokens in one pass and verifies it in a second, swapping the roles of draft and verifier from speculative decoding. ProVe‑lossless achieves 2.5× AR throughput, while a 3B verifier gives 2.7× speed with only 2.6 accuracy loss.
From Generic to Dedicated: A Novel Optimizer for Online Continual Learning
AI summarySPIN is an optimizer that splits plasticity and stability via averaging, acting as an implicit inverse‑Hessian regularizer for online continual learning. It replaces SGD/AdamW and consistently improves performance across rehearsal‑free and memory‑limited OCL methods.
SAPD: Learning from Agent Experience without Environment Interaction
AI summarySAPD trains LLM agents from offline trajectories by decomposing multi‑turn logs into decision contexts and applying relative step‑advantage signals for on‑policy distillation. It avoids environment interaction, reduces catastrophic forgetting compared to SFT, and is 4.7× faster than online OPD.
LaLa-World: Scaling Latent World Models with Natural Language Actions
AI summaryLaLa-World learns a single language‑conditioned latent world model from simulation, robotics, images, and videos, predicting future latent states via a language model. It outperforms domain‑specific LWMs, cuts training cost 6×, speeds planning 13.9×, and enables fast policy inference.
From Tokens to Files: Multi-Granularity Detection and Localization of AI-Involved Code
AI summaryThe paper introduces a token‑level authorship detector that aggregates evidence to classify lines, blocks, and entire files as AI‑involved, achieving 72.4 macro‑F1 at 1% FPR on 11,773 mixed files and 62.5 line‑level F1 on an external benchmark.
Enumeration-Guided Training of Neural Hamilton-Jacobi Safety Certificates
AI summaryEnumerates safety constraints to fine‑tune neural Hamilton‑Jacobi safety certificates, producing larger certified regions and faster verification. In high‑speed racing simulations and real experiments, the refined certificates enforce safety with zero constraint violations versus baseline controllers.
Elicitation or Expansion? Demystifying Gains from Agent Harness Optimization
AI summaryThe paper analyzes harness optimization for language‑model agents, distinguishing whether gains come from eliciting better performance on known tasks or expanding the set of solvable tasks. It finds that while harnesses improve pass@1 on easy tasks, they underperform stronger interventions on hard tasks, especially at larger attempt budgets.
Physical-Variable-Informed Deep Learning for Robust Diagnosis in Industrial Fluid Transport Systems
AI summaryPhyNet integrates physical operating variables as global constraints with local sensor signals using cross‑attention to handle sampling heterogeneity, achieving state‑of‑the‑art fault‑diagnosis accuracy on four industrial fluid transport benchmarks, including real‑world cavitation data.
The Shape of the DPO Hyperparameter Landscape: What to Measure, How It Moves with Scale, and When Not to Transfer It
AI summaryDPO hyperparameters such as learning rate and β depend on model size and training data, and cannot be transferred between setups. Short intermediate runs and dynamic benchmarks accelerate tuning, but validation loss is a weak proxy for downstream performance.
Near-Duplicate Image Selection Transfers Preferences in Vision-Language Models
AI summaryThis paper shows that vision-language models can implicitly transfer their preferences to student models simply by curating choices among near-duplicate images, even when captions and source photographs remain fixed. The results demonstrate that subtle data curation choices shift student model preferences toward targeted concepts without explicit prompting.
Unsupervised Head Network Personalization for Split Computing via Federated Learning
AI summaryUnsupervised federated learning trains a lightweight head network for split computing using small, unlabeled, non‑IID mobile data, with knowledge distillation for privacy. It matches baseline accuracy and cuts mobile communication by 95% with only 3–5 % accuracy loss.
Lost in Recovery: Demystifying Recovery Behavior of Computer-Use Agents via Goal-Oriented Action Replay
AI summaryReTrace is a framework that evaluates model‑based recovery in computer‑use agents by injecting controlled errors and replaying goal‑oriented actions, defining Transition‑Consistent Recovery. It shows ideal recovery unlocks hidden task completion, current agents underutilize this, and accurate recovery improves end‑to‑end performance.
Neural Closed-Loop Identification of Nonlinear Systems with Stability and Generalization Guarantees
AI summaryICI uses a nonlinear dual-Youla parameterization to identify unstable plants from closed‑loop data while guaranteeing stability and offering PAC‑Bayesian generalization bounds, outperforming direct methods under limited data, weak excitation, and feedback‑correlated noise.
