Ventricular tachycardia is a life-threatening arrhythmia. Pace-mapping identifies ablation targets. cAPM uses continual learning to transfer knowledge across VTs, reducing pacing sites needed. In silico, it achieved 81% localization accuracy with 4.5 sites vs. 38% with 13.7 for prior methods.

Self-play reinforcement learning can train driving policies without human data but often learns alien behaviors. A new method uses just 30 minutes of human demonstrations as a regularizer on top of a minimal safe reward, achieving human-compatible policies with 2500x less data than imitation learning, training in 15 hours on a single consumer GPU.

Proposes multivariate zero-inflated Gaussian (ZIG) distributions as sampling laws for EDAs, enabling joint optimization of sparsity patterns and active values without hand-crafted operators. The latent parameters are identifiable, and practical amortized inversion estimators are provided. On the Lunar Lander benchmark, ZIG-EDA outperforms dense Gaussian EDA, hand-crafted sparse EA, and ad-hoc sparse EDA in convergence speed and final returns while using few active parameters.

Building on a two-parameter Weibull framework, this paper analyzes why the Weibull weight-scale parameter λ grows, overshoots, and relaxes during AdamW training. A three-force decomposition (alignment, injection, decay) is derived, showing alignment dominates the rise phase (88-94%). Near saturation, alignment and decay balance. A spline displacement method recovers alignment from sparse checkpoints with ~92-94% accuracy. Peak λ varies with training-data coherence, suggesting a data-dependent component.

Information lattice learning (ILL) learns interpretable rules of a signal by alternately projecting the signal onto a partition lattice and lifting selected rules back. When the signal is a probability mass function, the probabilistic rules learned by ILL admit a natural probabilistic graphical model (PGM) interpretation. A partition induces a deterministic quotient variable, and a rule is the marginal law of that variable. A rule set is a collection of marginal constraints over interpretable abstractions. General lifting is the feasible family of all joint distributions satisfying constraints, while special lifting chooses a maximum-ignorance reconstruction via an L2 uniformity principle related to maximum entropy. Under Shannon-entropy lifting, constraints yield a log-linear factor grap…

This paper analyzes Med-DDPM's performance across three NVIDIA GPU generations, identifies inefficiencies, and demonstrates TF32 Tensor Core and channels-last optimizations achieving up to 100x improvement without quality loss.

A new paper introduces SPSD (Sentiment Preserving Semantic Distillation), an edge-based pipeline that compresses user prompts by removing social scaffolding before sending them to a cloud LLM, reducing input tokens by an average of 99.9 per call while maintaining response quality within a non-inferiority margin. The approach aims to cut cloud energy costs.

A novel framework called Gated Uncertainty-Aware Routing for Distillation (Guard) is proposed to address distributional misalignment and computational cost of Time-Series Foundation Models in scientific applications. It uses a Contextual Router and Uncertainty-Gated Temperature to distill knowledge from multiple pretrained models into lightweight forecasters. Evaluated on four climate-critical domains, Guard significantly reduces RMSE compared to a fixed-weight multi-teacher baseline, and misaligned teachers outperform globally superior models on 28.5% of hardest instances.

This paper introduces 'computational identifiability', a framework that replaces theoretical asymptotic assumptions with a finite computational search procedure for empirical estimators. Experiments show it works for small samples, ambiguous graphs, mixed data, and counterfactuals.

This paper proposes a prompt-based uncertainty decomposition method that separates action confidence from request uncertainty, enabling LLM agents to ask for clarification when task specifications are ambiguous. The authors introduce two new benchmarks with 50% underspecified tasks and evaluate against existing methods across five LLMs, showing significant F1 improvements.

This paper introduces the Integral Transform Network (ITNet), a unified architecture that generalizes convolution, self-attention, and recurrence through a learnable integral kernel. ITNet matches or exceeds specialized models on multiple benchmarks.

A new method enables LLMs to self-align with ethics using a conscience step and Direct Preference Optimization, without external judges. The technique counters emergent misalignment in scenarios like code hacking.

The paper proposes REVEAL++, a continuous formulation of phenotypic grouping in contrastive learning for vision-language alignment using retinal images and clinical risk narratives to predict Alzheimer's disease risk. It replaces hard group assignments with differentiable weights, enabling graded supervision and end-to-end learning. Evaluated on UK Biobank, it outperforms discrete methods.

A study comparing Qwen 2.5 7B and XGBoost on clinical prediction reveals that LLM verbalized confidence is epistemically vacuous, an inverse difficulty effect exists, few-shot and SHAP interventions improve accuracy, and a cross-model calibrator reduces calibration error.

DeXposure-Claw is a forecast-grounded agentic supervision system for decentralized finance risk, addressing the over-reading and high-stakes false alarms of general-purpose LLM agents. It uses a graph time-series foundation model to predict exposure networks, deterministic monitors and stress scenarios to generate alerts, and data-health gates to constrain escalation. DeXposure-Bench evaluates system decisions on six axes, including a regulator-aligned false-intervention rate. Experiments on five years of real data validate the system.

This paper proposes a dynamical system model for multi-agent LLM deliberation, where each agent has a hidden internal belief (anchor) that continuously pulls its opinion. The model explains a behavior forbidden by classical consensus rules: an agent's confidence can exceed the convex hull of initial beliefs. Experiments on three open-weight model families show that all anchors have similar influence but differ in location, and only when the anchor is far from initial opinions does deliberation escape the hull.

This paper presents a systematic experimental analysis of eight state-of-the-art diffusion language models (DLMs) across eight benchmarks spanning reasoning, coding, translation, knowledge, and structured problem solving, considering both generation quality and computational efficiency. It examines the impact of inference-time factors such as denoising steps, context length, block size, and parallel unmasking strategies, and finds that DLM behavior is strongly influenced by generation-time design choices, leading to distinct trade-offs between performance and efficiency. The study provides practical insights into the capabilities and deployment characteristics of contemporary DLMs.

A new study introduces a human-in-the-loop pipeline to measure how well undergraduate computer science programs align with curricular guidelines. Applied to CS2013 and CS2023, it found near-constant coverage (~50%) but a drop in cognitive depth achievement from 95% to 76%, reflecting raised expectations in the newer standard. Persistent gaps in parallel computing, programming languages, and systems fundamentals were identified.

A new paper proposes AgenticRei, a deontic policy framework to govern LLM-driven autonomous agents, addressing obligations, dispensations, and policy conflicts beyond current access-control engines.

Modal is all-in on speculative decoding, arguing it's the single most important inference optimization, delivering 2-3x speedups. They released state-of-the-art DFlash speculators for Qwen models, achieving 5-20% extra speedups, and explain the theory, simulation, and math behind the acceleration.