Researchers proposed Collab-Solver, a novel multi-agent-based policy learning framework for mixed-integer linear programming (MILP). This framework enables collaborative policy optimization for multiple modules, leading to improved solving performance and generalization abilities. This development may be of interest to researchers and developers working on AI optimization techniques and their applications.

Researchers introduced Row-wise Matrix AdaGrad and Column-wise Matrix AdaGrad, two matrix-aware adaptive optimization methods that improve optimization stability and trainability for neural networks. These methods are derived from an Online Mirror Descent framework and can be used to train larger and deeper networks.

A research paper proposes Learning-to-Optimize (L2O) as a missing architectural layer for AI-native networks, enabling systematic generation, transfer, and exploitation of optimization knowledge by AI models. The paper introduces a four-stage L2O workflow and demonstrates its application to a relay-selection problem in NR-V2X communication systems.

A new open-source neural network framework is presented, implemented from scratch without relying on pre-built deep learning modules. This framework serves as a pedagogical tool for understanding neural network mechanics and can be used as a baseline for educational purposes and future research.

A research paper analyzes the resource dynamics of AI agents, highlighting inefficiencies and opportunities for optimization. The study investigates the latency and resource usage of agents processing multiple requests and tasks concurrently, revealing task-dependent bottlenecks and proposing new optimization techniques to improve agent performance.

Researchers introduce BATON, a dual-axis policy optimization framework for LLM agents, combining Bayesian Feedback Attribution and Trajectory Mass Normalization. Experiments show that both axes provide independent gains and their combination achieves the strongest performance across model scales.

A research paper on optimizing key-value cache compression in reinforcement learning post-training to reduce memory usage, which is crucial for large language models.

A new self-distillation framework, VISTA, is introduced to address Trajectory Deviation in deep learning models. This issue occurs when models converge to suboptimal solutions despite high validation accuracy. VISTA uses a validation-informed Marginal Coverage score to identify expert anchors that retain specialized competence and integrates them online during training to regularize the loss landscape and preserve mastered knowledge.

This tutorial bridges the gap between operations research and management science (OR/MS) and deep learning for sequential decision making under uncertainty. It presents a framework for integrating OR/MS approaches with neural architectures, highlighting the complementary strengths of both fields.

Researchers proposed a reparameterization technique to optimize speech models during training, enabling joint optimization of performance and computational complexity using SGD-based methods. This method dynamically optimizes model size for a target performance-complexity trade-off without relying on heuristic criteria.

Visual Perception Engine (VPEngine) is a modular framework for efficient GPU usage in robotic vision tasks, enabling parallel execution of multiple machine learning models. It uses a shared foundation model backbone and task-specific heads, achieving up to 3x speedup and efficient GPU utilization.

Researchers proposed HOPE (Higher-Order Pruning of Experts), a method for reducing the parameter count in mixture-of-experts language models by preserving cooperative expert structure. HOPE outperforms existing methods, especially at high pruning rates and on challenging tasks like agentic coding.

This academic paper compares and contrasts different types of evolutionary algorithms, including Baldwinian, Lamarckian, and Darwinian evolution, and presents empirical and theoretical results.

Ave is a framework for guiding agentic GPU optimization using data-flow invariants. It provides a tile-based Pythonic DSL that exposes hardware instructions and compiler policies while abstracting complex memory layouts. Ave uses tag functions, an SMT solver, and an in-context reinforcement learning planner to optimize GPU kernels and achieve performance comparable to hand-optimized libraries.

Researchers explore a novel approach to code self-distillation, using a deterministic, judge-free filter to optimize verification and increase the clean-launch rate of generated projects. The study shows significant improvements in out-of-family generalization and coverage, with the gate's precision being the key factor in determining the loop's effectiveness.

FlashVector is an AI agent that optimizes performance across the model serving stack, achieving up to 2x throughput increase and 1.98x latency speedup on a real-world deployment. It generalizes the single kernel optimization agent paradigm to heterogeneous technical stacks.