BW-QAMOO applies Bayesian optimization to select weight vectors for multi-objective QAOA, achieving a 33% reduction in scalarized solves while preserving Pareto quality; efficiency grows with more objectives as the weight space expands. The approach uses a Gaussian-process model and EHVI to balance exploration and exploitation, drawing weights from a Dirichlet distribution and starting from simplex corners; validation with Wilcoxon test confirms savings.
LambdaMix jointly trains low-rank adapters for multiple objectives, then uses a preference vector to control their mixture and reward weights at inference. Across tests, it outperforms the evaluated baselines on two task groups while using fewer training completions and adapter parameters than specialist policies.
Harmony Harness presents a multi-objective, compositional approach to evolving agent harnesses, using decomposition-based search and factorized operators to balance general mechanisms with benchmark-specific adaptations, achieving notable gains over baselines on two benchmarks.
The paper argues that perfect value alignment for autonomous robots is unattainable and that misalignment risks will persist. It proposes a shift from maximizing alignment to ensuring misalignment remains acceptable, defines normative requirements, and introduces a framework of alignment mechanisms to meet these requirements.
Translated title. Original: ASCII.jp：時給30ドル、派遣ヒューマノイドに自宅の掃除を頼んでみた
A San Francisco service rents out a remotely operated humanoid robot for home cleaning. The robot works carefully but slowly and struggles on stairs.
The tutorial describes building and testing a low-latency gravity-compensation loop for a Seeed Studio reBot Arm with NVIDIA Holoscan 5 and Pinocchio. It explains the robot hardware, model requirements, and Holoscan tools for measuring control-loop latency and jitter.
A study shows EnsembleML uses an ensemble of four regressors trained on historical Pareto fronts to predict changes in dynamic multi-objective optimization, reinitializing the population with Gaussian perturbation and mutation. Evaluated on the CEC 2018 benchmark against eight methods, it often performs best on MIGD and MHVD, though results vary by problem and change severity, suggesting potential but limited real-world applicability.
IEEE provides a copyright and terms notice for content related to LEAN: Local Energy-Aware Navigation for skid-steer robots using deep reinforcement learning; the notice states IEEE's mission and 2026 copyright.
The paper introduces an adaptive-group process reward framework to improve reinforcement learning for GUI agents by partitioning trajectories into subgoals, evaluating actions within their groups, and aggregating these into refined process rewards that enhance training effectiveness; experimental results support its benefits.
RSI-Opt introduces a self-improving optimizer agent framework that stores prompts, procedures, and executable code, updated by a meta-agent from past runs. Across multiple generations and tasks, it yields reusable methods, enables cross-model policy transfer, and achieves substantial performance gains in optimizer-agent training and scientific machine learning benchmarks.