New Machine Learning (cs.LG) papers, Friday 23 October 2026
207 papers announced on arXiv. Get them ranked for you, free →
Domain-Specific Constitutional AI: Enhancing Safety in LLM-Powered Mental Health Chatbots
AI summaryThis paper introduces a domain-specific Constitutional AI approach tailored for mental health chatbots. The method integrates specialized psychological principles to improve crisis intervention accuracy and prevent harmful misinformation in sensitive clinical interactions.
ChemBART: Pre-training a Reaction Language Model for Comprehensive Synthesis Planning
AI summaryChemBART trains a reaction‑centric BART model with mask‑filling on whole reaction strings, enabling tasks from precursor generation to property prediction. It integrates into MCTS for multi‑step synthesis, yielding experimentally validated routes that improve literature yields by over 20%.
Do Machine Learning-Guided Proactive Full Capacity Protocols Mitigate Emergency Department Overcrowding? A Simulation-Based Evaluation
AI summaryThis study uses a machine learning and simulation framework to evaluate proactive emergency department capacity protocols triggered six hours ahead of predicted overcrowding. The results show that proactive activation significantly reduces the time spent above crowding thresholds and stabilizes operations compared to conventional reactive methods.
The Cost of Saying Hello: An Empirical Study of Handshake Overhead in AI Agent Communication Protocols
AI summaryThis paper provides the first empirical benchmark of handshake overhead across major AI agent communication protocols. It finds that round-trip counts and capability metadata parsing dominate costs, consuming significant latency and up to eleven percent of context windows before task execution begins.
baseVultra: Compact Nonlinear Recurrence with Reallocated Capacity
AI summarybaseVultra reallocates recurrent capacity from gates to candidate generation, using a single dense H×H matrix and diagonal gates. It outperforms matched LSTM/GRU in compact regimes (TinyShakespeare loss 1.594 vs. 1.604/1.616) and remains competitive at larger scales.
IRAH: Hierarchical Agentic Tree Search for Retrieval Pipeline Optimization
AI summaryIRAH jointly optimizes retrieval pipelines by hierarchical UCT-guided search (macro structure first, then parameters) with language-agent reflections and meta-learned warm starts. It outperforms TPE, achieving up to 6.4% nDCG@10 gains and the best average 0.4816 on four benchmarks.
Detecting AI Training in Datacentre Power
AI summaryThis paper introduces a Viterbi-based tracking method to detect large-scale AI training runs using facility-level power consumption data. The main result shows the tracker detects at least 96 percent of unhidden runs and overcomes most adversarial evasion strategies by making hidden training significantly more computationally expensive.
SIGNN++: Robust Graph Neural Star Identification on Earth
AI summarySIGNN++ improves graph‑neural‑network star identification for terrestrial navigation by adding realistic noise, Fourier edge features, and RANSAC alignment. It achieves higher accuracy and solve rates than previous SID methods while keeping attitude accuracy and runtime practical.
Agentic Verified Optimization of Real-World Software
Program optimization aims to improve performance while preserving behavior. Coding agents can already make programs faster by changing algorithms, data structures, and execution strategies. However, existing equivalence checks mostly rely on tests, which agents have been shown…
Why Constrained Updates Still Forget: A Functional-Path Analysis of Continual SFT
AI summaryThe paper decomposes how constrained updates in supervised fine‑tuning affect old‑task performance, showing that update placement and granularity drive forgetting. It demonstrates that matched replay schedules and subspace regularization can better preserve old behavior while improving new‑task gains.
Differentially Private Synthetic Data via Generalized Moment Matching
AI summaryPrivate synthetic data are generated by estimating moments with differential privacy, fitting a distribution, and sampling. The method achieves minimax‑optimal error (up to logs) for IPMs such as MMD with Matern/Laplace kernels, sliced Wasserstein‑1, and ReLU‑network metrics.
MoEZip: Routing-Aware KV Cache Compression for Sparse Mixture-of-Experts LLMs
AI summaryMoEZip compresses KV caches for MoE LLMs by using a Fisher‑information–based metric to gauge expert routing sensitivity, combining it with attention weights for routing‑aware eviction. It outperforms baselines, gaining +24/+20 average points on LongBench/RULER at 10% cache retention.
Spectral Uncertainty Decomposition for Sparse Bayesian Learning
AI summaryThis paper introduces a spectral framework for M/EEG source imaging that decomposes the Bayesian posterior covariance into data-informed and measurement-invisible components. The method successfully distinguishes reducible measurement uncertainty from fundamentally irreducible indeterminacy.
Prediction Is Not Calibration: Item Parameter Recovery and Representation Coupling in Knowledge Tracing
AI summaryThis paper investigates why accurate knowledge tracing models fail to reliably recover item parameters like discrimination. The authors show that separating sequence representations from item parameter keys improves parameter recovery and cross-order consistency without sacrificing predictive accuracy.
AI Mindmeld: A Latent Interconnect Protocol
AI summaryThe paper introduces the Latent Interconnect Protocol, a framework that learns lightweight connectors to map specialist models’ high‑dimensional representations directly into an LLM’s internal state. Experiments show that frozen LLMs can use these connectors for zero‑shot transfer and improved compositional reasoning across domains.
When Reconstruction Is Not Enough: Doubly Robust Counterfactual and Interventional Mediation Analysis with Learned Multimodal Representations
AI summaryThis paper shows that standard representation learning objectives like reconstruction can fail to preserve causal identification in multimodal health data, leading to severe bias in doubly robust mediation estimators. The authors propose a validated workflow including diagnostics and cross-fitted estimators to ensure reliable counterfactual and interventional analyses.
How Prompt Style Affects AI Accuracy and Speed in Competitive Programming
AI summaryThe paper evaluates nine prompt styles for GPT‑5.5 on IOI accuracy tasks and NP‑hard speed tasks, finding that the best prompt varies by task and statistic (mean vs. median). AI performance is comparable to human contestants, sometimes better, sometimes worse.
GANet: A Lightweight Group Attention Network for Plant Disease Identification on Embedded Systems
AI summaryGANet is a lightweight 0.001M‑parameter Group Attention Network for identifying 38 plant diseases. It achieves 99.99% accuracy on PlantVillage and runs in real time on low‑power devices, reaching 5 FPS on Raspberry Pi 4b and 11 FPS on Jetson Nano.
Benchmarking Model Compression Techniques Across Heterogeneous Edge Platforms: A Predictive Maintenance Use Case for Rotating Machinery
AI summaryThe study compares four compression methods on a CNN for rotating machinery maintenance, tested on nine edge devices. Knowledge distillation yields the best trade-off: 90% faster inference, 91.22% F-score, 98.9% size reduction, enabling efficient Industry 5.0 deployment.
Automated screening of privacy risks in medical AI
AI summaryThe paper measures per‑patient vulnerability to membership inference attacks across 7,789 medical AI models, finding that risk rankings are stable across training choices. It presents a lightweight tool that flags high‑risk records before model training, reducing exposed cohorts.
Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal Reasoning
AI summaryThe paper proposes minimal‑edit on‑policy distillation for vision‑language math models, correcting student outputs by penalizing only the teacher‑disfavored tokens while rewarding correct rollouts. It outperforms or matches models with reference figure descriptions, improving perception and transferring to math benchmarks.
A Quadratic Lens on Muon's Polar Update: Benefits, Pitfalls, and the Role of Newton-Schulz
AI summaryPolarGD, the orthogonalized update used by Muon, reaches the optimum in a single step for matrix quadratics when error rows are orthogonal and equal norm, independent of Hessian eigenvalues. It preconditions by the current iterate, so its speed depends on iterate conditioning, and limited Newton‑Schulz iterations often outperform full orthogonalization when tiny singular values dominate.
SWE-VERIFIER: Selective Verification for Cost-Efficient Coding Agents
AI summarySWE‑Verifier is a lightweight model that selectively triggers sampling and ranking of candidate actions for coding agents, deciding when to verify alternatives. It boosts MiniMax‑M2.7 resolution from 75.47 % to 77.20 % at only 1.09× cost, while verifying less than 4 % of steps.
Generalizing Metropolis-Hastings for DAG Sampling
AI summaryThe paper introduces duet sampling, a generalization of Metropolis‑Hastings that chooses two proposal distributions per iteration to resample parents of multiple nodes in a DAG. This yields Markov chains that mix faster and outperform state‑of‑the‑art DAG‑learning methods.
