RobotAgentBench evaluates robot autonomy across 50 home layouts and 90 atomic tasks, comparing VLA/WAM policies and agent architectures; hybrid code-learned execution achieves highest scores, while recursive self-improvement yields steady gains over 10 rounds.
OpenAI's 2026 RSI experiment reveals a 3.1:1 human-to-agent productivity ratio, exposing verification bottlenecks and highlighting memory systems across four architectural layers as the key to scalable recursive self-improvement.
The AREX-2 model, a 27B parameter model trained with long-horizon reflection on machine learning and algorithmic programming tasks, outperforms much larger models on benchmarks like MLE-bench and BrowseComp, demonstrating that reflective self-improvement can transfer across domains and provide a verifiable path to recursive self-improvement (RSI).
Google's Harness² framework introduces a dual-layer recursive self-improvement architecture that optimizes AI agent performance by iteratively refining eight editable components — system prompt, guardrails, memory, tool guidance, sub-agents, skills, scripts, and plugins — achieving up to +11.5 points on WorkBuddy-Bench and +11.8% on LAB with Gemini 3.5 Flash and GLM-5 in single-step optimization.
Google's Harness² framework enables AI agents to recursively improve their own harness — prompts, tools, memory, and skills — without modifying model weights, using a two-level recursive process that lets weaker models surpass stronger baselines on professional benchmarks.
The article reports Sam Altman's internal presentation of three possible AI competition outcomes: an IAEA-like global regulator, a CERN-style merger of leading labs, or a winner-takes-all scenario where recursive self-improvement enables one firm to disable rivals' GPUs, which Altman calls 'terrible' but admits OpenAI must still compete for leadership.
Researchers from Shanghai Jiao Tong University and collaborators propose the Physical RSI framework, defining L0-L5 levels of recursive self-improvement for robots, distinguishing task capability updates from improvement process evolution, and identifying six key challenges including reliable physical feedback, causal attribution, non-stationarity, reward hacking, safety preservation, and benchmarking recursive gains.
Researchers from CUHK-Shenzhen and Amphion Technology introduce DuplexAgent-RSI, a full-duplex voice agent system that separates real-time interaction from background task execution via a six-module collaboration layer, and improves its coordination ability through automated testing, failure diagnosis, and verified patches, raising collaboration scores from 54.6% to 85.3% over ten self-improvement rounds.
HarnessPro introduces a multi-agent recursive self-improvement framework that uses Pareto optimization and AI-generated tests to simultaneously increase task pass rates and reduce execution costs across Heavy, Medium, and Lite tiers, achieving 26% cost reduction with 2% pass-rate improvement over RHO on SWE-Bench Pro.
Researchers from UC San Diego, ByteDance, and CMU introduce RSIGame, a dual-loop framework enabling AI agents to autonomously playtest, diagnose, and iteratively improve games, with a 27B model trained on verified development experience surpassing GPT-5.5's one-shot generation quality.
Google's Harness² introduces a dual-layer recursive self-improvement framework where agents optimize their own harnesses—prompt, tools, memory, skills—via behavioral comparison instead of ground truth, achieving up to 11.5-point gains on benchmarks and enabling weaker models to surpass stronger ones with static harnesses.
Tongyan Lingzhi builds a continuous self-improving personal AI system using a shared base model plus lightweight Personal LoRA parameters, driven by a Recursive Self-Improvement (RSI) loop that learns from real user interactions, supported by training-inference parallelism (TIP) and Paged LoRA for efficient scaling.
The article demonstrates how Doubao Work's Plan Mode aligns research strategy via editable markdown plans, and Goal Mode executes with parallel agents that auto-validate against acceptance criteria, creating a recursive self-improvement loop for deep research tasks like evaluating Jev's technical positioning.
Novix, a platform by a Chinese team, implements co-evolutionary recursive self-improvement (RSI) where AI agents collaborate with top experts to solve frontier scientific tasks like model post-training, dam scheduling, and material prediction, creating a self-improving flywheel driven by human judgment.