Annealed correlation and quenched noise sensitivity in biased deep neural networks
AI summaryAdding independent Gaussian biases of natural scale, the paper proves quenched noise sensitivity still holds for deep fully connected nets with sign activations whenever the perturbed input is asymptotically unlikely to equal the original, while the annealed covariance stays bounded away from zero, driven entirely by output‑mean fluctuations.
On Johnson-Lindenstrauss Transforms for Byzantine-Robust Optimization
AI summaryThe paper uses Johnson–Lindenstrauss random projections to compress gradients in distributed optimization and weight worker updates by their alignment with a trusted reference, achieving Byzantine‑robust convergence guarantees and improved communication efficiency.
When Momentum Helps Polyak--Ruppert Averaging
AI summaryPolyak–Ruppert averaging lets SGD and momentum achieve identical asymptotic covariance; the paper quantifies convergence speed, parallelizability, and error decomposition (center, covariance, shape). Momentum reduces the sample size needed from Θ(√N/κ) to Θ(√N/κ²) and shows Nesterov’s variance is O(√κ/m) while heavy ball incurs Θ(κ³/2/m).
Building Telecom Language Models with RSOD: Domain Data, Specialist Post-Training, Memory-Conscious Distillation, and Networking Evaluation
AI summaryThis paper introduces Routed Specialist On-Policy Distillation (RSOD), a framework that compresses multiple domain-specific telecom teacher models into a single deployable student adapter. The resulting model nearly matches the performance of the multi-teacher ensemble while remaining practical for standard deployment.
RANGER: Retrieval Agent for Navigating Graphs with Semantic Evaluation Rewards
AI summaryRANGER trains a Transformer‑based RL agent to traverse knowledge graphs, using an LLM mentor to provide dense, semantic reward signals for path relevance. It achieves higher QA accuracy than classical search and heuristic‑reward RL baselines on open datasets.
From RL Papers to Framework Code: Inclusion and Review in Seven Open-Source Training Frameworks
AI summaryResearchers tracked 6,939 RL papers linked to seven open‑source frameworks and found that only 37 of 1,598 general RL methods were incorporated. Included papers tend to target large (≥30B) models and modify loss components, while most proposals remain unaddressed publicly.
When Does Training on Synthetic Data Provably Work?
AI summaryThis paper analyzes when classifiers trained on synthetic data successfully transfer to real distributions. The authors prove that synthetic training can fail catastrophically when real data falls outside the synthetic support, unless constrained by Lipschitz conditions.
Poisoning-Based Backdoors Are Not Cryptographic
AI summaryThe authors investigate whether poisoning-based neural network backdoors can achieve cryptographic asymmetry. They prove that any backdoor installed via data poisoning is inherently learnable, meaning defenders can always reconstruct the trigger verifier given observed triggers.
On-Policy Transition Distillation
AI summaryOn-Policy Transition Distillation augments standard final-layer matching with alignment of layer-to-layer log-probability changes on student rollouts. Across multiple backbones and benchmarks, it recovers ~90% of the teacher–student capability gap, outperforming OPD.
Universal Online Convex Optimization With Delayed Feedback
AI summaryUniversal online convex optimization with delayed feedback: a single algorithm adapts to convex, strongly convex, and exp‑concave losses without curvature knowledge. Regret ≈ √(T + d_tot) for convex and σ_max log T for curved losses, improving delay dependence to σ_max.
Auditing Where Concept Bottleneck Accuracy Comes From on CUB
AI summaryThis paper audits Concept Bottleneck Models on the CUB dataset and reveals that their high accuracy often stems from shortcut features and information leakage in internal embeddings rather than genuine visual grounding. The authors show that standard task and concept metrics are insufficient, and CUB's noisy annotations preclude overinterpreting accuracy gaps.
Quantifying the Threat: A Bayesian Probabilistic Framework for Heavy Metal Health Risks among Children in the Densu Basin, Ghana
AI summaryThe authors develop a Bayesian risk‐assessment framework for groundwater in Ghana’s Densu Basin, incorporating parameter uncertainty, co‑exposure, and spatial variability. They find Pb, As, and Cd drive cumulative risk, with ~31 % of sites exceeding a hazard index of 1, revealing highly heterogeneous, uncertainty‑aware risk surfaces.
