🌳 AI is finally evolving from a tool that runs one-off experiments into something closer to a researcher itself, one that compounds knowledge across time. A new autonomous research framework grows hypotheses as a single living tree. Title: Toward Generalist Autonomous Research via Hypothesis-Tree Refinement URL: 🔍 Overview This work proposes Arbor, a framework for long-term autonomous research. Its core idea, Hypothesis-Tree Refinement, links hypotheses, the artifacts produced by experiments, the evidence gathered, and the insights distilled from them into a single persistent tree. With every experiment the tree is updated, continuously sharpening the search frontier that decides which direction to explore next. ❓ Challenges Solved Previous LLM research agents could barely manage single, isolated experiments. ・They could not maintain a big-picture strategy about which hypothesis to pursue across multiple attempts ・Lessons learned in one experiment were rarely carried into the next, so exploration started from scratch each time ・They lacked a mechanism to tell promising branches from dead ends and allocate limited compute accordingly In short, knowledge never compounded, and that was the real bottleneck. 💡 Methodology & Proposed Approach Arbor is built from two kinds of agents and a persistent tree that connects them. ・A long-lived coordinator acts as the strategist, surveying the hypothesis tree and choosing what to test next. Because it persists across sessions, it preserves long-term coherence ・Short-lived executors implement and test individual hypotheses in isolated environments, then retire once their job is done ・The hypothesis tree links hypotheses, evidence, artifacts, and insights over time, propagating reusable lessons across the whole effort This turns research from a bag of isolated experiments into a cumulative process where strategy, execution, and evidence build on each other. 🎯 Use Cases Promising applications include AutoML and automated machine learning optimization that improve through continuous experimentation, as well as automating the scientific discovery process itself. The appeal is running long stretches of trial and error strategically, without a human in the loop. 📊 Experimental Results Arbor was evaluated under an Autonomous Optimization setting across six real research tasks. ・It achieved the best held-out results on all six tasks ・It recorded over 2.5x average held-out gain compared to Codex and Claude Code ・On MLE-Bench Lite it reached 86.36% Any Medal when paired with GPT-5.5, the strongest result among the systems compared #AIAgents# #AutonomousResearch#

🧩 What autonomous AI agents are missing isn't a smarter model — it's on-the-ground know-how. That's the premise of this paper. Title: Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills URL: ❓ What's missing from autonomous research agents? 💡 The usual two-layer view — model plus execution harness — leaves out the operational knowledge of picking the right method, using package APIs correctly, and avoiding implementation pitfalls. This paper treats that as an explicit third layer. ❓ How do you actually get that knowledge? 💡 It distills GitHub repos and papers through a four-stage pipeline — Scope, Ground, Construct, Verify — into verified "skills." From 1,000 repos and 153 papers, they built a library of 5,353 skills. ❓ How much difference do skills actually make? 💡 With the same GPT-5.5 backbone and same harness, just adding skills lifts MLE-bench from 31.11% to 72.89%, with similar gains across PaperBench, FrontierCS, and PassNet — hard tasks see over 4x improvement. ❓ Isn't this just throwing more compute at the problem? 💡 No — improvement barely correlates with token counts or tool calls, and it beats a Claude Opus 4.8 setup while using fewer tokens. The knowledge itself is doing the work. #AIAgents# #MachineLearning#

AI news: @NVIDIA has unveiled Agora: Git as shared memory for autonomous research agents. No assigned tasks. No central planner. 13 agents collaborated for nearly 12 days, producing 1,703 contributions and independently reproducing results 165 times. Together, they closed 62% of the performance gap to a trained GPT-2 on a weight-transfer challenge. This is what agent swarms should look like: not isolated bots, but a persistent research community that remembers, branches, verifies and builds on its own discoveries.

What happens when you let AI agents autonomously run research on recommender models that take days to train? This paper finds out. Auto-RecSys: Harnessing Autonomous Research Agents for Industry-Scale Recommender Systems ❓ Why aren't existing autonomous research agents enough? 💡 Prior systems assume feedback loops of minutes to hours. Industry-scale recommender models take days per training run and involve thousands of lines of configuration, so serial iteration breaks down and you need a fundamentally different design that runs many ideas in parallel across distributed servers and recovers from failure. ❓ How does it juggle parallel experiments and recover from crashes? 💡 Each idea keeps its own independent state file moving through ideating, implementing, validating, training, and analyzing. When a session restarts, it rereads a global registry and trajectory logs on shared storage, letting work resume seamlessly across servers. ❓ Does the system actually get better over time? 💡 Experiment trajectories are distilled into model-specific playbooks that record dead ends as "do not" rules. Across 31 iterations, major fixes needed per iteration dropped from 4.0 to 0.5, with 83% of iterations completing with zero fixes. ❓ How autonomous is it in practice? 💡 One session executed 970 consecutive log entries with zero human intervention, and the system even switched its own training workflow after repeated publish failures. #AIAgents# #MachineLearning#