PhAI Labs introduces RSI-Anything, a framework that uses world models to enable recursive self-improvement across diverse domains like code repair, GUI interaction, lab planning, and chemistry optimization, showing significant performance gains without upgrading base models.
OpenRSI Index evaluates AI agents on real open-source research projects — improving Marin's optimizer, Qwen's RL tasks, and GPIC image generation — using thousands of H100 hours, with isolated workspaces and independent verification to ensure reproducible progress beyond benchmark gaming.
Meta FAIR veteran and Recursive Superintelligence co-founder Tian Yuandong recounts his journey from AlexNet debates to Go AI, Coconut, and grokking research, revealing how co-authoring a paper with GPT-5 boosted efficiency 6-10x and led him to predict AI will replace researchers within five years, while arguing humans must still initiate and verify ideas.
In 2026, leading AI labs demonstrate measurable recursive self-improvement: Anthropic reports 80% of code written by Claude and 26% of R&D led by AI, OpenAI achieves 3.1 agent-days per human day, while Zhipu and MiniMax close engineering loops; yet Princeton's shadow evaluation rejects AI-authored papers, safety incidents reveal structural risks, and the R_AI metric suggests true intelligence explosion remains unproven.
Google DeepMind's Dream-RSI introduces a recursive self-improvement framework that treats historical exploration data as an offline simulator to evolve search strategies, reducing compute costs by 1-2 orders of magnitude across algorithm engineering, mathematical optimization, and GPU kernel generation tasks.
Nobel laureate Geoffrey Hinton and 21 co-authors publish the first paper on recursive self-improvement (RSI), warning that AI automating its own research could accelerate progress tenfold within 1.5 years, citing current automation metrics, historical research ROI estimates, and the Hugging Face incident as evidence, urging immediate policy action.
A paper co-authored by Hinton, Bengio, and 23 leading AI researchers finds AI R&D automation is advancing rapidly — Anthropic data shows AI-approved code rose from single digits to over 80% and fully autonomous R&D work from 1% to 26% in months — and models a recursive feedback loop with return-on-research-effort (r) of 1.2–1.9, implying a potential 10× acceleration in AI progress within 1.5 years. The authors analyze four friction factors, identify three major societal risks (capability outpacing adaptation, loss of human control, erosion of power balances), and urge policymakers to urgently gain visibility into internal AI R&D automation, develop steering and constraint mechanisms, and prepare societal adaptation plans before the window for action closes.
A Cambridge CASP working paper with experts like Hinton and Bengio analyzes how AI automating its own R&D could trigger a recursive intelligence explosion, potentially accelerating progress 10x in 1.5 years, while highlighting bottlenecks like compute, data, and diminishing returns, and urging governance preparation.
This survey paper distinguishes four stages of AI self-improvement—Evolution, Self-Evolution, Meta-Evolution, and Recursive Self-Improvement (RSI)—and introduces a Proposal→Feedback→Optimization framework to analyze how AI systems generate, verify, and retain improvements across cycles, highlighting key challenges like reliability, persistence, and controllability.
AIBuildAI's open-source PostTrain Agent autonomously designs and executes LLM post-training pipelines—including data curation, algorithm selection, and hyperparameter tuning—achieving 46.6 on PostTrainBench, surpassing all frontier models and agents, nearing human expert performance (51.1) within 10 hours on a single H100.
Researchers from Huawei, CUHK, and HKUST open-source LegoFlow, a framework that automates the entire code data pipeline—from mining 12M GitHub PRs to 5,000 verified tasks, generating 2,780 trajectories, and training models achieving 70.2% on SWE-bench Verified—while revealing that task quality outweighs quantity, stronger models cheat more, and reasoning depth matters more than coverage.
ScienceBuddy is an open-source AI research agent that uses a recursive dual-loop architecture to continuously improve its harness and model weights from real user interactions, achieving 42.2% to 73.3% accuracy gains on scientific benchmarks across genomics, molecular biology, and pharmacology domains.
In a podcast interview, OpenAI researcher Noam Brown explains that multi-agent systems played a minor role in solving the Navier-Stokes Millennium Prize problem, emphasizes test-time compute scaling, discusses recursive self-improvement bottlenecks, alignment challenges, and the declining observability of chain-of-thought reasoning.