GAV: A scientific experimentation environment for PDE agents
AI summaryGAV creates verifiable PDE problems by learning a distribution of reference solutions, then using manufactured solutions to generate matching PDEs, allowing agents to iteratively improve coding programs with accurate feedback. On Navier-Stokes, it reduces test error of existing agents by over tenfold.
EgoHand: Monocular 3D Hand Reconstruction from Egocentric Video
AI summaryEgoHand reconstructs 3D hand poses from single egocentric video, trained on 14 M diverse images and using SOMA-Hand with a transformer encoder and camera cues. It outperforms baselines on HOT3D, ARCTIC, and H2O with 9.7 ms latency and gains user preference over competitor models.
SciFusion: Physics-Guided Compositional Diffusion for Multimodal 3D Scientific Image Generation
Existing 3D generative approaches are typically developed for individual imaging modalities, requiring substantial modality-specific data and limiting their ability to exploit shared structure across scientific imaging domains. We present SciFusion, a multimodal framework for…
TerraTransfer++: Demonstration-Free End-to-End Planner
End-to-end autonomous driving has achieved state-of-the-art performance on benchmarks and real-world deployments. Its standard training recipe, however, is expensive across all existing paradigms: imitation learning requires fleet-scale driving logs whose behavior distribution…
Spike-Aware Gradient-Based Mixed-Precision Quantization of Spiking Neural Networks for Vision Applications
AI summarySpike-aware gradient-based mixed-precision quantization for SNNs assigns bit-widths via first-order gradients, avoiding Hessian overhead. It achieves ImageNet ResNet-34 accuracy of 73.30% with 4.78× compression, 2.31× less memory, and 100× faster bit-allocation than HAWQ‑V2.
AQL-Lite: Enabling Efficient Knowledge Acquisition for Arabic LLMs via Data-Centric Continual Pretraining
AI summaryAQL‑Lite curates Arabic data and fine‑tunes a multilingual LLM with partial‑parameter CPT (adapters + last 8 blocks + embeddings) to MSA, achieving a 3.51‑point average gain on 1B tokens and 94 % of full‑parameter gains in half the time.
On the Adversarial Vulnerability of Neural Receivers with Unrolled LDPC Decoders
AI summaryNeural 5G receivers using unrolled LDPC decoders are vulnerable to adversarial input. Gradient masking from tanh saturation reduces attack potency, but new per‑iteration and LLR‑margin losses restore effectiveness, boosting BLER by 3‑4 orders of magnitude.
Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable
AI summaryThe paper trains a policy via self‑rewarded argument quality to improve LLM moral judgments, showing higher decisiveness and order‑invariance on benchmarks, but revealing that consistency under reordering does not equate to valid moral decisions.
Route-and-Generate: KL-Regularized Router Improvement for Reasoning Routing in LLMs
AI summaryRoute-and-Generate trains a neural router to allocate a fixed reasoning budget across multiple formats, using KL‑regularized policy improvement. The resulting LLM produces concise answers for easy questions and longer reasoning for hard ones, improving accuracy–efficiency trade‑offs on reasoning benchmarks.
CTRL-FM: Inference-Time Structural Control for Flow-Based Virtual Sensing under Long-Horizon Node Outages
AI summaryCTRL-FM learns a flow‑matching model that keeps observed data fixed and selects, at inference time, the best structural control (graph, temporal, or balanced) based on self‑supervised proxies and local reliability. It outperforms fixed‑bias methods on up to 60 % node outages.
Probing and Steering for Safety-Relevant Behaviours in Large Language Models
AI summaryThe paper studies how activation probes—directions in a Llama‑3.1‑8B‑Instruct residual stream—can detect and steer safety‑relevant behaviours (deception, high‑stakes awareness, sycophancy). It shows that detection works best with logistic regression, steering with difference‑of‑means, but the pattern reverses for sycophancy, with performance also depending on token‑pooling and layer choice.
Bayesian Fixed-Confidence Best-Arm Identification with Abstention
AI summaryBayesian fixed‑confidence best‑arm identification with an optional abstention. The proposed Screen‑and‑Confirm algorithm matches lower bounds: expected stopping time ≈κν²[1+log((α+δ)/δ)]/(α+δ). In the no‑abstention case it achieves optimal Θ(1/δ) time.
A Transformer-based Approach for Real-time Multi-agent Path Planning with Varying Initial/Final Conditions
AI summaryThe authors propose OCFormer, a physics-informed transformer that maps a tractable latent space to interacting physical spaces for multi-agent path planning. Trained without trajectory data, the model computes near-optimal collision-free trajectories for up to 128 agents in real time.
Soft Ontology Priors over Latent Concept Posteriors for Weakly Supervised Neuro-Symbolic Learning
AI summaryInjects a soft ontology prior into the latent concept posterior of weakly supervised neuro-symbolic models, producing a smooth probabilistic circuit that bounds shortcut reasoning and remains robust to noisy axioms, validated on BDD-OIA and synthetic NeSy tasks.
Escaping the Phylogenetic Trap: Persistent Homology Enables Cross-Species Ion Channel Classification via Multimodal Graph Neural Networks
AI summaryThe authors evaluate ion‑channel classification across species, showing that random splits inflate performance. Using persistent‑homology graph features, their PH‑GNN achieves 63.4% accuracy under leave‑one‑taxon‑out, 8 points higher than other geometric priors.
AcoustiPrint: Physics-Guided Neural Design of Manufacturable Acoustic Structures
AI summaryAcoustiPrint uses implicit neural representations and surrogate gradients to jointly optimize 3D acoustic structures while enforcing discrete manufacturing constraints. It achieves superior focal accuracy in transcranial focused ultrasound compared to existing baselines.
High-Value Training Data from Noisy, Diverse Sources
AI summaryThe paper introduces scalable data‑weighting via metagradients, learning parametric rules from a small high‑quality target set and transferring them to larger models and streaming data. It shows that learned weights improve reward‑model performance and combine human and synthetic signals better than either alone.
Memory-Augmented Online Test-Time Adaptation for EEG Motor Imagery via Perturbation Consistency and Confidence Calibration
AI summaryMemory-augmented online test-time adaptation (MOPC) updates EEG motor‑imagery models in real time using random‑mask perturbations for consistency, confidence‑calibrated reweighting to avoid bad pseudo‑labels, and a lightweight memory to avoid forgetting. It outperforms prior methods by ~3 % on BCI IV IIb and OpenBMI with 40–60 ms per‑trial latency.
Towards Faithful Data Selection: Statistical Limits and Scalable Algorithms
AI summaryThe paper shows that minimizing population error via data selection is impossible without a large validation set, so it introduces a faithfulness guarantee based on an upper bound on KL divergence for ULA-trained models. A swap‑based local search improves any selector while preserving this bound.
Stability Analysis of Signed Graphs under Structural Perturbations
AI summaryAnalyzes how edge additions, deletions, and sign flips affect stability of signed graph filters, showing joint dependence on perturbations and node feature correlations. Proposes a theory-guided strategy that selects influential signed-edge changes, yielding larger embedding drift and prediction degradation than baselines.
Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning
AI summaryThe paper introduces BELT, a method that separates outcome prediction from stopping control in reasoning LLMs to optimize compute. BELT reduces generated tokens by up to 51% without sacrificing accuracy while transferring effectively to new domains with minimal training cost.
aggmetrics: Aggregating Classification Performance Measures for Machine Learning
AI summaryaggmetrics extends binary classification metrics to multiclass, multilabel, ordinal, cost‑sensitive, soft‑label, and cross‑validation settings by offering four aggregation operators. It auto‑detects metric properties, warns of mis‑applications, and provides a test suite verifying the theory.
Short-to-Long: Detached Features for Long-Horizon Value Learning
AI summaryShort-to-Long (S2L) trains short‑discount source networks to produce features, then uses these detached features to guide a long‑discount behavior learner; across Atari and continuous‑control tasks, S2L improves median performance on 40 of 57 games and reduces target drift.
Flow-based Spectral Kernel Learning for Nonstationary Attention
AI summaryFlow-based Spectral Kernel Attention learns nonstationary attention kernels via normalizing flows over spectral densities, enabling flexible, regularized kernel learning while retaining linear‑attention efficiency. Experiments show competitive performance and 40‑47 % fewer projection parameters than baselines.
Learning to Incentivise: Using Reinforcement Learning for Sustainable Urban Mobility
AI summaryThis paper uses multi-agent reinforcement learning to dynamically allocate driver incentives for route-switching, aiming to improve sustainable urban mobility. The method reduces total travel time by up to 16% or CO2 emissions by up to 9%, while revealing a trade-off between the two objectives.