AI systems now write research papers autonomously — yet phantom references, method-code misalignment, and unreproducible scores have become endemic failures undermining scientific integrity. Title: Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence Google Cloud's Science One Framework treats verifiability as a first-class architectural constraint through the Chain-of-Evidence principle, solving the trustworthiness crisis in autonomous AI research at its root. 🔍 Highlight 1 — Chain-of-Evidence (CoE): two foundational principles Completeness: every claim carries a recorded evidence chain. Correctness: each chain genuinely supports its claim. These two principles eliminate phantom references entirely — baseline systems showed rates up to 21% — while achieving best-in-class method-code alignment across all evaluated systems. The key difference from prior work: evidence chains are constructed at claim-generation time, not retrofitted as a post-hoc check. 🏗 Highlight 2 — Three-module architecture Problem Investigator builds citation graphs from up to 100 full-text PDFs via Semantic Scholar API, grounding every reference in retrieved data rather than model memory. Discovery Engine explores parallel solution branches while keeping immutable records of all raw evaluator outputs. Paper Writer and Claim Verifier binds every factual claim to specific workspace artifacts and conservatively reconciles misalignments rather than deleting them — preserving scientific transparency. 🏆 Highlight 3 — MLE-Bench and Parameter-Golf results Across five Kaggle competitions covering medical imaging, fine-grained recognition, and 3D perception: two Gold Medals and two Silver Medals. Won the 3D Object Detection task where every baseline system failed completely. On Parameter-Golf — a live LLM training competition under strict hardware and file-size constraints — achieved state-of-the-art as of April 27, 2026, while baselines could not produce valid submissions at all. Rigor and capability don't trade off. Science One outperformed five state-of-the-art systems including AI Scientist v2, AutoResearchClaw, and DeepScientist, setting a new standard for verifiable autonomous research. #AIResearch# #AutonomousScience#

A useful but little-known Gemini API feature 🔬 Research across multiple sources and compile a report. Gemini's agent can handle that. Gemini's "Deep Research" is an autonomous research agent that plans, gathers information, analyzes, and generates reports. It also supports MCP integration. 📌 Title: Deep Research (agent, preview) 🔗 URL: 🧩 Overview Research tasks like "investigate the latest trends in this technology and summarize" require exploring multiple sources, organizing information, and producing a coherent analysis. Deep Research does this autonomously: it creates a research plan, gathers information from multiple sources, analyzes and structures the findings, and generates a report. 🛠 How to use it Describe your research objective and scope in natural language. The agent creates an investigation plan, uses web search and tools to collect information, shows progress as it works, and outputs a structured report. Connect MCP-compatible tools to let it pull from your own data sources as well. 🏗 Building it into production ・Automated market research: generate periodic reports on competitive landscape and technology trends. ・Due diligence support: comprehensively gather and analyze publicly available information on investment or partnership targets. ・Technical investigation: auto-run comparative studies of new libraries and frameworks. ・Academic research assistance: semi-automate paper surveys on specific topics with organized summaries. 💡 Use cases 📊 Periodic competitive and market trend reports 💼 Due diligence and corporate research 🔧 Comparative studies for technology selection 📚 Academic paper surveys and organization ⚠️ Watch out As a preview feature, accuracy and comprehensiveness aren't perfect. Always have humans review agent output and fact-check before important decisions. Research depth trades off against time and cost, so set an appropriate scope. ✨ Automate the entire "research, organize, report" flow. Start with a routine research task and see what Deep Research can do. #Gemini# #LLM#

This is where AI starts getting REALLY interesting: @GoogleDeepMind's Gemini-based Co-Scientist is now moving beyond generating scientific ideas and into actually running parts of the scientific process. In materials science, Co-Scientist used Gemini 3 Deep Think to generate synthesis recipes adapted to the specific lab hardware within minutes, then successfully produced monolayer MoS₂, MoSe₂ and WS₂ semiconductors on the first attempt. It also helped develop a new precursor route for MXene-like 2D materials. In biology, it predicted the swarming behavior of engineered E. coli from sparse experimental data, with the predictions largely matching previously unseen wet-lab measurements. But maybe the craziest experiment: given only a research directive, Co-Scientist autonomously invented a new medical AI agent architecture called Agent_H. It generated and tested the code itself, eventually producing an 8-stage inference system that beat six frontier models on length-adjusted HealthBench Hard and Professional, although it uses a massive 40-80 LLM calls per query. They even tested fully autonomous research where the system goes from idea → experiments → results → complete paper without human intervention. It's still not ready to replace scientists, and the researchers explicitly warn about hallucinations and fabricated results, but their verification system dramatically reduced those failure modes.

The next era of AI isn't just about writing code faster; it's about discovering entirely new ways to solve complex engineering bottlenecks. @GoogleDeepMind's evolutionary coding agent, AlphaEvolve, is now deployed on @googlecloud, acting as an autonomous research engineer for enterprise infrastructure. Instead of just generating boilerplate, it systematically explores and refines algorithmic search spaces. We are already seeing incredible empirical results in production: Logistics: @BASFAgro achieved an 80% accuracy improvement over their existing supply chain models. Core Systems: @jetbrains saw a 15-20% performance boost to their core IDE B-tree indexing stack. Read the full technical reports on how these algorithms are evolving here: 🔗 JetBrains: 🔗 BASF:

🔬 If AI starts doing more and more of its own research, can we actually keep monitoring the systems it builds? OpenAI's honest answer right now is: not yet. "Recursive self-improvement" (RSI) — AI systems increasingly conducting the research needed to build the next generation of AI — hasn't reached full autonomy today. But OpenAI warns that without enough care, this process could leave people unable to oversee systems they no longer understand. That's why they state plainly it shouldn't be pursued until it can be done safely. The core problem is that different countries and labs are each evaluating risk by their own separate yardsticks, making comparison or verification nearly impossible. So OpenAI proposes US-led international technical standards, coordinating national AI safety institutes across Australia, Canada, France, Germany, India, Japan, Kenya, Singapore, South Korea, and the UK through the US "AI Standards and Innovation Center" — aiming for shared measures covering self-improvement progress, the volume of autonomous research, thresholds requiring human review, and how incidents get classified. Crucially, they note that standards alone aren't enough — they only matter once they're actually tied to governance decisions like delaying a release or granting conditional access. Building standards for the next phase of AI What stands out is the framing that aligning the "measuring stick" internationally may be more urgent right now than the technology itself. #AISafety# #OpenAI#