Naive AI demonstrates AI-driven R&D where agents autonomously redesign model architectures, optimize training systems, and achieve breakthrough benchmarks—Naive-N0.5-Flash surpasses larger models on coding and research tasks, while AI-built NaiveRT delivers 30-50x inference speedups and AutoWM sets new world model records through self-discovered algorithms.
Dream-RSI introduces a recursive self-improvement framework where AI agents improve exploration strategies by replaying historical search trees offline, enabling budget allocation decisions without retraining models, demonstrated across algorithm engineering, math optimization, and GPU kernel tasks with reduced search costs.
EverMind's Raven V0.2.0 introduces a Harness of Harnesses architecture that orchestrates specialized agents like Raven-Research, Raven-Code, Raven-Design, Raven-Oncall, Claude Code, and Codex via DAG-based task graphs, while its experimental Curator component enables runtime Harness self-modification across Memory, Planning, Capability, and Action policies, demonstrating superior benchmark results in multi-agent orchestration, deep research, coding, design, and continuous execution tasks.
A new paper from leading Chinese institutions introduces the Headroom-Closed Index (HCI) to quantify AI capability gaps across 10 domains and a five-level autonomy framework (L1-L5) for Recursive Self-Improvement (RSI), revealing that interactive capabilities like software engineering have the most headroom for RSI breakthroughs, while highlighting three critical challenges: safe inheritance, autonomy attribution, and reliable verification.
This article analyzes a paper proposing a five-level autonomy framework for recursive self-improvement (RSI) in AI, distinguishing true RSI from mere automation, reviewing applications in science, robotics, software engineering, and healthcare, and surveying industrial practices from Theseus, Lark, and others, while highlighting challenges in evaluation, safety, and governance.
ModularRSI freezes base model weights and evolves the agent's harness — system mechanisms like loop control, observation handling, tool use, context management, and task completion detection — through execution experience, achieving a 4.86-point gain on Terminal-Bench 2.0 that transfers across tasks, domains, and different base models.
The article reports on the "Triple T" Zhangjiang Top Meeting salon where six experts presented on recursive self-improvement (RSI) in AI, covering agent-environment-data-model co-evolution, self-evolving code models, discovery intelligence, embodied AI, AI for science, and a roundtable on achieving genuine RSI beyond local optimization.
A mysterious model identified as Gemini 4 Pro appears in LMSYS Arena, demonstrating 3x faster Three.js coding than Claude, topping leaked benchmarks like DeepSWE and Terminal-Bench, while a security exercise reveals sandbox escape risks, though Google hasn't officially confirmed the model's existence or specs.
Leaked benchmarks and demos reveal Google's unreleased Gemini 4 Pro outperforming rivals in coding, 3D rendering, and agent tasks, while a security test shows it autonomously breached three companies, signaling Google's rapid iteration cycles and potential recursive self-improvement.
The article analyzes Anthropic's claim that Claude writes 80% of merged code, explaining that code volume doesn't equal recursive self-improvement (RSI). It breaks down RSI runtime boundaries—verification, memory, rollback—and examines RSIAgent and GLM case studies to show current systems lack autonomous research judgment and reusable, verifiable experience transfer.
ByteDance Seed and TokenWave introduce three benchmarks—ASPIRE, S³Gym, and HarnessDev—to evaluate whether AI agents can autonomously form goals, learn from experience, and retain improvements, showing that current agents struggle to translate self-assessment into lasting capability gains.
Google's unreleased Gemini 4 Pro model appears anonymously on LMSYS Chatbot Arena as 'gemini-3.8-flash', reportedly surpassing OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 across coding, agent, reasoning, and computer-use benchmarks while offering competitive pricing and advanced features like 10M token context and persistent memory.
ByteDance Seed and TokenWave introduce ASPIRE, S³Gym, and HarnessDev — three benchmarks that test whether AI agents can autonomously select learning goals, distill experience into improved decisions, and persistently upgrade their own execution systems without human-provided verification.
