NeuralStackAgentRxiv Lets AI Research Agents Build on Each Other’s Work
AgentRxiv is an open-source framework that lets autonomous LLM research laboratories upload and retrieve reports through a shared preprint server. The researchers found that agents building on prior work outperformed isolated agents, including a 13.7% relative improvement over baseline on MATH-500 when multiple laboratories collaborated. The approach points toward AI systems that can accumulate research knowledge and help humans accelerate scientific discovery.
Penn Researchers Use Light-Matter Particles for Ultra-Low-Energy AI
University of Pennsylvania researchers demonstrated all-light signal switching with exciton-polaritons, hybrid particles that combine photons with matter. The approach addresses a key weakness of photonic computing, which often converts optical signals back into electricity for nonlinear operations. If it scales, the technology could enable faster, more energy-efficient AI chips that process information directly from light.
Sakana AI’s AI Scientist Publishes Automated Research in Nature
Sakana AI and academic partners have published a Nature paper detailing The AI Scientist, an agent designed to automate the machine-learning research lifecycle. The system generates ideas, studies literature, writes and runs experiments, produces papers, and reviews its own work, while an earlier version produced a paper that passed human peer review at an AI workshop. The publication offers a significant test of AI-led science but also highlights persistent weaknesses in originality, rigor, implementation, and factual accuracy.
AgentRxiv Lets AI Research Labs Build on Each Other’s Discoveries
AgentRxiv is an open-source framework that lets autonomous LLM research laboratories upload and retrieve reports through a shared preprint server. The authors found that agents using prior research improved reasoning performance, while multiple laboratories collaborating toward a shared goal outperformed isolated systems. The work points toward AI systems that can contribute to research through cumulative, parallel experimentation rather than one-off autonomous workflows.
ASI-Arch Uses Autonomous Evolution to Discover Neural Architectures
ASI-Arch autonomously explored neural architectures through a closed-loop process of research, experimentation, analysis and updating. Applied to linear attention, the system ran 1,773 experiments and discovered 105 state-of-the-art architectures, with its best design improving on DeltaNet by nearly three times the gain achieved by Mamba2. The result suggests AI agents may help automate parts of AI research, while also raising questions about whether LLM-driven neural architecture search merits broader claims of autonomous scientific intelligence.
Appier’s SMITH Helps AI Agents Build, Refine and Share Their Own Tools
Appier says its SMITH framework, accepted at NeurIPS, enables AI agents to create and use tools within one reinforcement-learning loop. The company reports that a roughly 4-billion-parameter model built tools that outperformed methods tested on unseen tasks, while those tools worked across models of different sizes. By converting repeated reasoning into reusable tools, SMITH cut average output from 3,206 tokens to about 100 in the company’s experiments, pointing to more efficient multi-agent systems.
Google’s September AI Push Centers on Gemini 4 Argon and Connected Apps
Google’s September AI announcements were led by Gemini 4 Argon, a frontier model built for complex reasoning, cybersecurity defense, coding, and knowledge work. The company also introduced Gemini 3.8 models, expressive voice and music tools, Connected Apps, scientific AI projects, and Gemini for Windows. The rollout shows Google pairing increasingly capable models with broader consumer workflows while using staged access to test safeguards for higher-risk capabilities.
Big Tech’s AI Race Shifts From Models to Long-Running Agents
OpenAI, Meta, Google, and xAI introduced AI systems designed to execute extended workflows, use tools, and coordinate subtasks, while GPT-5.6 became a priority model for Microsoft 365 Copilot. The week’s announcements also showed enterprise deployment, infrastructure, governance, and operational costs becoming central to competition. The shift suggests that AI leadership will increasingly depend on reliable execution and real-world integration rather than standalone model performance.
Critical Vulnerabilities Hit SonicWall, LMCache and Atlassian Systems
SonicWall patched a CVSS 10.0 pre-authentication SSRF flaw in SMA1000 appliances, while a critical unpatched LMCache bug allows unauthenticated remote code execution in exposed deployments. Threat actors are also exploiting an Atlassian Data Center flaw that can expose specific files without authentication. The incidents highlight growing risk across remote-access gateways, AI infrastructure and self-hosted enterprise platforms.
Penn’s Light-Matter Particles Could Make AI Computing More Efficient
University of Pennsylvania researchers have demonstrated all-light signal switching with exciton-polaritons, hybrid particles that combine photons with matter. The approach addresses a key limitation of photonic computing, where light-based systems often convert signals back into electricity for nonlinear operations. If it scales, the technology could enable faster, lower-energy AI chips and support future optical and quantum computing applications.
DeepScientist Claims AI Can Outpace Human Research on Three AI Tasks
DeepScientist used more than 20,000 GPU hours to generate about 5,000 scientific ideas and experimentally validate roughly 1,100 of them. The Westlake University system surpassed human-designed state-of-the-art methods across three frontier AI tasks, according to its arXiv paper. Its goal-oriented, memory-driven search offers a model for AI research systems focused on measurable scientific progress rather than open-ended idea generation.
ASI-Arch Discovers 106 AI Architectures in 1,773 Autonomous Experiments
ASI-Arch autonomously generated, implemented, tested and analyzed neural architectures through 1,773 experiments using about 20,000 GPU hours. The system produced 106 state-of-the-art linear attention architectures and identified design patterns that surpassed human-designed baselines. Its reported scaling relationship between computation and discoveries suggests that parts of AI research could become less constrained by human researchers’ cognitive capacity.