MuPACE: Learning to Pace Muon for LoRA Fine-Tuning
AI summaryMuPACE learns a recurrent policy to adjust a base optimizer’s step size during LoRA fine‑tuning, improving over fixed schedules on 64 tasks (64–78% better) and transferring one‑shot to deeper models, matching per‑task tuned peaks.
PULSE: Personalized Opportunistic Learning for Wearable Devices
AI summaryPULSE selects useful peer encounters for wearable personalization, using gradient compatibility and past experience to decide when to share models. It matches or improves accuracy on activity, fall, and stress detection while cutting peer exchanges by 80‑94 % and fitting in <600 KiB on-device.
From Complex Histories to Simple Predictors: Predictive Dimension in Dynamic Graph Learning
AI summaryThe paper shows that dynamic graph learning can focus on a low‑dimensional predictive state manifold instead of full histories, defines an intrinsic predictive dimension metric, and builds PreDyG, a tiny model that matches this dimension and outperforms heavy baselines with 0.07 % of their parameters.
Weak-$L^\alpha$ Noise, Strong Guarantees: Gradient Clipping under Affine Heavy Tails
AI summaryGradient clipping achieves high‑probability convergence for convex SGD with affine heavy‑tailed noise under a weak‑\(L^\alpha\) tail. For \(1<\alpha<2\) the rates are \(K^{-(\alpha-1)/\alpha}\); for \(\alpha=2\) a necessary log factor appears, matching lower bounds.
Causal Discovery for Linear Models with Latent Confounding via Multi-domain Identifiable Representation Learning
AI summaryLinADMG learns linear ADMGs from multi‑domain data by identifying noise variables via representation learning, avoiding non‑Gaussian or bow‑free assumptions. With a no‑cancellation condition on latent confounding, it recovers all direct causal edges and the entire ADMG.
Near-Optimal Regret Bounds for UCB-Style Gaussian Process Bandits via Data Thinning
AI summaryThis paper introduces a data thinning technique to Gaussian process upper confidence bound algorithms. It proves that this approach yields tighter confidence intervals and achieves near-optimal simple and cumulative regret bounds for GP bandits.
Fixed Sets and the Hardness of Computing Correlated Equilibria
We show that computing an approximate correlated equilibrium of a polytope game is PPAD-complete, and thus polynomial-time equivalent to finding a Nash equilibrium. This challenges the conventional view in game theory that correlated equilibria are more tractable objects than…
Curriculum Learning for Generative Flexible Docking
AI summaryCurriculum learning progressively introduces large apo‑to‑holo conformational changes during training, improving generalization for flexible docking. The resulting GCLDock model achieves 46.52% top‑1 success (RMSD <2Å) and 40.26% with PoseBusters validity on the PDBBind ESMFold benchmark.
Predicting silent failures of learned inference under restricted access: what measurement physics and prior information explain
AI summaryThis paper investigates predicting silent failures in learned image reconstruction under restricted measurement access. The main result is that while a new score can anticipate these failures, their occurrence is driven primarily by measurement geometry and physics rather than the learned prior.
From Code Generator to Engineering Companion: A Plant-aware Agent for PLC Programming
Programmable Logic Controller (PLC) coding remains a manual, IDE-centric task that demands constant context switching between documentation, project files, and code editors. Recent work has demonstrated that retrieval-augmented large language models can generate vendor-compliant…
Coincidence-Maximizing Customer Clustering Using Smart Meter Data
AI summaryThe paper proposes a fixed-peak refinement algorithm that iteratively reassigns customers to maximize weighted peak‑coincidence in smart‑meter data. On four public datasets it improves coincidence by 1–3.5 percentage points over existing benchmarks.
Foreword: Pre-Generation Prediction and Characterisation of LLM Test Failures using the Jacobian Lens
Testing LLM-based software is expensive: each input needs several sampled responses, each judged by a costly oracle. Further, every prompt or model change calls for re-testing. Pre-generation adequacy measures cut this cost by ranking inputs by failure risk based on the model's…
Offline Makes Online Better: Calibrated Cache Compression for Video Generation
AI summaryThe paper proposes CALM, a method that calibrates token merging ratios across layers and heads for autoregressive video generation. By offline profiling the risk map and using dynamic programming online, CALM reduces memory usage while improving video quality, achieving a 21% LPIPS reduction compared to uniform merging.
Adaptive Pullback: Reweighting Supervision at Penalty Jumps in On-Policy Distillation
AI summaryAdaptive Pullback OPD reweights teacher‑student supervision at sudden penalty jumps by scaling the following token block so its average penalty matches the preceding block. This reduces penalty‑confidence tension, improving results on math reasoning and code generation.