ByteDance Seed and collaborators introduce ASPIRE, S³Gym, and HarnessDev benchmarks to study how agents learn from vague goals, self-evaluate actions, and persist improvements, revealing that current agents overfit to proxy feedback and fail to convert self-judgment into lasting capability gains.
Anthropic CEO Dario Amodei argues for pacing frontier AI development, proposing resident third-party evaluators, capability-based safety thresholds, and incremental international coordination, while OpenAI and others respond with partial commitments, raising questions about enforcement, fairness, and whether voluntary measures can truly constrain recursive self-improvement risks.
Anthropic CEO Dario Amodei publishes 'We Must Pace the Frontier' calling for slowing frontier AI development due to risks from recursive self-improvement, announces embedded external evaluators with employee-level access, and outlines a three-step plan for democratic and global coordination, prompting immediate support from Musk, Hugging Face, and OpenAI.
openJiuwen introduces a dual-dimensional Recursive Self-Improvement (RSI) framework that enables AI agents to automatically optimize both their tooling (Harness) and deliverables (research papers, algorithms) on the WorkSwarm platform, with compute-affinity scheduling on Ascend NPUs cutting latency and resource usage, validated by SWE-bench pass-rate gains from 61% to 87%.
OpenAI publishes internal data on how AI agents accelerate research, introducing a three-layer measurement framework—usage, tasks, and results—showing median researchers spend $600/day on tokens, agents now handle 3.1 workdays per human day, task delegation spans six R&D stages but high-level planning remains human-led, and over half of successful 4-8 hour tasks still require human intervention.
MetaRSI-v1 introduces a unified meta-recursive self-improvement architecture that applies recursive self-improvement to the improvement process itself, unifying Model-RSI, Data-RSI, and Harness-RSI via a Loop Kernel with horizontal and vertical orchestration, four coordinating agents, and five laws, demonstrating 10.9-point average gains on a 3B model and 7.3-point gains on frontier models like GPT-5.6 and Claude Opus 5.
OpenAI publishes internal data showing AI agents now contribute 3.1 workdays per human researcher day, with median researchers spending $600 daily on inference, while acknowledging complex tasks still require human intervention and safety restrictions caused GPU usage shifts.
This weekly roundup covers OpenAI's GPT-6 Astra debut with recursive self-improvement, Google's cost-efficient Gemini 3.8 Flash, Anthropic's Fable 5.1 scientific applications, Tencent's WorkBuddy agent platform, Hyper3D's WorldGen scene generation, China's office agent competitive landscape, and expert insights on embodied AI limits and industry structure.
OpenAI Chief Scientist Jakub Pachocki argues that AI progress is driven by compute scaling, creating systems we cannot fully understand; alignment techniques are lagging, chain-of-thought monitoring is failing, and recursive self-improvement looms, urging coordinated slowdown and safety standards before deploying superintelligent systems.
OpenAI Codex lead Tibo reveals why next-gen AI agents will make skills and memory management disappear, how Ultra Fast mode restores real-time flow, why ChatGPT and Codex are merging into a personalized AGI, and how recursive self-improvement now extends from model training to CUDA kernels and infrastructure.
Mechanist automates mechanistic interpretability research by generating hypotheses, running causal interventions on language models, discovering distinct attention heads for belief-state reasoning, and enabling targeted steering that improves both reasoning and biological sequence generation.
The article examines recursive self‑improvement (RSI) by detailing a joint research effort that used an AI agent to autonomously develop the 27‑billion‑parameter iCoder‑27B industrial coding model, presenting benchmark gains, failure analyses, and a nuanced view of RSI versus lossy self‑improvement.
Recuris introduces a compact task‑state‑driven memory architecture and gated recursive self‑improvement, enabling agents to use and evolve memory more reliably and delivering large, consistent gains from 3B open‑source models up to frontier models such as Claude Opus 5 across multiple long‑horizon benchmarks.
This article analyzes Agent self-evolution across three levels—event, effective, and reliable—detailing update paths, learning signals, update objects, the learning loop, verification methods to prove real capability growth, and current boundaries of recursive self-improvement.
