Hexo Labs released SIA, an open-source self-improving loop, under an MIT license. A Feedback-Agent reads each run's trajectory, then either rewrites the scaffold or triggers a LoRA weight update on gpt-oss-120b. Combining both levers beat scaffold-only iteration on LawBench, TriMul GPU kernels, and scRNA-seq denoising.

AI may change cross-platform app development from one UI framework to one product graph with native surface outputs.

PromptLayer is AI observability for developers, offering a unified timeline and waterfall view to trace requests, workflows, token usage, latency, costs, and failures across multi-step AI systems. Free beta is now available.

Enterprise AI initiatives treat design as a finishing step. Carsten Wierwille, Chief Product & Design Officer at HTEC, argues that this is a strategic mistake, explaining why many AI investments produce tools that work technically but fail to change how people actually work.

Comedian Chieng delivered a humorous yet earnest warning about AI to Harvard graduates during the 375th Commencement ceremony.

Drafted is an AI tool that allows users to instantly design home spaces.

Exploring the implications of AI-generated systems code.

The article presents multiple lines of evidence, including statistical analysis of punctuation and word usage, and results from an AI detection tool, to argue that Pope Leo's first encyclical on AI contains substantial portions written by AI, likely Claude. The author acknowledges each piece of evidence might be explained away but argues the consilience is hard to dismiss.

CodePulse is an open-source codebase indexer that saves 60-80% of token budget for AI coding assistants by maintaining a persistent, git-diff-aware index and injecting a compact snapshot at session start. It supports Claude Code, OpenAI Codex CLI, Cursor, and other tools, with features like task-aware ranking, git-aware ranking, and auto budget. It offers CLI, MCP server, and multiple integration methods.

Lithium is a hierarchical versioned storage engine built on PostgreSQL ltree, offering deterministic, scoped retrieval, built-in versioning, and zero runtime dependencies. It integrates with AI tools via MCP server, suitable for AI agent memory, decision tracking, and more.

The author, frustrated by clipboard sync issues under Wayland, used Claude Code to rewrite the Java project ClipCascade in Rust, creating the lightweight binary clipboardwire. The key insight: the bottleneck was the quality of feedback the AI received, and UI tests became the guardrails that enabled reliable iteration.

Xerolith is a working platform that achieves persistent identity, autonomous belief formation, and substrate-independent knowledge consolidation through a hierarchical fractal vault architecture. Over 80 days of continuous operation, it has compressed 2,817 raw entries into 1,218 beliefs, with complete genealogical tracing and internal alignment.

This article describes a macroeconomic research agent built with Deep Agents, LangSmith, and the You.com Finance Research API. It analyzes GDP data across all 27 EU member states, detects anomalies, and produces a cited briefing in approximately 45 minutes. The report details the anomalous growth in Ireland and contraction in Germany, emphasizing the importance of traceability and auditability.

This paper presents a dual-drone docking platform where two quadrotors operate in a leader-follower formation and dock using a lightweight modular frame with passive magnetic latching. A progress-aware mission supervisor manages phase transitions: approach, alignment, capture, and settle. The platform integrates a complete hardware-software stack (ROS 2 with Crazyflie/PX4 interfaces) and is evaluated in simulation and real-world experiments using quantitative metrics such as formation error, docking success rate, and time-to-dock.

This paper presents a phase-conditioned, force-aware framework for robust deformable object manipulation. Using FiLM-conditioned ACT encoder and multi-modal phase predictor, the system autonomously detects and recovers from contact failures, improving T-shirt hanging success rate from 56% to 87%.

This paper presents a decentralized framework that combines large language models (LLMs) with acoustic mobile robots for contactless object manipulation. Using Whisper speech recognition, LLM semantic parsing, and JSON task scheduling, the system converts spoken commands into coordinated multi-robot actions. Experiments with two TurtleBot3-based acoustic robots achieved success rates of 96% for sequential, 86% for parallel, and 70% for synchronized tasks, showcasing the potential of LLM-driven automation for human-robot interaction.

The Open Motion Planning Library (OMPL), first released in 2008, has become a cornerstone of the motion planning community, providing implementations of a wide range of state-of-the-art sampling-based algorithms. Over almost two decades of continuous development, OMPL 2.0 targets real-time motion planning through hardware acceleration and integrates seamlessly with modern AI research workflows.

