#machine-learning News & Analysis

Aligning Data-Driven Predictors with Allocation: A Decision-Focused Approach to Survival Analysis

AUGUSTE: Online-Learning dApp for Predictive URLLC Scheduling

Re-Evaluating Continual Learning with Few-Shot Adaptation

Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

Introduction to optimization methods for training SciML models

Physics-informed diffusion models in spectral space

New Microsoft tool lets devs spin up AI behavior tests using text descriptions

Microsoft has released Adaptive Spec-driven Scoring for Evaluation and Regression Testing (ASSERT), an open-source framework designed to help developers create and run AI behavior evaluations using natural language descriptions. This tool simplifies the process of testing AI systems by reducing the technical complexity required to set up comprehensive evaluation protocols.

U.S. Soccer is using AI to scout 70 million teenagers. The former consulting CEO running the federation calls it a ‘paradigm shift’ for the sport

U.S. Soccer is implementing AI technology to scout talent across 70 million teenagers, marking a significant shift in how the sport identifies and develops players. The federation's new consulting-led approach represents a modernization of traditional talent discovery methods ahead of hosting the World Cup.

Agents on a Tree: Pathwise Coordination for Multi-Objective Molecular Optimization

Researchers introduce ATOM, a multi-agent framework that treats molecular optimization as tree-structured search where specialized agents coordinate across different pathways rather than enforcing consensus. The method demonstrates improved performance on multi-objective molecular design benchmarks by maintaining diverse trade-offs and exploring multiple promising trajectories simultaneously.

CAST: Non-Privileged Clipped Asymmetric Self-Teaching with Advantage Flipping for GRPO

Researchers propose CAST, a new self-distillation method for reinforcement learning in large language models that improves upon existing approaches by using answer-free teacher scoring and bidirectional advantage flipping. The method addresses limitations in Group Relative Policy Optimization (GRPO) by providing denser token-level guidance while maintaining alignment with trajectory correctness, demonstrating improvements in mathematical reasoning tasks.

From Noise to Control: Parameterized Diffusion Policies

Researchers propose Parameterized Diffusion Policy (PDP), a machine learning framework that enables diffusion models to learn controllable behaviors through low-dimensional parameters mapped to a semantic behavior manifold. This approach transforms diffusion models from stochastic noise generators into precise policy control tools, allowing smooth interpolation between strategies and adaptation to novel constraints without retraining.

VESTA: Visual Exploration with Statistical Tool Agents

VESTA is a new AI framework that enhances vision-language models with dynamically generated statistical tools to automate scientific model fitting tasks. The system outperforms prior approaches by actively exploring data through adaptive tool creation rather than relying solely on iterative critique, with particular strength on complex, domain-specific modeling problems.

EnergyMamba: An Uncertainty-Aware Graph-Enhanced Selective State Space Model for Energy Consumption Prediction

Researchers introduce EnergyMamba, a machine learning framework that combines graph neural networks with state-space models to predict energy consumption while quantifying prediction uncertainty. The system achieves 5% accuracy improvement over existing methods by simultaneously modeling spatial grid relationships and temporal patterns, with enhanced reliability during abnormal conditions like extreme weather.

Efficient Test-time Inference for Generative Planning Models

Researchers introduce an optimized inference method for generative AI planning models that combines classical Open-Closed List search with learned generative and heuristic components. The approach demonstrates superior computational efficiency and solution quality compared to existing neurosymbolic and classical solvers across combinatorial planning domains.

Medication-Aware Financial Exploitation Detection for Alzheimer's Patients Using Edge-Aware Interaction Risk Modeling

Researchers propose a medication-aware AI framework that detects financial exploitation of Alzheimer's patients by combining transaction monitoring with medication adherence data. The interaction-aware model significantly improves detection of fraudulent transactions during periods of cognitive vulnerability, suggesting that clinical context enhances fraud detection accuracy beyond financial patterns alone.

Regularized Offline Policy Optimization with Posterior Hybrid Bayesian Belief

Researchers propose Posterior Hybrid Bayesian Belief (PhyB), a new method for offline reinforcement learning that efficiently manages uncertainty in policy optimization. The approach reformulates complex Bayesian objectives into tractable convex combinations of dynamics models, achieving state-of-the-art performance while providing theoretical guarantees for convergence.

Property Prediction of Stacked Bilayer Materials: A Multimodal Learning Approach

Researchers propose a multimodal machine learning approach to predict properties of stacked bilayer 2D materials, addressing a significant gap in AI-assisted materials discovery. This work aims to accelerate the design of novel materials with engineered functionality by modeling how different material layers interact when vertically integrated.

Can AI Review Improve Paper Drafting? An Empirical Study on 20 Computer Architecture Submissions

Researchers developed AI-Paper-Review, a tool that generates structured peer review feedback for academic papers using multiple AI reviewers, and conducted a case study on 20 computer architecture submissions to measure how well AI review aligns with human review. The study finds that AI review can identify significant portions of human-raised issues while also surfacing problems missed by human reviewers, raising important questions about AI's role in academic peer review without endorsing its use for formal publication decisions.

Deft Scheduling of Dynamic Cloud Workflows with Varying Deadlines via Mixture-of-Experts

Researchers introduce DEFT, a new deep reinforcement learning architecture using a mixture-of-experts approach to optimize cloud workflow scheduling with varying deadline constraints. The system uses a graph-adaptive gating mechanism to route scheduling decisions through specialized experts, demonstrating improved performance in reducing execution costs and deadline violations compared to existing DRL baselines.

"Skill issues'': data-centric optimization of lakehouse agents

Researchers present a data-centric optimization framework for AI coding agents operating on branching lakehouses, demonstrating that agent skills can be systematically improved through task-verifier pairs and sandboxed execution. The approach treats agent evaluation as state verification rather than output matching, achieving 31.9% accuracy improvements on preliminary tasks.

Advanced Mathematics Learning Behavior Prediction and Academic Early Warning Model Based on Multimodal Data Analysis

Researchers have developed an AI system using multimodal data analysis to predict at-risk mathematics students and provide early academic warnings. The framework combines knowledge graphs with temporal modeling to identify students struggling with complex concepts and enable timely interventions that improve learning outcomes.

SIRIUS-SQL: Anchoring Multi-Candidate Text-to-SQL in Execution Feedback

SIRIUS-SQL introduces a multi-candidate approach to Text-to-SQL generation that addresses redundancy, execution error classification, and selector limitations through difficulty-smoothing reinforcement learning, targeted repair mechanisms, and hybrid confidence-gated selection. The system achieves 75.88% accuracy on BIRD dev and 91.20% on SPIDER test, surpassing previous state-of-the-art multi-candidate systems.

Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence

Researchers present a category-theoretic framework for agentic AI systems that can revise their own representational structures during scientific discovery, rather than merely generating answers within fixed assumptions. The work demonstrates how self-revising discovery systems can be engineered for materials science through two instantiated systems: Builder/Breaker and CategoryScienceClaw.

Transferring Information Across Interventions in Causal Bayesian Optimization

Researchers present graph-coupled causal Bayesian optimization, a method that improves expensive system optimization by sharing information across related interventions through a causal kernel. The approach demonstrates logarithmic information gains and cleanly separates optimization, causal estimation, and intervention selection errors, with strongest performance when direct interventions are unavailable.

ReSkill: Reconciling Skill Creation with Policy Optimization in Agentic RL

Researchers introduce ReSkill, an RL-in-the-loop framework that improves how AI agents create and refine reusable skills during policy learning. The method synchronizes skill evolution with policy optimization, enabling agents to automatically develop, test, and prune strategies that generalize across tasks more effectively than existing approaches.