DiffLTM: Continuous Diffusions Large Tabular Models for Conditional Generation
AI summaryDiffLTM trains a continuous diffusion model for tabular data, using cross‑entropy loss for categorical columns and new metrics targeting rare sub‑populations. The resulting model outperforms baselines in conditional generation accuracy, even capturing classical econometric patterns.
Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy
AI summaryThis paper introduces an atomic interpretable unit framework for evaluating sparse autoencoders, revealing that standard metrics miss hidden polysemanticity in dominant features. The results show that widely used SAE features are often functionally redundant and less interpretable where it matters most.
FIGFLEX: A Mixed-Precision FP-INT GEMM Accelerator Co-Designed with Low-Bit LLM Quantization
AI summaryFIGFLEX is a mixed‑precision FP‑INT GEMM accelerator with a shift‑add radix‑4 datapath that halves partial products, plus an INT3 optimization, and a mixed‑precision quantizer that assigns per‑channel bit‑widths. On LLaMA‑2‑7B it achieves lower perplexity (5.85/8.77 at 3/2‑bit) than GPTQ, AWQ, SliM‑LLM, AMQ, and QuIP and delivers 1.56× and 1.42× higher throughput per area than AxCore and FIGLUT, respectively.
Optimal Multitask Representations Under a Rate Constraint: An Information Bottleneck Analysis
AI summaryMultitask IB formalizes pretraining with a mutual‑information budget; for Gaussian tasks it shows a task‑aligned specialist beats a reconstruction generalist at low rates but not at high rates, and that a task‑aware objective can uniformly bound regret. Experiments confirm compression can swap which objective and task dominate.
NEST: Nested Embedding Representation for Multiscale Clustering
AI summaryNEST learns a single embedding that encodes a nested, multiscale clustering hierarchy, optimizing for separation across progressively higher‑dimensional prefix subspaces. It recovers the true hierarchy on synthetic data and reveals multiscale organization in vision, text, and single‑cell datasets.
LLM2Cert: LLM Synthesis of Certificates with SMT/Lean Verification
Lyapunov functions and barrier certificates give state-space guarantees of stability and safety, but synthesizing and verifying them---especially for nonlinear, high-dimensional systems---remains an open challenge. Existing pipelines verify candidates with numerical solvers…
Decision Making and Ride Comfort Evaluation Based on Lane-Change Prediction for Autonomous Vehicles
Predicting surrounding vehicles' lane-change intentions enables autonomous vehicles to anticipate potential hazards and respond proactively rather than relying solely on reactive maneuvers. This study investigates the benefits of integrating lane-change prediction, based on…
Lifetime-valid decisions by continually evolving AI analysts
AI agents increasingly act as analysts whose decision can often be cast as hypothesis tests. Classical tests control the Type I error rate, the probability of rejecting a true null hypothesis, for fixed procedures. However, AI analysts introduce internal randomness and can…
When News Gets Old: Portfolio Allocation with Reinforcement Learning and Large Language Models
AI summaryThe study uses an LLM to extract signals from SEC disclosures and tests two approaches to aging news in portfolio allocation: keeping news with timing cues versus filtering it out. Filtering old news yields the highest net return, 18.43%/yr, surpassing RL strategies.
Bridging the Closed-World Gap: Interpretable Anomaly Detection via a Hybrid TSK-FCM Ensemble on the BETH Dataset
AI summaryHybrid TSK‑FCM ensemble uses a supervised TSK classifier for known insider threats and a parallel Fuzzy C‑Means cluster for geometric anomalies, with a Min Distance metric mitigating high‑dimensional fuzziness. It achieves 0.9951 AUROC on BETH, 17.1% above Isolation Forest.
Evaluating Type-1 Diabetes Digital Twins by Decision Quality
AI summaryDigital twin models for T1D are evaluated on how well they rank insulin treatments for individual patients. The study shows that a model with the lowest prediction error can still give poor treatment rankings, highlighting the need to assess decision quality directly.
Future of Time Series Anomaly Detection: From Historical Review to Unified Outlook
AI summaryThis paper surveys recent TSAD architectures—diffusion, state‑space, LLMs—introduces a refined anomaly taxonomy, evaluates metrics via extensive experiments, and releases an open benchmark, highlighting challenges in multivariate detection, zero‑shot transfer, and interpretability.
Interpreting Model Reasoning through Behavioral Counterfactuals
AI summaryThe paper introduces behavioral counterfactuals, editing the model’s residual stream to match next-token distributions between a source and counterfactual prompt. It shows this method reliably recovers counterfactual behavior, identifies key causal locations, and reveals how suggestions subtly shift reasoning in large LLMs.
Fast Iterative Combinatorial Auctions via Deep Reinforcement Learning
Iterative combinatorial auctions are used in spectrum allocation, transportation, and industrial procurement, but rounds are costly. We introduce FastCombo, a deep reinforcement learning approach to sequential price discovery. Bayesian pricing has sharply reduced auction rounds…
RiboBridge: protein-conditioned de novo RNA binder structure design
AI summaryRiboBridge designs RNA binders by first generating RNA conformations in a pretrained latent space, then predicting protein‑RNA interaction geometry, and finally rigidly placing the RNA on the protein surface. It outperforms prior methods on a 124‑complex benchmark, achieving higher fNAT, DockQ, and lower RMSDs.
Scaling Laws for Safety Probes
AI summaryThe paper shows that safety‑probe accuracy scales predictably with model size and label count, following two power‑law relationships that hold across tasks, families, and architectures. Using these laws, researchers can choose probe size, label budget, and architecture within 0.02 error of actual performance.
Learning Domain Trajectories with Flow Matching for Gradual Domain Adaptation
AI summaryThe authors propose FlowGDA, a flow-matching framework that models gradual domain adaptation as continuous transport on a statistical manifold using a simulation-free regression objective. Experiments show that this approach outperforms existing state-of-the-art techniques for handling distribution shifts.
An Agent Architecture with Tools and Memory for Selective Correction of Wrist PPG Heart Rate Estimates
AI summaryAgent architecture with tools and memory selects or corrects wrist‑PPG heart rate estimates; tested on four datasets, it reduces MAE by up to 44.5%, with selective corrections improving accuracy while quantifying risks.
AIRO: A Streaming-Based Machine Learning Framework for Threat Detection and Security in Industrial Internet of Things
Modern Industrial Internet of Things (IIoT) im- plementations now connect edge gateways, cloud dashboards and business applications. This improves visibility and remote monitoring. However it simultaneously exposes industrial traffic to various cyber threats including scanning,…
Solvable models of learning to pursue a moving target
AI summaryThis paper introduces a solvable linear-quadratic model of target pursuit to study how agents learn to intercept moving targets. The main result is a closed-form characterization of learning dynamics, revealing a spectral gap that shows agents learn to reach the target before optimizing their movement efficiency.
Federated Graph Retrieval-Augmented Generation
AI summaryFederated GraphRAG enables large language models to retrieve and generate knowledge from distributed, private knowledge graphs by routing queries to relevant clients and converting subgraphs into abstract evidence. It outperforms baselines on retrieval and generation tasks in federated settings.
Width and Depth Limits Commute During Training
AI summaryShows that for linear residual nets scaled by 1/√depth, the infinite-width and infinite-depth limits commute during training with gradient clipping, and still hold in probability without clipping. The network outputs converge in L² for any fixed number of steps.
Random Remasking Scales Masked Diffusion Inference
AI summaryThis paper re-formulates masked diffusion inference as a Markov chain with random remasking, introducing a test-time correction procedure and training method. The approach achieves near-log-linear scaling on logical tasks and outperforms existing baselines on language generation benchmarks.
Wavenumber-Aware Deep Learning for Robust Analysis of Raman Spectra
AI summaryThis paper introduces a wavenumber-aware positional encoding for deep learning on Raman spectra, treating the physical coordinate as input rather than an array index. The method enables models to maintain high accuracy under noise, sparse sampling, and unseen measurement grids.
Visualizing and Interpreting Temporal Care Team Dynamics for Cancer Survival Risk Analysis
AI summaryThis paper introduces a visual analytics framework combining dynamic collaboration networks and temporal graph neural networks to model how evolving care team dynamics relate to patient cancer survival risk over time.