Attractive Faces Skew Text-to-Image Reward Models and Leaderboards
AI summaryThe paper shows that human annotators rank attractive faces higher than unattractive ones, even when image quality is identical, biasing text-to-image preference models and leaderboard rankings. Attractiveness dominates model scores and can override explicit prompts, suggesting the need to separate quality from cognitive bias.
Reframing Object-Centric Reinforcement Learning: From Absolute to Relative Representations
AI summaryThe paper replaces absolute object coordinates with anchor-relative coordinates in object-centric RL and proposes ORBiT, a transformer architecture that uses these relative inputs. ORBiT consistently outperforms baselines, achieving IQM >1.0 on 13 Atari games and mastering Foosball and Tag from scratch.
Parameter-Efficient Adaptation of Time Series Foundation Models for Electrical Load Forecasting
AI summaryThis paper evaluates parameter-efficient adaptation techniques for time series foundation models in electrical load forecasting. The authors demonstrate that low-rank adaptation reduces error by 7.5% while updating only one percent of the model parameters, offering a strong accuracy-efficiency trade-off.
JAXbounce: A Continuous-Control Pong-Like Benchmark for Active Inference Beyond Low Dimensional Environments
AI summaryJAXbounce is a continuous‑control pong‑like benchmark that tests active inference in high‑dimensional, stochastic settings. Experiments show that state‑based agents with Kalman filtering and expected free energy planning outperform deterministic baselines, scaling with planning horizon, observation type, and noise knowledge.
ExeBench-FHD: Addressing Fidelity Issues in Decompiler Evaluation
ExeBench is one of the most widely used benchmarks for evaluating LLM-based decompilers. We analyze ExeBench and identify nine issues that affect how faithfully a test measures a decompiler. Three of them prevent LLM decompilers from recovering C code from ExeBench's function.…
TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization
AI summaryThe paper introduces TPO, a tri-level distributionally robust framework that addresses hierarchical out-of-distribution preference shifts in Direct Preference Optimization. Results show that TPO provides theoretical convergence guarantees and outperforms existing robust DPO methods.
Convergent Performance, Divergent Mechanisms in LoRA-Adapted Vision Foundation Models for Brain MRI
AI summaryThis paper investigates how vision foundation models with similar downstream performance achieve brain MRI adaptation through radically different internal mechanisms. The authors show that restricting LoRA adaptation to early-to-mid layers matches full-layer performance while improving parameter efficiency and cross-task robustness.
Geometry-Aware Block Selection for Sparse Variational Gaussian Processes
AI summaryThis paper introduces a geometry-aware algorithm to select block partitions for structured scaling matrices in sparse variational Gaussian processes. The method improves predictive performance and tightens the variational lower bound with minimal computational overhead.
Local Work, Global Progress: A Convergence Theory of AI Agent Collaboration
Teams of large language model (LLM) agents are increasingly applied to complex tasks, but their collaboration protocols are mostly designed empirically, without convergence guarantees. We formalize collaborative problem solving as block optimization: each agent improves one part…
Adaptive Replay Refinement: Allocating Test-Time Compute Between Revision and Resampling
AI summaryAdaptive Replay Refinement (ARR) automatically detects when a refinement step has converged and reallocates the freed compute to sampling a new solution, improving inference efficiency; on MATH500 it reaches 84.6% accuracy using only 41% of the serial passes of majority voting.
Learning Normal Coordinates on Riemannian Manifolds with Flow Charts
AI summaryThe paper introduces flow charts, conditional invertible neural networks that approximate normal coordinates on learned Riemannian manifolds. This approach enables fast Euclidean approximations of exponential and logarithmic maps, achieving a 60-70x speedup over numerical baselines with comparable accuracy.
Recommending Bacterial Cultivation Conditions from 16S rRNA
AI summaryThis paper introduces DSNI, a benchmark for predicting bacterial cultivation conditions from 16S rRNA sequences, and a Variability-Gated Encoder that leverages sequence structure. The method achieves 44.2% exact-match joint accuracy and recovers 90.2% of ground-truth protocols in top-25 recommendations.
