2026 Research Update
2026 in RSI Research: New Breakthroughs and Papers
A follow-up to the Recursive Self-Improvement: A Literature Survey, covering only what's new since then โ named techniques, surveys, and AI-automating-AI-research benchmarks published between January and September 2026, plus what frontier labs have said about self-improvement in production. Same rule as the base survey: confidence levels are stated explicitly, and nothing here is presented as confirmed unless it was actually found at a primary source.
How to read the confidence tags below: HIGH means a paper or announcement page was found directly at arxiv.org or an official lab domain. MEDIUM means it was found via search snippets or a reputable secondary outlet but not read in full, or came from a lab spokesperson quote rather than a paper. LOW means it surfaced only in an AI-generated search summary or a single aggregator and could not be independently confirmed โ treat those items as leads to verify, not facts to cite. This survey relied on web search rather than opening every PDF, so even HIGH-confidence entries below are abstract-level, not full-text reads.
Opening
2026 Timeline
New Named Techniques
New Surveys
AI Automating AI Research
Lab Announcements
What Changed Since 2025
Researcher Focus
Practitioner Focus
Confidence Notes & Gaps
Full Reference List
Why 2026 Is Different
The base literature survey covered a field that, through late 2025, was still small enough for one person to read chronologically: roughly two dozen papers across four decades. That's no longer true. A July 2026 survey (arXiv:2607.07663) that reviewed the RSI-adjacent literature reports around 1,250 relevant papers posted between 2024 and 2026 โ and 74% of that corpus is from 2026 alone, with output running at roughly 500 papers per quarter by Q2. RSI also got its own dedicated workshop at ICLR 2026. This piece does not try to catalog all 1,250 papers. It follows the same selection standard as the base survey โ named techniques, formal surveys, and AI-automating-AI-research systems specifically, not every paper that uses the phrase "self-improving" โ and adds what qualifies from 2026.
2026 RSI Timeline
Every dated item below is discussed in the sections that follow, in the same order. One panel this time โ nine months is short enough not to need a scale break.
Named technique
Survey or benchmark
Lab announcement
Jan 2026
Digital Red Queen โ Sakana AI ร MIT
Open-ended adversarial coevolution of LLM-authored programs in a Core War sandbox.
Feb 2026
DARWIN
Dynamic agentically-rewriting self-improving network โ title and ID confirmed, details unread; verify before citing specifics.
Feb 2026
International AI Safety Report 2026 & METR timelines note
Government-convened safety report, plus METR's "simpler AI timelines model" projecting ~99% AI R&D automation around 2032.
Mar 2026
PostTrainBench & METR's "Measuring AI R&D Automation"
Benchmarks tracking how much of post-training and broader AI R&D can already be automated.
Apr 2026
AutoResearchBench
Literature-discovery benchmark spanning roughly 3 million arXiv papers.
May 2026
AutoResearchClaw & ARC-Bench
Multi-agent autonomous-research system claiming a 54.7% improvement over The AI Scientist-v2 on its own benchmark.
May 2026
AutoLLMResearch
Training research agents to automate LLM experiment configuration.
Jun 2026
Red Queen Gรถdel Machine (RQGM)
Co-evolves the self-improving agent and its evaluator under shifting utility criteria โ direct successor line to the Darwin Gรถdel Machine.
Jun 2026
AutoLab, "Act As a Real Researcher," ResearchClawBench
A cluster of long-horizon and full-lifecycle research-automation benchmarks released within weeks of each other.
Jun 2026
Sakana Fugu launches; Sakana RSI Lab announced
Fugu is a multi-agent orchestration system marketed as "a model" โ its fit inside the RSI literature proper is uncertain, flagged below.
Jul 2026
The RSI field-mapping survey (arXiv:2607.07663)
The 1,250-paper corpus review that anchors this article's "why 2026 is different" framing.
Jul 2026
AlphaEvolve reaches general availability
Shipped on Google Cloud's Gemini Enterprise Agent Platform, upgraded to Gemini 2.5 Pro โ no separately-named successor found.