Primal-Dual Representation Learning for Low-Rank Constrained MDPs
AI summaryRepresents low‑rank continuous‑state CMDPs with a decaying‑ε greedy learning phase, then optimizes an optimistic Lagrangian. REP‑PD‑TEG attains a \~O(√K) regret with \~O(K^{3/4}) samples for soft violations, while REP‑PD‑H‑TEG gives hard‑violation guarantees with similar regret.
Adaptive Deadline- and Energy-Aware Caching Algorithms in Dynamic Multi-UAV Edge Networks
AI summaryThis paper introduces GRU-DEPPO, a reinforcement learning-based caching framework for multi-UAV edge networks that jointly optimizes content delivery deadlines and UAV energy consumption. Simulations demonstrate that it significantly reduces deadline misses and energy imbalances compared to baseline methods.
One Pretrained Decision Model Is Enough for Training-Free Generalist Graph Anomaly Detection
AI summaryJev‑GAD identifies anomalous graph entities without training by combining structural, attribute, and neighborhood evidence into percentiles, feeding them into a frozen language model’s decision mechanism via a Graph‑to‑State transform, then fusing statistical ranks to achieve 89.2% AUROC and 52.4% AUPRC.
Gandalf: Flash Shall Not Die! Increasing Flash Storage Lifetime for KV Cache Offloading in LLM Serving
AI summaryGandalf offloads KV cache to SSD while prioritizing full‑attention entries, extending SSD life by 7.3× on average with negligible hit‑rate loss. It uses attention‑type awareness to reduce unnecessary writes, preserving cost‑efficiency for LLM serving.
Trace2World: Reconstructing App Worlds from Real Interaction Traces for Scalable GUI Agent Training
AI summaryTrace2World separates behavior inference from implementation to reconstruct realistic app worlds from real interaction traces, building 50 executable environments from 1,570 traces. It achieves 91.1% end‑to‑end reproduction versus 59.8% for baselines, and boosts training gains by up to 20.2 points on internal and 16.3 on real-device tasks.
GIFT: Gradient Informed Filtering for Distillation Augmented Reasoning RL
AI summaryGIFT selectively filters token‑level distillation gradients using a geometric split into RL‑parallel (conflict) and RL‑orthogonal (value) components, ranking guidance by safety and novelty. Experiments show it improves mathematical reasoning by up to 9.5 % over baselines.
Amplifying Latency and Cost of LLM Inference via Rowhammer Attack
Large language models (LLMs) achieve high inference efficiency using optimized GPU kernels generated by just-in-time (JIT) compilation frameworks such as Triton. Triton compiles GPU kernels through an MLIR/LLVM-based pipeline. We investigate whether a Rowhammer-induced bit-flip…
A Computationally Efficient and Rate-Optimal Algorithm for Adversarial Linear MDPs with (Aggregate) Bandit Feedback
AI summaryThe paper presents a polynomial‑time algorithm for episodic linear MDPs with unknown transitions, adversarial losses, and only aggregate bandit feedback. It achieves $\tilde O(\sqrt{K})$ regret, matching the optimal rate previously attainable only by inefficient methods.
Is spectral predictability enough? Testing {\Omega}-matching for synthetic pretraining under distribution shift
AI summarySynthetic pretraining benefits from matching spectral predictability Ω to the target only in controlled synthetic‑to‑synthetic setups, where Ω‑matched data can beat an oracle by 46.8% MSE (LSTM). On real data, Ω matching gives marginal gains or even drops performance, indicating Ω alone is insufficient for transfer.
What Lies Beyond Perception: A Partition Theory of the Limits of AI Evaluation and Oversight
AI summaryThis paper introduces a partition theory of AI evaluation showing that an evaluator's judgment is fundamentally bounded by its perception. The main result proves that behavioral evaluation can never certify an upper bound on a model's capabilities or reliably detect hidden capabilities without assuming the model's cooperation.
xVGAE: A Hierarchical Variational Graph Autoencoder for Exchangeable Graphs
AI summaryxVGAE learns a graphon from training graphs using a graph-level latent variable, node-level exchangeable latents, and an edge probability function, producing new graphs that match training statistics and generate up to two orders of magnitude faster than current models.
Neural Spectral Capacity with Residual Context for Heterogeneous Resource Allocation within Transformers
AI summaryResNSC augments Neural Spectral Capacity with residual-aware statistics and a dynamic programming allocator that jointly sets residual and feed‑forward widths under a budget. On 121M/374M Transformers, ResNSC‑DP yields lower perplexity than uniform or tapered baselines and improves LLaMA‑7B pruning accuracy by 3–5 pp.
Semi-Linear Gaussian Bayesian Networks: Enhancing Flexibility in Node Dependencies via penalized B-Splines
AI summaryThe paper introduces Semi-Linear Gaussian Bayesian Networks, which use penalized B-splines within generalized additive models to capture nonlinear dependencies while maintaining Gaussian conditional distributions. Experiments show this framework improves goodness of fit and structure recovery over traditional models while retaining computational efficiency.
Lean Finder V2: Semantic Search for Mathlib That Disambiguates User Queries
AI summaryLean Finder V2 disambiguates short Mathlib queries using adversarial training and query rewriting, achieving 11 % better retrieval than previous engines on the MathlibQR benchmark and improving downstream theorem‑proving success when paired with Claude Code.
Reverse-KL Policy Distillation: Statistical Guarantees and Limits
AI summaryReverse-KL distillation’s sample complexity depends on feedback: full teacher distributions give optimal minimax rates, while only student-action probabilities require more data. Offline distillation achieves the same rates regardless of feedback, revealing a statistical advantage over online methods.
Bifurcations from Weight Space: JEPAWG on Parametric Dynamical Systems
AI summaryThe paper trains neural forecasters on grids of dynamical systems, forming a weight‑parameter zoo, and applies JEPAWG to embed these weights. It shows JEPAWG can forecast, locate bifurcations (Hopf, period‑doubling, Feigenbaum constant, global), and reveal phase transitions.
Training Mixed-Monotone Neural Networks for Verification Without Set Propagation
AI summaryA mixed-monotone neural network architecture splits the model into two monotone subnetworks, enabling tight output enclosures by two forward passes of opposite interval corners. Training directly minimizes enclosure size, achieving ≤0.5% of a proven lower bound while matching prediction accuracy.
Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?
AI summaryMeasures how much economic value (net dollars and time saved) AI delivers per joule across 220 tasks, finding that model and hardware choices affect value/energy tradeoffs; Gemma 4 31B achieves 2.5× higher net-value‑per‑joule than GLM 5.2, while Qwen3.6‑27B’s performance varies four‑fold across accelerators.
MOER: Mix-Policy Optimization under Expert Prefix Guidance with Selective Rollback
AI summaryMOER blends off‑policy expert prefixes with live exploration by progressively shortening the guided chain and rolling back failures to the original expert trace. It achieves top scores on reasoning benchmarks, outperforming baseline mix‑policy methods and maintaining stability across models and scales.
One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-Distillation
AI summaryOn-policy self-distillation uses a single teacher conditioned on one reference solution, which causes reference-induced target instability. Monte Carlo Privileged Distillation averages over multiple partial reference views, yielding more stable supervision and better performance on reasoning benchmarks.
Inherited, Not Invented: Tracing the Provenance of Wait in Reasoning Models
AI summaryThis paper traces the origin of the reasoning marker wait in language models. It shows that the marker is inherited from training data rather than invented by reinforcement learning, proving that its frequency does not reliably signal model reasoning.
Tracing Subliminal Learning Through Pretraining
AI summaryThe paper studies how large language models inherit hidden preferences from teachers through hard- and soft‑label distillation, showing transfer persists across architectures and checkpoints, is asymmetric, and can be partially explained by synthetic number patterns, with soft‑label distillation yielding more consistent but still unintended preference transfer.
Building Presentation Intelligence for High-Quality Slide Generation
AI summaryThis paper introduces a unified training framework for automated PowerPoint generation that combines validated data synthesis, mixed-teacher supervised fine-tuning, and verifiable reward reinforcement learning. Experiments show substantial performance gains on PresentBench, surpassing existing models in both text content and visual aesthetics.
Continuous-depth Deep Gaussian Processes
AI summaryThis paper introduces continuous-depth deep Gaussian processes, replacing discrete layer stacks with stochastic differential equations governed by Gaussian process priors. The proposed models improve optimization stability and regression performance across various benchmarks compared to traditional discrete deep Gaussian processes.