Debiased Prior-Data Fitted Networks for Amortized Bayesian Semi-Parametric Inference
AI summaryThis paper analyzes the statistical behavior of prior-data fitted networks (PFNs) for semi-parametric Bayesian inference and introduces efficient influence function-based debiasing strategies. The authors prove that the proposed debiased PFNs achieve optimal convergence rates that match traditional frequentist estimators, establishing a bridge between amortized Bayesian inference and semi-parametric efficiency theory.
Disclosure Begets Disclosure: A Self-Reinforcing Privacy Loop in LLM Agents with Persistent Memory
Large language model (LLM) agents increasingly use persistent memory to personalize responses, and personal attributes in this memory pose a privacy risk if exposed. We ask how consulting memory during dialogue changes what users go on to say, and thus what the memory comes to…
ABACUS: A Benchmark for Image Editing Models to Assess Object Category and Counting Under Scene Conservation
AI summaryABACUS benchmarks image‑editing models for adding, removing, or changing a specified number of objects while preserving the rest of the scene, using a judge‑free, instance‑level check against panoptic ground truth. It shows that current models fail increasingly with higher counts and complex scenes.
Stabilizing Filters for Graph Signal Generation
Generating graph signals is an important task in machine learning, and incorporating the underlying graph into a generative model enables structurally-informed distribution modeling. However, said graph is often noisy or imperfectly estimated, and structural errors can affect…
Manifold-Informed Neural Networks for Low-Label ECG Electrolyte Inference
AI summaryThe paper introduces Manifold-Informed Neural Networks (MINNs), which replace the differential operators of physics-informed neural networks with graph manifold residuals to enable low-label inference. MINNs achieve superior electrolyte classification performance from ECGs, outperforming twelve baselines especially in extremely low-label regimes.
Observation Assimilation Transformer: Probabilistic Weather Analysis from Sparse Station Observations
AI summaryThe Observation Assimilation Transformer (OAT) uses a geometry‑aware transformer to refine NWP fields with sparse station data, producing probabilistic high‑resolution predictions at arbitrary locations. OAT outperforms baselines on regional and continental datasets, and downscales ERA5 to 1 km with lower MAE and CRPS.
Identifying Dependent Components from Multi-domain Linear Mixtures
AI summaryThis paper introduces a multi-domain linear mixing model with dependent latent components that uses domain-specific scalings to achieve identifiability from second-order statistics alone. The authors propose the Multi-Domain Covariance Matching algorithm, successfully recovering latent variables and mixing functions without relying on independence assumptions or parametric dependencies.
Characterizing Cross-Relay Generalization Failure in Machine Learning-Based Power System Protection
AI summaryMachine learning fault detectors trained on one relay’s data generalize poorly to others: F1 remains ~0.73 but false-positive rates rise dramatically (0.088→0.341), especially across bus groups, revealing location‑dependent error patterns.
Beyond the Point Mass: Acceptance-Derived Weights and Distributional Losses for Speculative Decoding
AI summaryThis paper introduces D-PAL, a training method for speculative decoding that replaces point-mass approximations with distributional overlap to weight token positions and guide drafter optimization. It improves mean acceptance length and generation throughput across six benchmarks compared to prior baseline methods.
CountVideoBench: A Benchmark to Evaluate Prediction Biases on the Object Counting task by Video-Language Models
AI summaryCountVideoBench evaluates video‑language models on object counting by presenting 2,600 synthetic videos with balanced counts from 5 to 50; none of the 16 models studied achieves over 27.6% exact match, and accuracy patterns vary, revealing prediction biases rather than pure difficulty.
Confessions of a Masked MDP
AI summaryThis paper studies reinforcement learning in Markov Decision Processes where the set of valid actions is initially unknown and must be learned alongside the optimal policy. The authors propose computationally efficient, minimax-optimal algorithms that actively discover action masks and apply them to transfer learning settings.
Relative Representations Enable Portable White-Box Monitoring Across Language Models
AI summaryThis paper introduces relative representations using angular distances to shared anchors, enabling white-box probes trained on one language model to transfer directly to other models without retraining. The transferred probes retain at least 85% of their original performance across diverse monitoring tasks.
Don't Skip the Baseline: A Controlled Comparison of Dynamic Expert Skipping in MoEs
AI summaryThis paper compares dynamic expert-skipping methods in MoEs against a static reduction of active experts. It finds that the simple static baseline is consistently difficult to outperform across multiple models and modalities.