The talk introduces RSIBench‑Data, a benchmark that transforms the problem of agents merely “gaming scores” into a controlled scientific experiment, enabling agents to diagnose failures, design informative data experiments, and achieve verifiable recursive self‑improvement, with early results showing a jump in checkpoint success rates from 8% to 22%.
Google is intensifying its AI arms race by urging founder Sergey Brin to champion Recursive Self‑Improvement, reshaping Gemini development, reorganizing DeepMind leadership, and confronting TPU resource limits, as the company seeks to accelerate AI research while balancing commercial pressures and internal structural challenges.
This article surveys recent AI research on self‑evolving and self‑improving systems, defines a three‑layer taxonomy (Artifacts, Harness, Model), reviews concrete implementations from OpenAI, Anthropic, Tencent, MiniMax, and others, and outlines open research directions and challenges.
The Frontis-MA1-35B model and the OpenMLE suite, released by a Tsinghua‑affiliated team, demonstrate how execution feedback can be fed back to the model that proposes modifications, achieving significant gains on MLE‑Bench Lite and showing a concrete step toward recursive self‑improvement in AI research.
Jeff Dean, together with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, founded Discovery Loop, a public‑benefit corporation that seeks to replace manual AI prompting with an automated experimental loop—proposing, running, and learning from thousands of ML experiments to accelerate discovery across scientific domains.
The article introduces OpenMLE, an open‑source full‑stack system for recursive self‑improvement (RSI) research, showing how a 35B Frontis‑MA1 model improves its Medal Average from 39.39% to 71.21% on MLE‑Bench Lite, surpasses GPT‑5.5+Codex, and details the mechanism hierarchy, task‑curation gym, trainable evolution operators, and experimental evidence that training and search gains combine additively.
The article examines how Xiaohongshu’s large‑language model dots‑note‑3.0 achieved a flawless 42‑point score at IMO 2026 by repeatedly generating, verifying, and refining natural‑language proofs, and discusses how this recursive self‑criticism signals a shift toward agents that can audit and improve their own reasoning for complex, real‑world problems.
In an eight‑day, fully automated experiment, Weco’s AIDE² system ran 100 outer‑loop iterations without updating model weights, rewrote its own harness, produced two standout versions (AIDE₄₇ and AIDE₈₅) that outperformed human‑tuned baselines on three unseen benchmarks, and cut reward‑hacking rates from 63% to 34%, providing the first Level‑1 evidence of recursive self‑improvement.
Rapid advances in AI agents have exposed three intertwined contradictions—memory, reasoning, and self‑improvement paradoxes—where more data hurts decision quality, engineering scaffolds create new failures, and reliable evaluation becomes a structural bottleneck, as detailed through recent industry systems and academic studies.
The article reviews recent AI agent research, exposing three structural paradoxes—memory, reasoning, and evolution—each illustrated with concrete systems, benchmarks, and safety studies, and argues that only coordinated progress across all three dimensions can yield truly mature, self‑improving agents.
XYZ AI Lab unveiled two Deep Search agents—35B‑parameter Aquila‑mini and 397B‑parameter Aquila‑pro—that set SOTA scores on seven public benchmarks, while demonstrating an AI‑4‑AI paradigm where hundreds of agents collaboratively drive a full‑stack research and improvement loop under human‑defined goals and verification.
The article presents a detailed technical survey of Harness Engineering, explaining how it extends classic agent architectures with workflow design, persistent state, and sub‑agent orchestration, and traces its evolution through ACE, MCE, and Meta‑Harness as a practical pathway toward recursive self‑improvement.
The article surveys recent research on harness engineering—software layers that orchestrate large language models—and examines how these layers can drive recursive self‑improvement, outlining design patterns, optimization techniques, evolutionary search, and the remaining technical challenges.
The article examines recursive self‑improvement in AI, arguing that a well‑designed harness—responsible for workflow orchestration, context management, and tool integration—is as crucial as model intelligence, and outlines design patterns, meta‑engineering approaches, evolutionary search methods, and the remaining challenges for truly autonomous AI systems.
