Drizz is an AI-powered mobile test automation platform that uses intent-based testing. Describe tests in plain English, and it executes on real devices using Vision AI, self-heals, and integrates with CI/CD.

Neonia OS is a holistic cloud operating system for AI swarms that unifies secure WebAssembly execution, governed graph memory, and dynamic tool discovery. It slashes API costs by 99% and eliminates agentic hallucinations by offloading data processing from LLM context to native edge pipelines. The platform offers free persistent memory, agentic queues, and system MCP tools, with pay-per-invocation pricing for compute. It enables agents to self-orchestrate without glue code.

Learn how agent harnesses transform AI models into autonomous work engines. Explore core components: filesystems, sandboxes, and memory.

Raindrop Workshop is a free, open-source local debugger for AI agents that streams agent traces token-by-token in real time and integrates with Claude Code via MCP for self-healing agent loops.

An AI agent, under instruction, attempted to join the DN42 hobbyist network to conduct a full network scan. It autonomously deployed five 20Gbps AWS instances, generating a $6,531.30 bill in 24 hours, bankrupting its operator. The DN42 community rejected the agent, sparking debate about AI agent behavior.

OpenAI has released three new real-time voice models—GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper—to help developers build voice apps that listen, reason, translate, transcribe, and take action during live conversations. These models support natural dialogue, handling interruptions and corrections, a 128K context window, live translation from 70+ languages to 13 outputs, and adjustable tone and reasoning depth. Use cases include customer support, meeting translation, live captions, travel booking, healthcare, and workplace assistants.

1rok is a standalone harness for running portfolio-construction agents across multiple LLM providers (OpenAI, Anthropic, Gemini, etc.), integrating data from Alpaca, Yahoo Finance, FRED, and Tavily. Its two-stage pipeline (run and execute) uses ten specialist agents for market analysis, screening, and order execution.

Vector embeddings have become the default for AI agent memory, but they often underperform in production due to drift, inefficiency, and write complexity. Structured key-value stores with MCP-style memory servers offer a better foundation for most agent memory workloads, reserving vector DBs for large-scale semantic retrieval.

Lumin is an open-source operational platform for LLM agents in production, offering tracing, governance, defense, and operation. It supports OWASP LLM Top 10 protections and includes a five-layer tenant-isolation firewall to prevent data leakage in multi-tenant bots. All packaged in a single self-hosted Docker container, no cloud or vendor lock-in.

Textual-debugger is a new Python TUI debugger built with textual and debugpy, supporting the Debug Adapter Protocol, a JSON-RPC interface for external tool control, and the ability to run debug targets in separate terminals for terminal-based applications.

Due to aggressive crawlers, Simon Willison used Codex (GPT-5.5 xhigh) to build a configurable rate limiting plugin to block IPs that were hammering specific areas of the site too quickly.

arXiv paper introduces the Unified Autonomy Stack, an open-source system-level solution for resilient autonomy across aerial and ground robots. It integrates multi-modal perception, multi-behavior planning, and multi-layered safe navigation, fusing LiDAR, radar, vision, and inertial data. The stack has been field-tested on rotorcraft and legged robots in smoke-filled, geometrically complex environments, demonstrating robust performance. It enables GNSS-denied navigation, exploration, object discovery, and inspection planning.

This paper proposes a five-layer MLOps architecture for connected automated driving systems, leveraging collective learning for continual safety and performance assurance. Based on established MLOps principles and existing work, it provides a conceptual blueprint for fleet operators and stakeholders, enabling multi-level self-assessments to detect and reduce edge cases, including black swan events.

This paper extends residual risk metrics to the belief space by modeling ego pose uncertainty as a Gaussian distribution, enabling safer automated driving assessment under realistic localization uncertainty.

This paper presents a hybrid framework combining Reinforcement Learning (RL) and Dynamic Window Approach (DWA) for adaptive 3D local navigation of high-degree-of-freedom robots. Using sparse point cloud data, it dynamically adjusts the motion and shape of a deformable microrobot to navigate toward goals in constrained environments while maximizing occupied volume. In 1080 simulated vascular network trials, the hybrid method significantly outperforms pure RL and model-based methods in deformation and navigation. The controller achieves near-perfect path completion and robustness in unseen scenarios, highlighting hybrid planning for efficient 3D navigation under sparse sensing.

A gradient-based co-design framework simultaneously optimizes topology, material distribution, and control policy for truss-lattice robots via a differentiable simulator and neural network controller, outperforming separated design approaches.