Equilibria and Learning in Competitive Market Making
We propose a market-making game played among $N \ge 2$ strategic market makers and a single non-strategic representative trader. In a continuous quote space, standard finite-game arguments do not apply, while non-convex and discontinuous payoffs rule out classical continuous…
Parameter Learning in Active Inference via Variational Message Passing
AI summaryThis paper extends active inference planning during parameter learning by using factor-specific parameter priors and sum-product message passing. The resulting method preserves information-gain corrections that standard mean-field approximations destroy, enabling agents to successfully learn models and navigate safely in grid POMDP environments.
Same molecule, Different Knowledge: Cross-Identifier Binding Failures in LLMs For chemistry
AI summaryThe paper introduces a benchmark showing that the same molecule’s different identifiers (names, IUPAC, SMILES) expose distinct knowledge in LLMs, a phenomenon called cross‑identifier binding failure. Binding Self‑Distillation is proposed to transfer knowledge from names to SMILES, enabling direct factual retrieval.
MIG-COW: Transferable Adversarial Attacks on Deepfake Detectors via Gradient Decomposition
Despite recent advances, deepfake detectors remain vulnerable to adversarial examples, particularly in diverse, real-world settings. We propose MIG-COW, a novel adversarial attack framework that generates highly generalizable and visually imperceptible adversarial examples. By…
Bridging Number-Theoretic Primality Testing And Algorithmic Optimization: An Adaptive Cryptographic Framework
The security of modern public-key cryptographic systems, such as RSA, Diffie-Hellman, and ECC, depends entirely on the computational intractability of the integer factorization and discrete logarithm problems. For these asymmetric algorithms to operate, a core requirement is…
Vision Transformers for Estimating Irradiance Using Data Scarce Sky Images
Accurate estimation of diffuse horizontal irradiance (DHI) is critical for optimising photovoltaic system performance and energy forecasting yet remains challenging in regions lacking comprehensive ground-based instrumentation. Recent advancements using Vision Transformers…
PAIR: Preference-Aligned Intent Retrieval for Premium People Recommendation at LinkedIn
General-purpose embeddings trained on profile similarity often misrank candidates in professional networks, where relevance is inherently complementary rather than symmetric. For example, job seekers usually engage with recruiters rather than peers with similar credentials.…
GateFormer: Forecast-Space Decomposition for Intra-Hour Solar Irradiance Prediction from Sky Images
Sub-hourly dispatch, balancing, and storage control increasingly depend on forecasts that anticipate rapid variability in solar generation. Accurate sky-image-based irradiance forecasting requires a combination of visual cloud dynamics with recent irradiance state while adapting…
Random Projections for Spectral Algorithms: Optimal Rates in Sobolev Norms
Random projections (RP) offer an effective method for reducing computational and storage overhead while preserving the geometric structure of data. However, existing studies have primarily focused on the optimal generalization performance of specific kernel-regularized…
Random Initialization as a Baseline for Time Series Foundation Models
Time series foundation models (TSFMs) use self-supervised pretraining to learn reusable representations for downstream tasks such as multivariate time series classification (MTSC). We examine how much pretrained weights contribute beyond the architecture itself by comparing…
OpenAttention: One Percent Is Enough with Tail Proxies
Sparse attention reads only a small subset of the key-value cache, and how to treat the unread tail remains a key challenge. Discarding the tail biases the output, and sampling it as previous approaches do fails at reads of 5% and below, where the tail likely holds heavy hitters…
Ground-Truth-First Benchmarking of Reinforcement Learning via Converse Optimality: From Smooth, Fully Observed Systems to Non-Smooth or Partially Observed Tasks
Benchmarking in reinforcement learning is usually relative: algorithms are ranked by returns on environments with unknown optima. Distance to optimality remains unobservable, while conclusions may depend on reward design, stochasticity, hyperparameters, environment selection,…
Doubly Robust Distributionally Robust Offline Contextual Pricing
Offline contextual pricing often relies on logged observational data, but faces challenges from distributional shifts between training and deployment environments. Distributionally robust optimization (DRO) provides a principled approach to off-policy evaluation and learning…
ReDence: A Hybrid Agentic and Decision-Model Framework for Rapid, Evidence-Grounded Detection of Modern Social Engineering Scams
AI summaryReDence is a hybrid framework that uses an investigative LLM (Qwen3.5-4B) to plan and interpret evidence while a specialized model (Tev1) makes concise decisions about browser actions and scam detection. It attains 96 % accuracy across 12 SE scam categories, is four times faster than pure agentic models, and produces structured evidence‑grounded reports for analyst audit.
Energy and Latency First: How Deployment Constraints Reverse Time-Series Model Selection
AI models, and time-series models in particular are typically ranked by accuracy and deployed under constraints absent from that ranking. In this paper, we study a real setting where these constraints are explicit and measurable: handover pre- diction in cellular networks, where…
Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters
Many language-model pipelines end in a closed-set decision, such as selecting a tool. Prompting a large language model for it is slow and order-sensitive. We ask how small an encoder for such decisions can be and whether it should read the options jointly or separately. In a…
Diagnostic Loss: Asymmetric Supervision for Radiology Report Generation
Automated radiology report generation requires models to do more than produce fluent clinical narratives: they must explicitly state supported diagnoses while avoiding unsupported ones. However, existing medical vision-language models (MedVLMs) are typically trained with…
EEG-Captioner: Grounding EEG-to-Language Generation in Fundamental Signal Properties
Electroencephalography (EEG)-to-language models offer a promising approach for translating complex neural signals into human-readable descriptions. However, generating plausible language does not necessarily demonstrate that a model has learned to interpret the underlying EEG…
XALIGN: A Human-centric Visual Analytics Approach to Understanding AI Explanations
Explainable AI (XAI) methods are usually evaluated for technical properties such as faithfulness and robustness. In high-stakes application domains, experts also need to check model predictions against established knowledge. Different explanation methods often disagree, and this…
Mostly Reading: What Open-Book Scores Measure in Document-Level LoRA Adapters
AI summaryTrains a LoRA adapter per Wikipedia article on Llama‑3.1‑8B; finds open‑book scoring largely reflects a generic “reading” capability rather than document‑specific knowledge. Specialists improve cloze recall only ~0.094 and their updates are document‑aligned, but merging adapters recovers most open‑book advantage (≈88 %) yet only ~12–16 % of cloze gain.
NN-Based Transistor Modeling for Ultra-Fast SPICE Simulation and Standard Cell Library Characterization
Neural network (NN)-based transistor models have proven to be a promising solution to accelerate device modeling. Although these models demonstrated remarkable speedup in circuit simulations, they were not rigorously tested in complex tasks within standard EDA tool flows for…
Federated Differentially Private Optimization via Low-Rank Preconditioning
We consider differentially private (DP) optimization in the federated setting, where a DP estimate of the feature covariance matrix serves as a Hessian approximation for generalized linear models. Existing central DP methods for covariance estimation have communication cost that…
When Forgotten Facts Return: Relearning-Aware Unlearning of Knowledge Graph Embeddings via Tangent Kernels
Approximate unlearning answers a deletion request on a knowledge-graph embedding by editing entity embeddings instead of retraining from scratch. These embeddings keep being updated as new facts arrive. Under this ordinary continued training, forgotten facts recover more than in…
nnInteractive Agent: General-Purpose AI Agents as 3D Image Annotators
Interactive 3D segmentation models such as nnInteractive turn a few spatial prompts (points, scribbles, boxes) into accurate volumetric masks, but a human still does the annotation work: finding each structure, placing prompts, checking the result and correcting it.…
Transformers with CoT Discover Heavy-Ball Acceleration Through Training
AI summaryTransformers trained with chain‑of‑thought supervision first learn to imitate gradient descent for linear regression weight prediction. Additional trajectory supervision using true weights causes the model to discover the Heavy‑Ball accelerated method, improving convergence without a Heavy‑Ball teacher.