Self-Credit: Credit Assignment for Long-Horizon Agentic RL via Intra-group Rescaling
AI summarySelf‑Credit generates dense supervision from sparse rewards by having an LLM agent cite crucial past actions, then rescales these self‑generated credits across successful and failed rollouts to emphasize truly critical actions. It achieves up to 37‑point gains on BrowseComp‑Plus and outperforms state‑of‑the‑art baselines on GAIA and xBench‑DeepSearch.
Rewards for Budget-Aware Multi-Turn Agents That Adapt to Arbitrary Test-Time Constraints
AI summaryThe paper trains multi‑turn language model agents to learn budget allocation across tokens, API calls, and turns, using new reward penalties for overspending. Agents achieve higher within‑budget success and natural compliance, generalizing to unseen budgets and tasks.
Brain2Vec: A Self-Supervised 3D Foundation Model for Structural Brain MRI
AI summaryBrain2Vec pre‑trains a self‑supervised 3D neural network on 74,425 T1‑MRI scans, yielding embeddings that improve brain‑age prediction accuracy and speed convergence versus random init, and outperform volumetric features on most downstream tasks.
DISTRACT: Degradation-Indexed Study of Textual Reference Across Competing Targets
AI summaryDISTRACT evaluates how textual descriptions influence vision‑language trackers under visual degradation. The study finds that only Molmo2 and JointNLT use text to guide tracking, with decodable effects at early decision stages, while other trackers show negligible or non‑significant text influence.
Region-Specific EEG Classification of Hypnotic Susceptibility Using a Sliding-Window 2D CNN: A Subject-Independent Evaluation
AI summaryClassifies hypnotic susceptibility using a 2D CNN on sliding-window EEG from seven brain regions. Subject‑independent LOSO testing shows only frontal and prefrontal regions remain above chance (0.646 and 0.624 accuracy), proving anterior cortex is a reliable marker.
Do Neural Models Prefer Their Submodels?
AI summaryThe paper analyzes whether loss minima of neural networks lie within their natural submodels. It shows that shallow MLPs never favor subnetworks with inactive neurons, deep MLPs rarely do, and most self‑attention submodels are not preferred, suggesting geometry alone doesn’t explain submodel preference.
CNN-based classifier for subject-independent MBTI Extraversion/Introversion preference classification from short-time EEG signals
AI summaryCNN-based EEG classifier predicts MBTI extraversion/introversion from 4‑second EEG segments; the compact SSCNet architecture achieves 64.8% accuracy, 0.706 AUC, 0.628 F1, with lowest fold variance and fastest training among tested models.
A Bochner Weitzenböck Decomposition for Cellular Sheaf Laplacians, Curvature and Neural Sheaf Diffusion
We extend the Bochner-Weitzenböck decomposition to cellular sheaf Laplacians, splitting it into a positive semidefinite term, the Bochner sheaf Laplacian, and a block diagonal term, the Forman curvature. We show that the Bochner Laplacians form a class of cellular sheaves that…
LocaLeJEPA: Maximum Entropy Representations under Geometric Priors
AI summaryLocaLeJEPA extends LeJEPA by replacing the isotropic Gaussian target with a locally Euclidean distribution that respects non‑Euclidean embedding geometries. On ImageNet, it yields higher linear and k‑NN accuracy for hyperbolic ViTs, equivariant ResNets, and standard ViTs.
Claims of emergence in AI should be grounded in information decomposition
AI summaryClaims of AI emergence are formally defined using information decomposition: a system is emergent only if it contains synergistic information about outcomes, separating unique, redundant, and synergistic contributions. The framework is applied to attention heads, agents, and human‑AI teams, linking and extending prior observations.
Geometry-Agnostic Data-Driven Discovery of Extreme Scenarios
AI summaryThis paper introduces a two-stage neural framework that infers geometry-agnostic dynamics from observational data and uses physics-informed search to discover unobserved extreme scenarios like solitons and instantons.
Model- and Layer-Level Closed-Loop Forecasting with SSM-Based Spiking Neural Networks
AI summaryClosed‑loop forecasting in spiking neural networks is achieved by feeding the network’s own predictions back as past inputs, at both model‑ and layer‑levels. On ETT‑small, these methods match or outperform state‑of‑the‑art continuous‑activation models, showing competitive performance with low‑bit activations.
Present but Hard to Read: Probing Latent Reasoning with Minimal Pairs
AI summaryProbes that read a model’s state only after the answer token miss some hidden reasoning signals: a digit copied from the question is detectable only when the full prompt is present, and while an intermediate sum can be read just before the final digit, it is not used in generating the answer.
BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models
AI summaryBitMTP jointly trains ultra‑low‑bit LLMs with an MTP objective, enabling confidence‑only, speculative‑only, and hybrid decoding without separate draft models. It achieves 2–3× speedups on 1.5–4 B models across NVIDIA H20, RTX 3090, and Jetson Orin while maintaining task quality.
ECG for laboratory values:A comprehensive physiological and technical analysis
AI summaryThe study trains multimodal deep learning models to predict NT‑proBNP from ECGs and metadata, comparing 1‑lead to 12‑lead setups. The 12‑lead model achieves MAE≈514 pg/mL and AUROC≈0.876 for binary classification, showing ECG can classify broad NT‑proBNP ranges but precise regression is still difficult.
Variance Reduction with Delayed Feedback for Non-Convex Optimization
We study variance reduction for non-convex optimization when gradient evaluations return with a delay. The server can update the parameter while earlier evaluations remain unfinished. We first analyze an ordered recursive momentum estimator for stochastic optimization. Under…
GraphCast-Mamba: Learning Non-Markovian Corrections for Weather Forecasting
AI summaryGraphCast-Mamba augments a frozen GraphCast backbone with a trainable Mamba correction branch that retains temporal memory across successive 6‑hour steps, learning non‑Markovian corrections. It reduces GraphCast’s ten‑day loss by 14.4% and day‑10 RMSE by up to 14.6% after only 2,500 updates.
Binary Discrete Diffusion: Candidate-Set Diffusion for Logic and Planning Tasks
AI summaryBinary Discrete Diffusion lets a diffusion model maintain candidate token sets as binary vectors, gradually adding/removing options to represent partial uncertainty. It outperforms masked and uniform baselines on hard planning tasks (Sudoku-Extreme, Boxoban, graph coloring) and remains competitive on language modeling.
Discount Yourself: Self-inclusive Freshness-weighted Gossip for Decentralised Learning under Client Outages
AI summarySelf‑inclusive Freshness‑weighted Gossip (SFG) discounts a client’s own stale model in decentralised federated learning by using effective lag and steep decay, guaranteeing bounded error during outages. Experiments on image benchmarks and real smartphone traces show SFG‑L and SFG‑Cut outperform existing gossip and age‑based rules.
Learning Computational Structure for Sample-Efficient Prediction
AI summaryLearned Wiring learns a low-dimensional embedding per neuron, connecting only nearby neurons while jointly training weights and connectivity, achieving 27 % lower sample‑curve area than dense nets on 32 synthetic compositional targets. It also outperforms dense baselines on 16/17 real-world model‑prediction tasks.
When Do LLMs Help with Feature Selection? A Systematic Empirical Study
AI summaryLLMs can use semantic feature names and summary stats for tabular feature selection, but a systematic study on 18 datasets shows classical baselines outperform 7B LLM selectors on ROC‑AUC and R², though LLMs give more compact subsets and modest gains when feature names are informative.
Convergence of stochastic gradient descent for learning linear neural networks
AI summaryThe paper proves that SGD, with either fixed or adaptive step sizes, converges almost surely to critical points in deep linear networks under regularized squared loss, showing bounded iterates without requiring Lipschitz continuity or whitened data.
OptScientist: Automated Discovery of Optimization Laws
AI summaryOptScientist automatically infers symbolic optimization problems from data, using an LLM‑guided discovery pipeline. It recovers ground‑truth physics laws on 80 %+ of a 99‑problem benchmark and outperforms prior methods on real‑world tasks while using far fewer parameters and tokens.
Asymmetric Bidirectional Predictive Dynamics in Google Search and COVID-19 Time Series: A Multi-Horizon Forecast-Error Analysis
Online search activity is widely used as a digital trace of information seeking during epidemics, but its predictive value is often reduced to correlation, a selected lag, or a single forecast-error score. We examine eight pre-specified Google Trends-COVID-19 pairs in both…
The Dynamics of Collapse and Recovery in Clustered Federated Learning
AI summaryThe paper models clustered federated learning as a controllable dynamical system and shows that greedy client routing can cause cluster collapse. By using blocked EXP3 to intervene and redirect updates, the method recovers specialist models and improves performance after bad initializations.
Continual Concept Personalization in Diffusion Models via Residual-Space Adaptation
AI summaryContinual concept personalization for diffusion models learns new subjects sequentially via residual‑space adaptation, achieving superior concept fidelity and almost no forgetting. The method identifies key layers, allowing 72% fewer trainable parameters without sacrificing performance.