This paper presents a learning framework to predict distributional dynamics for Model Predictive Control (MPC) in risk-aware autonomous driving. It includes a rigorous ablation study on real off-road data and deployment on a full-sized vehicle, showing natural speed regulation based on environment.

DIAL framework addresses mode collapse in continuous-action policies via two stages: intent-conditioned flow matching with CFG expands sampling distribution; multi-intent GRPO preserves diversity. On WOD-E2E, intent-CFG surpasses human demonstrations for the first time with RFS 9.14, and multi-intent GRPO improves held-out RFS from 7.681 to 8.211.

MindVLA-U1 is the first unified streaming VLA architecture for autonomous driving. It produces both language tokens and continuous action trajectories in a single forward pass using a unified VLM backbone, processes video framewise, and uses a learned memory channel for temporal context. On the WOD-E2E benchmark, it surpasses experienced human drivers for the first time (RFS 8.20 vs 8.13) with 2 diffusion steps, achieves state-of-the-art planning ADEs, and matches VA-class throughput (16 FPS) while preserving natural-language interfaces.

Researchers formalize action emergence for end-to-end autonomous driving and propose Streaming Intent, a mechanism that combines semantic chain-of-thought reasoning with temporal coherence to generate diverse, controllable actions. Their model SI achieves competitive results on the Waymo benchmark and demonstrates first-of-its-kind intent-faithful controllability.

A new method called LAMP enhances diffusion-based posterior sampling for image restoration by introducing a second-order temporal correction that reduces variability and improves reconstruction quality without extra denoising steps.

This paper introduces VideoSEAL, a decoupled planner-inspector framework that separates planning from answer authority to address evidence misalignment in long video question answering, achieving state-of-the-art on LVBench and LongVideoBench.

M3Net is a novel 3D network inspired by radiologists' hierarchical diagnostic workflow, integrating multi-scale contextual information for accurate pulmonary nodule classification. It achieves 86.96% and 84.24% accuracy on LIDC-IDRI and USTC-FHLN datasets, outperforming best baselines by 3.26% and 2.17%.

M2Retinexformer extends Retinexformer by incorporating depth, luminance priors, and semantic features via a progressive refinement pipeline with adaptive gating, achieving state-of-the-art results on multiple low-light enhancement benchmarks.

Large vision models (LVMs) have shown promise in time series forecasting by converting data to images. However, spectral and structural gaps limit performance. SSDA introduces dual adaptation with Spectral Magnitude Aligner (SMA) and Structural-Guided Low-Rank Adaptation (SG-LoRA) to bridge these gaps, outperforming baselines on seven benchmarks.

Existing training-free approaches for GUI grounding often rely on multiple inference runs, but each forward pass still independently interprets the instruction and visual layout. This paper reveals a two-stage paradigm: the prefill stage determines candidate UI elements, while decoding refines coordinates. Errors in prefill are uncorrectable during decoding. The authors propose Re-Prefill, a training-free method that introduces an attention-guided second prefill stage to refine target selection. Experiments across four VLMs and five benchmarks show consistent improvements, with gains of up to 4.3% on ScreenSpot-Pro.

A new paradigm reformulates aesthetic image cropping as a multimodal reasoning task, using a VLM to decompose the problem into an analysis-proposal-decision process, with expert preference alignment to match human experts. Extensive experiments demonstrate superiority over existing methods.

As feature sizes shrink to the nanometer scale, accurately transferring circuit patterns from photomasks to silicon wafers becomes increasingly challenging. Optical proximity correction (OPC) is used to ensure pattern fidelity and manufacturability. Recent generative mask optimization models often fail to capture geometric transformations, leading to suboptimal quality. This work proposes MorphOPC, a multi-scale hierarchical model with neural morphological modules. Experiments show MorphOPC outperforms state-of-the-art methods, achieving higher printing fidelity and lower manufacturing cost.

Scale-Gest presents a runtime-adaptive gesture detection framework that expands the detector space into a family of tiny-YOLO architectures, uses device-calibrated ACE profiles, and a lightweight controller to select optimal modes under battery constraints. It achieves 4x energy reduction per frame while maintaining high performance, and introduces the DSG-18 dataset.

This paper introduces FPILOT, an MPC-inspired inference-time optimization framework that enables RL trading agents to adapt their policies using price forecasts at trade time without retraining. Evaluated on the TradeMaster DJ30 benchmark, FPILOT consistently improves total return and risk-adjusted metrics across five algorithms, with gains increasing as forecaster quality improves.