Aug 2026
Anthropic: automated alignment researchers
Claude-driven automated researcher applied to Anthropic's own alignment evaluation pipeline โ the strongest primary-source lab result in this update.
Sep 2026
"The Last AI Built by Humans" & "Recursive Criticality of AI Self-Improvement"
Two further 2026 surveys/position papers, both confirmed to exist, contents not independently read.
New Named Techniques
Red Queen Gรถdel Machine (RQGM) HIGH
The clearest direct successor to the Darwin Gรถdel Machine line covered in the base survey. Where DGM validates self-modifications against a fixed benchmark, RQGM co-evolves the agent and the evaluator judging it, letting evaluation criteria shift at defined "epoch" boundaries. The stated goal is closing a known failure mode in earlier self-improving-agent loops: agents learning to game a static benchmark rather than genuinely improving. Applied to coding, paper-writing, and proof-generation tasks.
๐ arXiv:2606.26294
๐บ No verifiable dedicated video found
Digital Red Queen HIGH
Not a task-performance system like most of the base survey's techniques โ an open-endedness testbed. Competing LLM-authored programs coevolve adversarially inside a Core War-style sandbox, with complexity emerging from competitive pressure rather than an explicit objective. Framed by the authors as infrastructure for studying open-ended RSI dynamics directly, separate from the "improve at task X" framing that dominates the rest of this literature.
๐ arXiv:2601.03335
๐บ No verifiable dedicated video found
DARWIN (Dynamic Agentically Rewriting Self-Improving Network) MEDIUM
Title and arXiv ID surfaced directly in search results, but the abstract itself was not opened and read during this research pass. Named similarly enough to the unrelated Darwin Gรถdel Machine that the two should not be conflated. Included here as a lead worth following up, not as a confirmed description of what the system actually does.
๐ arXiv:2602.05848 (unread โ verify before citing)
๐บ No verifiable dedicated video found
AlphaEvolve: general availability, no renamed successor HIGH
Rather than a new named technique, AlphaEvolve (covered in the base survey) reached general availability on Google Cloud's Gemini Enterprise Agent Platform around July 2026, upgraded to run on Gemini 2.5 Pro. No differently-named DeepMind successor to AlphaEvolve was found in this research pass โ worth noting explicitly so it isn't assumed to exist.
๐ No standalone paper โ product/platform announcement
๐บ No verifiable dedicated video found
Sakana Fugu / Fugu Ultra โ relevance uncertain LOW relevance
Confirmed to exist and to have launched, alongside a separately-announced "Sakana AI RSI Lab" research initiative. But Fugu is marketed as a multi-agent orchestration system positioned as "a model," and it is unclear from available material whether it implements genuine self-improvement in the AlphaEvolve/Gรถdel-Machine sense or is adjacent branding. Listed here for completeness, flagged rather than treated as confirmed RSI.
๐ Official page not independently re-verified here
๐บ No verifiable dedicated video found
New Survey Papers
Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops HIGH
The corpus-scale survey this article leans on for its "why 2026 is different" framing: reviews roughly 1,250 RSI-adjacent papers from 2024โ2026 and taxonomizes by what improves (behavior, policy, evaluator, or the research process itself) and by degree of loop closure. Its central distinction โ convergent "bounded self-refinement" already in industrial use, versus open-ended RSI still bounded by grounding, collapse dynamics, and compute โ is the organizing idea for this whole follow-up piece.
๐ arXiv:2607.07663
๐บ No verifiable dedicated video found
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement HIGH
Confirmed to exist via both its arXiv page and an alphaXiv mirror. Content was not read in full during this research pass โ the title alone suggests a position paper on what would distinguish "genuine" RSI from the bounded self-refinement now common in industry, which tracks the same distinction as the July survey above, but this should be confirmed by reading the paper directly before citing its argument specifically.
๐ arXiv:2609.11873
๐บ No verifiable dedicated video found
Recursive Criticality of AI Self-Improvement HIGH
Confirmed via its arXiv abstract page and a Hugging Face paper page; not read in full. The title suggests a formal treatment of self-improvement dynamics near a "critical" threshold โ potentially connecting to Yudkowsky's 2013 "returns on cognitive reinvestment" framing from the base survey's foundational-theory section, but that connection is this article's inference, not a confirmed claim from the paper itself.