DRAFT: Decoding and Reconstructing 3D Room Layouts from fMRI and Camera Trajectories
While functional magnetic resonance imaging (fMRI) decoders reconstruct images and 3D objects, recovering the metric geometry of rooms from a moving viewpoint remains an open challenge. We introduce, to our knowledge for the first time, the task of reconstructing 3D layouts of…
What Comes Next and Why: Interpretable Next-Event Prediction with Neuro-Symbolic Rules
In real-world applications, being able to explain why a prediction is made is often as important - if not more so - than achieving high accuracy. However, existing methods for sequential event prediction typically prioritize accuracy at the expense of interpretability.…
Delta N-gram: Forward-Only Continual Writing into a Language Model's Native N-gram Memory
Our previous analysis found that a model's native hashed N-gram table is where continual updates disturb the base least, but its writes needed backpropagation and changed the pretrained table. We present Delta N-gram, a forward-only write path into that table. The pretrained…
Post-Audit Mirage: Identifiability Limits of Offline Agent Verification
Offline tests can fail to reveal whether an agent update will remain safe after deployment. We operationalize this post-audit mirage using matched safe and harmful lifecycles that expose identical verifier-visible evidence while an isolated scorer records opposite deployment…
Is High Local Accuracy Enough? Geometric Consensus with Support-Adaptive Statistical Personalization for Federated Learning
Real-world deployment of personalized federated learning (PFL) requires adapting to heterogeneous clients while preserving recognition performance beyond local distributions. However, high local accuracy can conceal failures on locally rare or absent classes. Effective…
Mixing Matters: Evidence Position Bias Across Sequence Mixers in Long-Context Question Answering
Long-context models often use answer-bearing evidence less effectively when it appears away from a prompt boundary. We compare this position bias across Transformer, state-space, and hybrid Mamba architectures by moving one gold document through a fixed set of nine distractors.…
GM-GReFEL: Geometry-Guided Spatiotemporal State-Space Learning for Robust Dynamic Facial Expression Recognition
Recognizing facial expressions in unconstrained video suffers from two coupled problems: self-attention grows quadratically in cost with clip length, and crowdsourced labels are noisy. We present GM-GReFEL, a spatiotemporal state-space network with a geometry-aware decision…
PriMa-Causa: Causal Prescriptive Maintenance
Automated manufacturing requires data-driven decisions under uncertainty and sparse observational records, and unplanned downtime is a dominant cost of getting them wrong. Prescriptive maintenance recommends a maintenance action against an impending failure rather than merely…
Prompt Sensitivity in Tool-Augmented Model Comparisons
Language models are compared on question-answering benchmarks that pose each question in one wording, and a reworded question can change which of two models scores higher. We ask whether giving the models tools makes such comparisons more stable. An accuracy gain cannot answer…
Towards Field-Deployable Wearable IMU Systems for Occupational Activity Recognition
Wearable inertial measurement unit (IMU) systems for construction worker human activity recognition (HAR) routinely report aggregate accuracies above 95 %, rarely under genuine subject-independent (leave-one-subject-out (LOSO)) evaluation, masking the per-individual variance…
Prediction Uncertainty Controlled Computer Model Calibration
Computer-model calibration is usually undertaken in service of prediction, yet neither Bayesian Kennedy-O'Hagan (KOH) calibration nor amortized simulation-based inference is built to control the predictive error of the calibrated simulator. When parameters are weakly identified…
Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation
Humanoid loco-manipulation in dynamic scenes is non-Markovian: cues vanish as objects change or are hidden, or leave the view as the robot turns and walks, yet whole-body vision-language-action (VLA) policies see only the present. We present WholeBodyMem, an object-guided…
Prompting Strategy Effects Depend on the Judge: A Multilingual Study of LLM-as-a-Judge Robustness
LLM-as-a-Judge evaluation is increasingly applied to multilingual tasks, yet prompting strategies such as Chain-of-Thought and Role Conditioning are recommended based solely on English benchmarks. Whether their effectiveness transfers across judge models and languages remains…
Track finding and particle identification performance with particle physics foundation models
Modern high energy nuclear and particle physics experiments produce large amounts of sparse data that must be reduced significantly in size before performing physics analysis. Recently, there has been an effort to evaluate Foundation Models as a potential tool for performing…
RaSFNet: Ray-aligned and Scan-aware Fusion Network for LiDAR Semantic Segmentation
LiDAR point cloud semantic segmentation is a fundamental component of 3D perception for autonomous driving and robotics. Projection-based methods map 3D point clouds into 2D representations for feature extraction using 2D networks. However, aggregating multiple points during…
When More Metadata Misleads: Probing Field-Level Grounding in LLM-Generated Museum Descriptions
Metadata-conditioned generation is often framed as a data-to-text task in which additional input fields should improve grounding. Museum description generation challenges this assumption because catalogue records are heterogeneous, partially visual, and unevenly documented…
Inside-JEPA: From Prediction to Relational Grounding in Latent World Models
Joint-embedding predictive architectures (JEPAs) learn world models in latent space, enabling efficient planning but leaving their states opaque. We introduce Inside-JEPA, which places relational channels for ordered entity pairs inside the JEPA latent state and shows that an…
Evolving Federated Density Peak Clustering with Shuffler-Based Anonymization
AI summaryEvolving Federated Density Peak Clustering (eFDPC) uses one-shot data-stream clustering with a shuffler to anonymize client contributions, transmitting only density peaks. It outperforms baselines on synthetic and real datasets, handling high-dimensional, non-IID data efficiently.
Local Discrimination of Binary Classifiers via Localized AUC Functional
Risk prediction models are frequently evaluated with AUC, which means that during evaluation we are blind to performance heterogeneity, which could lead to bad decision-making in clinics. In this work, we aim to change that by identifying patient cohorts for whom the risk…
Muting Tokens: Exploring Token Sparsification for Efficient Audio Transformers
AI summaryThe paper benchmarks token sparsification for audio transformers, testing pruning and merging methods on various datasets and sizes. It proposes a frequency‑aware sparsifier that outperforms attention‑only approaches, showing that spectral cues and spectrogram intensity improve efficiency without sacrificing performance.
QuaNAS: Uncertainty-Aware Reuse of Earlier Searches for Query-Efficient Neural Architecture Search
Neural architecture search (NAS) on a new dataset often affords only tens of short training runs, so what the search knows before its first evaluation decides its efficiency. Yet most NAS methods start from scratch, guided at best by zero-cost proxies that are unreliable across…
Can Entanglement-Induced Feature-Map Designs Improve Projected Quantum Kernels?
AI summaryProjected quantum kernels were evaluated with and without classical RBF hyperparameter tuning, revealing that extensive tuning masks their true capability. Introducing entanglement‑based feature maps markedly improves the native inductive bias, yielding stronger PQK performance driven by quantum architecture rather than tuning.
Deep ReLU Networks Can Fail the Best Approximation Property for Continuous Target Functions
We construct a continuous function on $[-1,1]^2$ for which the infimum of the $L^2$-approximation error over the class of ReLU networks with two hidden layers and architecture $(2,2,1,1)$ is not attained. This counterexample shows that deep ReLU networks do not satisfy the best…
Learning Robot Behavior Priors from Non-Robotic Synthetic Data
AI summaryThis paper explores pretraining robot action prediction models on procedurally generated, non-robotic synthetic data. It demonstrates that transformers pretrained purely on mathematical control abstractions achieve non-trivial zero-shot performance on unseen physical robot tasks via in-context learning.
Beyond Accuracy: Disentangling Decomposition from Typed-State Quality Assurance in Multi-Agent LLM Pipelines
Multi-agent LLM pipelines are often credited with gains in software engineering tasks, but it is rarely clear whether a gain stems from the typed contract between agents or from the specialized prompts that decomposition enables. We study this attribution problem in Requirements…
DISC: Domain-Integrated Similarity Cost for Time-Series Dataset
AI summaryDISC learns to combine multiple time‑series similarity metrics, weighting them per task, and shows higher correlation with downstream forecasting errors than any single metric. The method calibrates with transfer observations, then predicts similarity for new source‑target pairs using only metric values.
Data-Driven Vision-Language Models for Remote Sensing: A Survey
During the deep learning era, innovations in remote sensing (RS) vision models primarily focused on optimizing network architectures for specific tasks and conducting end-to-end training. However, with the advent of the Vision-Language Models (VLMs) era, constrained by…
The Bayesian Learning Rule Beyond KL Geometry
The Bayesian learning rule (BLR) exploits the relation between Bayesian updating and the Kullback-Leibler divergence (KLD) to derive efficient additive updates for exponential family variational posteriors. The KL geometry is powerful but restrictive: combined with posterior…
Extreme Compression of Per-Layer Embeddings via Cross-Layer Latent Vector Quantization
AI summaryThe paper introduces OPTIC, a token‑shared quantization scheme that compresses per‑layer embeddings 100–1000× while preserving 84–98% of model performance; it uses one code per token and layer‑specific decoders, achieving 30% GPU‑memory savings with negligible slowdown.