Evolving Semantic Indexes for Generative Retrieval
AI summaryThe paper uses LLM-driven evolutionary search to mutate generative retrieval components, guided by a fast 1M‑parameter proxy decoder. Across 21 SID programs on NQ320K, evolution boosts Recall@1 from 55.8% to 59.3% and Recall@100 from 85.8% to 90.6%.
The Token-Energy Law of LLM Decode: Active Bytes and Realized Time Account for Joules per Token
AI summaryThe paper introduces the Token‑Energy Law, modeling GPU energy per output token as an affine combination of active bytes (shared weights vs private KV history) and realized decode time. Calibrated on BF16 runs on Hopper/Blackwell GPUs, it predicts held‑out energy within 3.73%.
Continuous-Time Policy Gradient from Discrete Data: Convergence Analysis and Model-Free Algorithms
AI summaryThe paper builds a continuous‑time policy gradient method that works with only discrete observations, proves exponential convergence of an entropy‑regularized flow, and gives a bound separating critic, regression, and step‑size errors. It integrates this with PhiBE to produce model‑free actor‑critic algorithms that maintain continuous‑time structure, validated on LQR and wheeled‑robot tasks.
Two $\operatorname{ReLU}$ Layers is All You Need
We prove that every $\operatorname{ReLU}$ network can be exactly simulated by a $\operatorname{ReLU}$ network with two hidden layers. We present two proofs: one that is based on fillings in homological algebra and a direct combinatorial proof.
WildMAE: Structure-Aware Masked Autoencoding for Wildlife Image Reconstruction
AI summaryWildMAE extends MAE with a decoder-side, structure-aware refinement that uses zero-gated convolutions and gated cross-attention to recover masked patches. It achieves 27.90 dB PSNR, 0.840 SSIM, and 0.160 LPIPS on Snapshot Serengeti, outperforming vanilla MAE.
PaPi: Parallel Picard-Duhamel Integration for Fast Neuro-Spectral Architectures
AI summaryPaPi introduces a time‑parallel integration method for neuro‑spectral architectures, using exponential integration and Duhamel‑Picard iteration. It converges exponentially for any flow and reduces training/inference time by up to 100× while retaining accurate predictions.
Neural Implicit Energy Optimization for LiDAR Mapping
AI summaryNeural implicit LiDAR mapping is formulated as optimizing a curvature-based energy over the SDF’s shape operator, yielding smoother surfaces while preserving fidelity on million-point scans, with no extra inference cost.
Turning TS2Vec into a Foundation Model for Time Series Classification
AI summaryTS2Vec is adapted into a foundation model by training a single 1.6‑M‑parameter dilated CNN once on synthetic data, using dual stems for amplitude and shape views, and distilling them into one encoder. It attains 83.3% on UCR‑128 and 75.3% on UEA‑27, matching state‑of‑the‑art MantisV2 within 0.3 pp.
Efficient Edge-aware Attention Network for Graph Generation
AI summaryThe paper introduces an edge‑aware attention layer that integrates graph structure into pairwise scores and updates edge features with a lightweight residual mechanism, enabling use of hardware‑optimized FlexAttention. It matches baseline generation quality while cutting GPU memory by ~65% and boosting throughput.
Adaptive Multi-Token Prediction: How far should we go?
AI summaryAMTP learns a context‑dependent draft horizon online, predicting how far a multi‑token draft will survive and selecting the horizon that maximizes tokens per second. It speeds up decoding up to 4.1× and matches or exceeds fixed‑horizon performance in most tested settings.
Planning under Noise with Neural Operators
AI summaryPUN is a neural operator that learns noise‑aware value fields for stochastic planning. It outperforms baselines, achieving up to 100% success on 4‑D replanning tasks while running 3‑32× faster across noise levels.
Learning Adaptable Dexterous Skills via Hybrid Dexterous Mimicking
Dexterous multi-finger hands hold promise to unlock diverse skills like functional grasps, in-hand reorientation, and dynamic handover that the human-designed world assumes, yet existing learning recipes force a trade-off. Goal-reaching reinforcement learning generalizes to…
Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs
AI summaryThis paper introduces a model that jointly generates synchronized video and 3D motion for arbitrary skeletons using a decoupled architecture. The approach outperforms video-first pipelines in faithfully animating diverse rigs and placing them into consistent 3D scenes.