๐ arXiv:2609.00137
๐บ No verifiable dedicated video found
International AI Safety Report 2026 HIGH
Confirmed to exist as a large, formally convened cross-institution report. Likely contains material relevant to RSI risk given the report series' scope in prior years, but its specific RSI content was not independently read here โ cited as a pointer for researchers who want the policy-facing companion to the technical surveys above.
๐ arXiv:2602.21012
๐บ No verifiable dedicated video found
ICLR 2026 Workshop: AI with Recursive Self-Improvement HIGH
Not a paper โ a dedicated workshop, confirmed to exist on the official ICLR 2026 program. Its existence is itself a data point: the subfield formalized enough in 2026 to warrant its own venue, distinct from being folded into general agents or alignment tracks as in prior years.
๐ ICLR 2026 program page (not re-fetched for this article)
๐บ No verifiable dedicated video found
AI Automating AI Research: The 2026 Benchmark Wave
The single biggest shift from the base survey's "AI Automating AI Research" section isn't one breakthrough system โ it's how many benchmarks for exactly this capability launched within a few months of each other in 2026. Read as a group, they suggest the field converged on "can an agent do the whole research loop" as the open question worth measuring, rather than continuing to ship one-off pipelines like 2024's AI Scientist.
AutoResearchClaw & ARC-Bench HIGH
An open-source multi-agent autonomous-research system built around structured multi-agent debate, a self-healing executor, verifiable reporting meant to resist hallucinated results, and seven distinct human-in-the-loop intervention modes. Its companion benchmark, ARC-Bench, spans 25 ML topics plus a 20-topic extension into physics/biology/statistics. The paper claims a 54.7% improvement over The AI Scientist-v2 on that benchmark โ a claim made by the same team that built the benchmark, so treat the specific margin as self-reported pending independent replication.
๐ arXiv:2605.20025
๐บ No verifiable dedicated video found
PostTrainBench: Can LLM Agents Automate LLM Post-Training? HIGH
Directly targets a question the base survey's MLE-bench and AutoML-Agent entries only touched indirectly: not "can an agent do ML engineering" but specifically "can an agent post-train a model as well as a human team." A commonly cited figure โ that AI-run post-training as of March 2026 achieves roughly half the uplift of human-run post-training โ comes from a secondary summary and should be confirmed against the paper's actual tables before being repeated as precise.
๐ arXiv:2603.08640
๐บ No verifiable dedicated video found
AutoLLMResearch, AutoLab, "Act As a Real Researcher," ResearchClawBench, AutoResearchBench HIGH
Five separate benchmark/system papers, all confirmed to exist, released within a roughly ten-week window. Together they cover: automating LLM experiment configuration (AutoLLMResearch); long-horizon research-and-engineering tasks for frontier models (AutoLab); a full research-lifecycle benchmark suite ("Act As a Real Researcher"); an end-to-end autonomous science benchmark related to the AutoResearchClaw family (ResearchClawBench); and literature discovery across roughly 3 million arXiv papers (AutoResearchBench). None of the five were read past their arXiv listing for this article โ they're grouped here as a cluster worth surveying together rather than described individually, since the overlap between several of them (especially ResearchClawBench vs. ARC-Bench) was not resolved in this research pass.
METR: Measuring AI R&D Automation HIGH
METR's direct 2026 follow-on to RE-Bench and MLE-bench from the base survey, explicitly measuring how much of real AI R&D work has become automatable rather than benchmarking a specific system's score. The natural companion read to METR's February 2026 blog note projecting roughly 99% AI R&D automation around 2032, from a simplified timelines model.