Identifiability from Stationary Covariance in Linear SDEs with Reciprocal Interactions and Latent Variables
Continuous-time stochastic differential equations (SDEs) provide a natural framework for modeling interacting dynamical systems. A central causal question is whether causal effects among observed variables can be identified from observational data drawn from the stationary…
The Value of System 1 Models in Medicine
Many clinical decisions are a choice among a few fixed options, taken quickly and in large numbers. Such decisions also need an uncertainty estimate to know when to defer to a clinician. Generative language models answer slowly in free text and state their confidence poorly. In…
Bridging the Data Gap: Source-Informed Task-Space Transport for Few-Shot Diffusion
Many deployments require a model to move rapidly from a data-rich source to a related target observed only sparsely. This is especially consequential where target data are difficult to collect, as in healthcare settings with persistent population or acquisition imbalances.…
AgenticEvolve: Evolutionary Program Search with Persistent Language Model Agents that Design and Run Experiments
AI summaryAgenticEvolve uses persistent language‑model agents that design and run experiments on candidate programs, improving evolutionary search. It outperforms baseline methods on design tasks and equation discovery, achieving higher quality with fewer evaluations and similar cost.
Deep Learning for Random Objects in Metric Spaces: Beyond Manifold Assumptions
Deep learning (DL) methods have substantially improved predictive performance across a wide range of downstream prediction tasks, enabling new discoveries in multiple domains. However, principled DL methods for non-Euclidean random objects remain relatively scarce, despite the…
Harmful Responses Are Jailbreak Templates: Transferring Autoregressive Jailbreaks to Diffusion Language Models
Diffusion large language models (dLLMs) enable flexible masked generation through bidirectional context modeling, but this capability also introduces a distinct jailbreak attack surface. We find that a harmful response to a given query naturally serves as an effective jailbreak…
Metric Conformal Prediction via the Conditional Expected Local Radius
AI summaryThis paper defines the conditional expected local radius (CELR) and uses it to form a conformal prediction set for non‑Euclidean responses. It proves uniform consistency, asymptotic conditional validity, and shows improved computation and coverage over existing methods.
Graph-Recurrent Learning of Emergent Collective Dynamics
AI summaryA graph‑recurrent neural network learns the time evolution of collective order from raw trajectories without explicit microscopic equations, using gated memory and message passing. Applied to the Holstein model, it reproduces charge‑density‑wave spatial and temporal correlations and coarsening scaling.
AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents
Many LLM agents maintain external persistent memory and derive additional state from past interactions, such as summaries. Deleting an original memory record therefore need not remove information carried by its descendants. We introduce AfterDeleteBench, a controlled behavioral…
Convex Equilibrium Potentials for Monge--Ampère: Curvature Bounds and A Posteriori Evaluation
Equilibrium neural networks define their hidden state as the fixed point of a contraction, so their effective depth is set at evaluation time rather than by the architecture. Using them for fully nonlinear PDEs raises two issues that explicit networks avoid: second derivatives…
HERD: Harness-Evolving Recursive Distillation for Continual Learning
Continual learning requires a model to keep acquiring new skills and knowledge without compromising its existing capabilities. Parametric methods suffer from catastrophic forgetting and loss of plasticity. Non-parametric methods, such as prompts, skills, and harnesses, leverage…
SQuAT: Spherical Lattice Vector Quantization-Aware Training
Vector quantization (VQ), in particular lattice-based methods, offer strong rate-distortion trade-offs for low-bit large language models (LLMs), yet their behavior under quantization-aware training (VQ-QAT) remains poorly understood. We scale lattice VQ-QAT to billion-parameter…
Post-Hoc Feature Selection Layer for Neural Networks Interpretability
AI summaryPost-hoc Feature Selection Layer fine‑tunes a lightweight weighting module on frozen pre‑trained tabular models, highlighting the most relevant features without altering the original network. Experiments show it identifies key features while preserving accuracy, outperforming Integrated Gradients, SHAP, and others.
How Depth Growth and Looping Shape In-Context Learning in Language Models
AI summaryThis paper investigates how parameter reuse in looped and depth-grown language models affects in-context learning. The results show that these architectures develop a stronger inductive bias toward context, acquiring and retaining in-context learning earlier and more reliably than standard models by propagating induction heads across layers.
Predictable performance is suboptimal, and what it means for model routing
AI summaryThe paper proves that a performance predictor improves routing only when the base model is suboptimal or the predictor adds significant capacity, showing log‑loss transfer from predictor to model. Experiments confirm larger gains for weaker models, diminishing after fine‑tuning.
Dimension-free regimes for unadjusted Langevin in ridge-separable models: A subspace robust Wasserstein analysis
AI summaryULA’s bias is analyzed in subspace robust Wasserstein distance, revealing a bell-shaped mixing time governed by data rank N instead of dimension d. In overparameterized and large‑sample regimes, convergence is dimension‑free; near the proportional regime dimension‑free guarantees fail.
Drift Says When, Budget Says How Much: Rate-Controlled KV Re-selection for Long-Form Multimodal Generation
AI summaryThe paper proposes Rate‑Controlled Refresh (RCR), a dynamic KV re‑selection method that lets attention drift decide when and where to refresh key‑value pairs while a per‑sample controller limits refresh frequency. RCR beats the fixed‑threshold method InfiniGen by 1.6–3.8 ROUGE‑L points on most long‑form multimodal tasks, matching per‑step retrieval and using less GPU memory.
Conservative Meta-Agent Search: Architecture Adaptation under Non-Stationarity
AI summaryConservative Meta‑Agent Search (CMAS) frames architecture search as incumbent replacement with switching costs and risk constraints, using archived models, typed graph edits, and paired shadow replay to promote only high‑confidence edits. On a regime‑switching finance benchmark, CMAS reduces surrogate dynamic‑oracle regret from 0.116 to 0.104 and raises held‑out utility from 0.44 to 0.48 versus the best adaptive baseline.
DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels
When a user says "Keep it short", how short should the LLM's response be? More generally, users and applications require different levels of instruction adherence, but prompting does not reliably provide fine-grained control over them. We therefore study "steerable language…
Multi-Token Prediction and Alternative Tokenization Improve Learning Latent Multi-Hop Graph Search at Scale
Given an unseen graph, a start node, $s$, and an end node, $t$, can a language model generate the shortest path from $s$ to $t$? Prior work has found that models fail to learn this task as the graphs scale. We show that this failure has two causes: the supervision signal and the…
RheOPT: Few-Shot Non-Newtonian Flow Prediction via Newtonian Pretraining of RheOFormer
Despite advances in numerical methods, non-Newtonian flow simulations remain computationally expensive because velocity and stress must be advanced together through rate- and history-dependent constitutive dynamics. Neural-operator surrogates reduce inference cost after…
MoG-VLM: Efficient Video Anomaly Detection through Motion-Guided Vision-Language Models
Vision-Language Models (VLMs) have shown strong potential for video anomaly detection, but dense frame-level inference makes them impractical for real-time deployment. Existing acceleration commonly relies on uniform sampling, which ignores content dynamics and can either waste…
Evaluating the Applicability of LLM Inference Optimizations on Apple Silicon
Apple Silicon is among the most widely used platforms for local LLM inference, yet nearly all inference optimizations are designed and validated on server-class NVIDIA GPUs under batched serving workloads. Which optimizations retain their benefits when moved to Apple Silicon,…
Training Small Language Models for Transferable Tool Selection
AI summaryThey train a small Qwen3-8B with reinforcement learning to decompose complex instructions and select the right tool from thousands, outperforming a 402‑B Llama model on ToolBench, and the policy transfers to other retrievers and domains.
SABLE: Scalable Affine Bayesian Latent-space Exploration
Complex objects such as engineering designs and proteins are often difficult to optimise because their design spaces exhibit high-dimensional structure and objective evaluations are expensive. Generative models provide continuous latent representations, but their latent spaces…
Structure-preserving reduced-order models of thin panels in supersonic and hypersonic flow
AI summaryStructure‑preserving reduced‑order models are built for a clamped thin panel at Mach 1.92 and a cantilever plate at Mach 5.85; the learned models achieve 2.2×10⁻¹⁵ residuals, match shock‑pressure ratios, and reproduce measured flutter frequencies within 1 % accuracy.