MonoDuo proposes a framework for learning bimanual manipulation policies using single-arm robot demonstrations paired with human collaboration. By collecting data through teleoperation with role swapping and generating synthetic demonstrations via hand pose estimation, segmentation, and inpainting, MonoDuo achieves zero-shot success rates up to 70% on five tasks, with few-shot finetuning further boosting performance.

Researchers propose dynamic symmetry, quantified by dynamic isotropy, as a measure of uniformity in a robot's attainable center-of-mass accelerations. Through simulations and physical experiments, high dynamic symmetry improves trajectory tracking, task success, robustness, resilience, and energy efficiency. The Argus family of spherical robots, especially a 20-legged variant with near-extreme dynamic isotropy, demonstrates orientation-invariant locomotion, agile terrain traversal, rapid self-stabilization, and resilience to actuator failures.

This paper proposes a data-driven approach using recurrent neural networks and one-step-ahead predictive control for bead geometry control in Wire Arc Additive Manufacturing (WAAM). By updating the model online to account for changing thermal conditions, it significantly improves bead height and width consistency.

Researchers propose a multi-resolution end-to-end deep neural network to balance latency and safety in autonomous driving. By selecting input resolution at runtime, the network improves safety metrics like lane invasions, red-light infractions, and collisions in CARLA simulations compared to fixed-resolution baselines.

This paper introduces the 'Bionic Swarm,' a human-in-the-loop system that lowers barriers to real-world validation of swarm robotics. It uses a smartphone web-app, Bluetooth sensors, and a centralized server to direct human users. The Score-Biased-Search algorithm for soil mapping demonstrates superlinear map reconstruction in both simulations and outdoor experiments.

This paper proposes a target-informed self-supervised pretraining and model-ensemble strategy that leverages unlabeled target-domain data to improve cross-device generalization of medical imaging AI. Applied to pediatric wrist fracture assessment using point-of-care ultrasound, the method achieves over 6% Dice improvement on the target domain, demonstrating a label-efficient and privacy-preserving approach.

This paper introduces GeRaF 2.0, a unified framework integrating Line-of-Sight (LoS) and Non-Line-of-Sight (NLoS) neural geometry reconstruction, leveraging LoS geometry to guide RF propagation for stable and physically consistent 3D reconstruction of hidden scenes, achieving state-of-the-art results.

GeRaF is the first method to use neural implicit learning for near-range 3D geometry reconstruction from RF signals, addressing challenges like noise and specular reflections through novel rendering and sampling strategies.

This paper proposes two lightweight face forgery detectors, LFWS and LFWL, built on Xception (21.9M params) by adding a fusion module with only 292 extra parameters. They combine wavelet-denoised features with phase spectrum or local binary patterns, boosting AUC by 3.8% and 4.4% on FaceForensics++ and DFDC-Preview, respectively, outperforming larger models like F3Net and SRM across eight benchmarks.

This paper proposes a self-supervised enhancement framework for Sentinel-1 Stripmap SAR imagery using azimuth subaperture decomposition. It generates training data without external sensors or simulated ground truth, integrates single- and multi-frame learning, and employs iterative inference. Experiments show it outperforms MERLIN in PSNR and SSIM, while MERLIN achieves higher ENL, highlighting a trade-off between structural fidelity and speckle smoothing.

Embodied3DBench targets low-level spatial intelligence in embodied 3D environments, with 6 task categories and over 21k QA pairs. Evaluations of 13 models show strong high-level reasoning but weak interaction-oriented perception. A synthesized dataset of 1.3M QA pairs significantly improves performance after fine-tuning.

This paper introduces TRACE, a training-free trajectory-constrained reconstruction framework that stabilizes the reconstruction path by coupling adjacent states, improving reconstruction quality for imaging inverse problems.

This paper audits evaluation protocols for training-free shape descriptors by introducing Diffused Geodesic Moments (DGM). Experiments show that Geometric Moment Shape Descriptor based on Heat Kernel Signature (GMSD-HKS) achieves the highest scores on FAUST-Reg and TOSCA, while Wave Kernel Signature (WKS) remains strong. DGM is valuable for sparse or non-spectral applications. The work provides a reproducible protocol-cascade analysis, cross-shape alignment diagnostic, and recommendations for designing and reporting training-free descriptors.