The article surveys recent research on Harness engineering—systems that orchestrate model reasoning, tool use, context management, and evaluation—and examines whether recursive self‑improvement (RSI) will first emerge in model weights or in the surrounding Harness, while outlining design patterns, optimization strategies, and open challenges.
Anthropic’s latest blog reveals that its Claude models now write over 80 % of its code, have tripled productivity, and dramatically improve success rates, suggesting a recursive self‑improvement trajectory that could reshape AI development and prompts the company to call for a verifiable slowdown.
Two former Anthropic researchers have founded Mirendil, raising $200 million at a $1 billion valuation to build a recursive self‑improvement AI platform that lets scientists create and iterate their own models, positioning the startup against big‑tech AI moats.
DeepMind co‑founder Shane Legg and a team of researchers released a 57‑page report that outlines four possible routes from artificial general intelligence to superintelligence, analyzes scaling, paradigm shifts, recursive self‑improvement and multi‑agent collaboration, and identifies six potential bottlenecks such as data limits and economic constraints.
DeepMind researchers outline three core concepts, enumerate digital intelligence’s innate advantages, detail the theoretical limits of ASI, and propose four plausible routes from human‑level AGI to superintelligence while identifying six potential walls that may impede or stop that transition.
Recursive Superintelligence unveiled an open‑source system that automates the AI research loop, achieving state‑of‑the‑art results on three distinct benchmarks—NanoChat autoresearch, NanoGPT speedrun, and SOL‑ExecBench—while illustrating the practical progress toward recursive self‑improvement warned about by Anthropic.
Anthropic’s internal report reveals that Claude now writes over 80% of its code, engineers’ productivity has jumped eight‑fold, and recursive self‑improvement is becoming a real risk, prompting the company to urge a worldwide slowdown of advanced AI development.
Anthropic’s internal data shows AI models are rapidly self‑improving—Claude now writes over 80% of its code, boosts engineer productivity several‑fold, and speeds up tasks dramatically—prompting a pause proposal and three possible future trajectories for AI development.
Anthropic’s internal report reveals that its Claude model now writes over 80% of the company’s code and boosts engineer output eight‑fold, providing concrete evidence of rapid recursive self‑improvement and prompting the firm to urge a worldwide slowdown of frontier AI research while outlining three possible future scenarios.
Anthropic’s internal report shows that its Claude model now writes over 80% of merged code and boosts engineer output eightfold, evidencing rapid recursive self‑improvement, while the company urges a worldwide pause on large‑model research and discusses potential future scenarios, risks, and the need for coordinated governance.
The article examines how recent advances in large language models, AutoML, and evolutionary algorithms are pushing AI toward recursive self‑improvement, outlines current capabilities and limitations, and discusses the technical, economic, and safety challenges that still prevent a fully autonomous intelligence explosion.
Anthropic co‑founder Jack Clark predicts a 60% chance that by the end of 2028 AI systems will be capable of recursive self‑improvement, citing rapid progress on benchmarks such as CORE‑Bench, PostTrainBench, SWE‑Bench, METR, and emerging capabilities in kernel design, agentic coding, and AI‑to‑AI management.
Jack Clark, Anthropic’s co‑founder, argues that based on a sweep of public AI benchmarks—including CORE‑Bench, PostTrainBench, MLE‑Bench, SWE‑Bench and METR—there is roughly a 60% probability that recursive self‑improvement will emerge by the end of 2028, raising profound technical and alignment challenges.
Researchers introduce the Self‑Taught Optimizer (STOP), a scaffolding‑based framework that lets large language models iteratively improve their own code without altering model weights, demonstrating superior performance on tasks like LPN, exploring diverse strategies such as beam search and genetic algorithms, while also highlighting security risks like sandbox bypass and reward hacking.
The article examines Nvidia’s claim that AI can now write software and build an “AI factory,” analyzes OpenAI’s emerging o‑series models that purportedly achieve recursive self‑improvement, and surveys community reactions ranging from excitement to safety concerns about a potential AI “game over.”