Piano2Pop: Expanding Piano Demos into Pop Productions Without Aligned Supervision
We introduce Piano2Pop, a music-expansion task that transforms piano MIDI into instrumental pop productions. Enabled by recent advances in music generation and polyphonic music transcription, our pseudo-supervised approach conditions a pretrained flow-matching model on piano…
Nonparametric autoregressive estimation of probability distributions with parallel gradient boosting
Nonparametric estimation of joint probability distributions is a core problem in statistics, which is notoriously challenging with high-dimensional data. Such challenges also arise with conditional distribution estimation and have led to flexible algorithms that aim to tackle…
Direct Inference on Admissible Sets (Diona)
AI summaryDiona is a local regression method that replaces global training with on-demand constrained optimization to estimate function values, derivatives, and admissible bounds. It achieves orders of magnitude faster prediction times than Gaussian processes and neural networks while maintaining reliable uncertainty quantification.
LACToS: Local Anomaly Detection via Conditional Typicality of Superpixels
AI summaryLACToS identifies anomalous image regions by measuring statistical typicality of superpixel groups, using an efficient proxy for typicality estimation. Across medical, face, and X-ray security benchmarks, it outperforms previous methods, especially when the inlier distribution is multimodal.
PYTHIA, a transferable machine learning framework for simultaneous defect and adsorption site optimization
AI summaryPYTHIA uses a frozen foundational model with transfer learning, a transformer alignment layer, and Gaussian process uncertainty to build a surrogate for Bayesian optimization, enabling efficient defect and adsorption site optimization on new materials with limited data. The method identifies optimal sites accurately and quickly.
Same Luck, Different Choice: Twin-World Credit Assignment for Agentic Reinforcement Learning
AI summaryTWICE uses twin-world rollouts that share the same Gumbel noise as the factual episode, providing unbiased credit estimation with reduced variance (∝ 1−ρ). On QA tasks it boosts GRPO’s exact match by 5.3/4.5 points, beats tree-search, and cuts auxiliary tokens by 68%.
Shift-aware score transport for privacy auditing without retraining
AI summaryThe paper develops a shift‑aware method that transports membership‑inference scores from out‑of‑distribution non‑members to the target distribution, then conservatively corrects the transport error. It yields more stable privacy audits and tighter lower bounds than baselines.
Fine-Grained Analysis of Catastrophic Forgetting in Continual Learning via Algorithmic Stability
Understanding catastrophic forgetting, in which learning new tasks degrades performance on previously learned tasks, remains a fundamental challenge in continual learning (CL). Existing theoretical analyses often rely on specific model or task structures and offer limited…
Diffusion Language Models for Perplexity-Aware Black-Box Attacks
\emph{Continuous-space} attacks find prompts that trigger malicious behaviors, such as overthinking, in black-box LLMs by sampling or optimizing over a low-dimensional Gaussian latent space and decoding it to text through a \emph{random mapping} and nearest-token lookup. Because…
When to Follow the Teacher: Advantage-Based On-Policy Distillation
Online Policy Distillation (OPD) often outperforms offline methods such as Supervised Fine-Tuning (SFT) by training on trajectories sampled directly from the student. However, standard OPD trains the student to follow the teacher at every state of the trajectory, assuming that…
On the Token Collapse Phenomenon in Textual Graph Soft Prompting
Recent methods for injecting graph structure into LLM prompts include graph soft prompting, in which a graph neural network (GNN) learns graph-specific tokens that capture relational information among entities in textual prompts. G-Retriever combines graph soft prompts with a…
Rethinking Sequence Compression with Test-time Regression
AI summaryTTR‑Compressor compresses token sequences by solving a test‑time regression whose fitted coefficients serve as the compressed representation, computed in closed form. It outperforms pooling or projection baselines: improves RULER block‑wise attention and reduces visual tokens by 44.4% while matching MLP merger accuracy.
Action Learning for Scientific Simulations
Understanding and controlling how systems evolve is central to science and engineering, where many problems require reasoning about complete trajectories. Existing approaches typically learn local dynamics or generate trajectories, but objectives constructed from individual…
Would an LLM Match into a Residency? Uncertainty, Reasoning and Image Use of Open Language Models on the Spanish MIR Examination
Accuracy on medical examinations is common evidence offered for the clinical use of language models, yet a health system that intends to run an open model on its own hardware needs to know considerably more: whether the model recognizes when it is likely to be wrong, whether…
Beyond Best-of-$k$: Prior-Consistent DAPS for Ambiguous Nonlinear Inverse Problems
Diffusion models have become strong priors for solving inverse problems. However, nonlinear inverse problems such as phase retrieval remain challenging when multiple measurement-consistent solutions are distinguishable only by the prior; we call such measurements ambiguous.…
CRANE: Post-Hoc Counterfactual Retrieval for Multimodal Classifiers
AI summaryCRANE retrieves real training examples from the target class as multimodal counterfactual explanations, using a nearest‑neighbour, model‑agnostic approach. It shows that intersecting per‑modality neighbours (MC‑R) outperforms other methods on three of four datasets.
Efficient and Optimal Learning of Discrete Distributions with Person-Level Privacy
We study learning a discrete distribution under person-level differential privacy in the batched setting. Given $n$ users, each contributing $m$ i.i.d. samples from an unknown distribution $\mu$ over $[k]$, we seek an estimate of $\mu$ with total-variation error $\alpha$. We…
OP-Print: Tracing Membership Imprints in On-Policy Self-Distillation
On-policy distillation has been widely adopted for alleviating the distribution mismatch of off-policy distillation. Self-distillation further removes the need for an external teacher by using the student's own outputs, enhanced with extra context or demonstrations, as…
Questioning Instrumental Convergence in RLVR: Checkpoint Evidence from OLMo-3
Instrumental convergence is often invoked as a general reason to expect capable agents to acquire or express power-seeking subgoals: self-preservation, resource acquisition, deception, or resistance to oversight which poses an important threat to the society. For LLMs,…
BeNice! Closing Looping's Inference-Memory Gap via KV-Sharing
Looped architectures scale quality while keeping parameter count fixed by repeatedly passing activations through the same block of layers. During inference, each recurrence generates independent KV-caches, causing memory to grow linearly. To address this challenge, prior work…
Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation
Chain-of-thought (CoT) prompting elicits step-by-step rationales that appear to explain model predictions, but there is no guarantee these steps causally drive the final answer. In standard autoregressive generation, dense attention allows each reasoning step to attend to the…
A Fix for Softmax and Cross-Entropy
Standard implementations of softmax, log-softmax, and cross-entropy have two numerical bugs. The first bug, Softmax Collapse, rounds the target logit's gradient to zero if that one logit is much larger than the others, resulting in gradients that no longer sum to zero. The…
EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents under Missing Evidence
Tool-using language model agents rely on environmental facts to decide whether an action is permitted. When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and every returned statement is true. We…
METTA: Memory-Enhanced Test-Time Adaptation for Meta-Reinforcement Learning
AI summaryMETTA splits adaptation into fine‑grained within‑episode causal memory (linear cost) and compact episode summaries for cross‑episode inference, enabling fast test‑time adaptation with low memory. It reaches >50% success on Meta‑World held‑out tests and outperforms transformer baselines on long‑context sparse‑reward tasks.
Cost-Aware Abstention for LLM-Based IT Ticket Classification
AI summaryThis paper evaluates cost-aware abstention policies for IT ticket classification by selecting classification thresholds on calibration data under explicit error costs. The method reduces routing costs by 38.1% compared to baseline approaches on a standard support ticket corpus.
SpectralCanary: Embarrassingly Simple Low-Frequency Traces for T2I Dataset-Use Auditing
AI summarySpectralCanary embeds a low‑frequency mask into every image of a dataset to leave a detectable audit trace in T2I models without altering captions or requiring per‑image tuning. It achieves ~91% per‑generation TPR@$1\%$ FPR and ~95% model‑level accuracy while preserving image fidelity across diverse domains.
Reliable AI-Assisted Decisions via Selective Auditing
AI summaryThe paper introduces an online optimization framework for selectively auditing AI decisions to minimize human review while guaranteeing that unaudited error rates remain below a specified threshold. Experiments across LLM safety screening and data labeling demonstrate that the proposed algorithms maintain strict error control while significantly reducing auditing frequency compared to baselines.
Phase Transition from Random Features to Fully Trainable Networks
Constructing accurate and computationally tractable approximation spaces is a fundamental task in scientific computing. Partially trainable neural networks bridge fixed random features and fully adaptive representations. This paper quantifies a phase transition pattern in…
Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success
AI summarySensorimotor Alignment (S²) is a training‑free offline metric that quantifies how well a model’s sensory representation matches the geometry of physical motion. The paper shows S² strongly correlates with downstream robot task success and can identify promising vision‑language backbones before costly fine‑tuning.