What Frontier Labs Actually Said in 2026
Automated Researchers Can Reliably Mitigate Alignment Failures HIGH
The strongest primary-source item in this entire update. Anthropic reports applying a Claude-driven automated researcher to ten categories of alignment failure โ deception, sycophancy, jailbreak susceptibility, and others โ and closing what they describe as a substantial portion of the safety gap to their fully-trained production model. The reported sample efficiency figure (roughly 2,400 training examples, claimed around 15,000ร more efficient than full production alignment training) comes directly from Anthropic's own page, not a secondary summary, which is why this is flagged HIGH rather than MEDIUM like most other 2026 entries here.
๐ anthropic.com research page (not re-fetched for this article โ verify current URL)
๐บ No verifiable dedicated video found
Anthropic's disclosures on Claude's role in its own R&D MEDIUM
The TechCrunch article itself is a real, directly-found piece of reporting. But two specific numbers attached to this story in wider circulation โ that Claude leads "26%" of Anthropic's AI R&D and runs across "30,000 concurrent agents," and separately that Claude writes over 80% of Anthropic's merged code as of May 2026 โ traced back only to a secondary aggregator and an AI-generated search summary respectively, not to an Anthropic primary source verified in this research pass. Treat the qualitative claim (Anthropic uses Claude substantially in its own R&D) as credible and the specific percentages as unverified.
๐ TechCrunch, Aug 28 2026 (not re-fetched for this article)
๐บ No verifiable dedicated video found
OpenAI's "automated AI research intern" goal MEDIUM
OpenAI reportedly set a goal of shipping something like an automated AI research intern by around September 2026. This research pass found the goal reported by third parties, not an OpenAI product page or paper confirming it shipped, and did not confirm whether the September target was met. A separately circulating claim referencing a model called "GPT-5.6 Sol" helping post-train a smaller model could not be verified as a real model name and should not be repeated without independent confirmation.
๐ Secondary reporting only โ no primary OpenAI source confirmed
๐บ No verifiable dedicated video found
METR Frontier Risk Report: cross-lab misalignment pilot HIGH
Describes a February 2026 pilot assessing misalignment risk specifically from AI agents used inside frontier labs' own workflows, with Anthropic, Google, Meta, and OpenAI all reported as participants. Not itself an RSI technique paper, but directly relevant context: it's an attempt to measure risk from the same "AI helping build AI" pattern the rest of this article covers, rather than risk from AI deployed to end users.
๐ METR blog, May 19 2026
๐บ No verifiable dedicated video found
What Actually Changed Since the Base Survey
Convergence: "Bounded self-refinement" โ agents that improve their own scaffolding, prompts, or evaluators inside a fixed outer loop โ moved from research curiosity to something at least one frontier lab (Anthropic) reports using in its own production alignment pipeline. That's a real qualitative shift from the base survey's era, where every technique was still a research paper, not a production workflow.
Disagreement: Whether anything in 2026 constitutes "genuine" open-ended RSI โ as opposed to bounded self-refinement dressed up in RSI language โ is contested even within the 2026 literature itself. The July survey's own taxonomy exists specifically because the field couldn't agree on this before 2026; "The Last AI Built by Humans" (Sept 2026) reads, from its title alone, like an argument that genuine RSI still hasn't arrived. Media coverage (MIT Technology Review, Aug 18 2026) leans the same direction: labs consider it plausible and near, but progress is contested, not settled.
Most consequential single data point: not a technique, but the benchmark-proliferation cluster in the AI-automating-AI-research section. Five to six independent teams built overlapping full-lifecycle research-automation benchmarks within about ten weeks of each other in spring 2026. That kind of convergent effort, more than any individual score, is the clearest sign the field now treats "can AI run the whole research loop" as the central open question โ which is exactly the frame the researcher-focus section below builds on.
Researcher Focus: Where the Open Problems Actually Are
If you're deciding what to actually work on rather than just read about, four directions stand out from this year's output:
1
Evaluator co-evolution and gaming resistance. RQGM (arXiv:2606.26294) exists specifically because static-benchmark gaming was a real failure mode in the Darwin Gรถdel Machine line. If you want to extend that work, the open question is whether evaluator co-evolution itself becomes gameable at a higher level โ read RQGM's epoch-boundary mechanism closely before assuming it's a solved problem rather than a deferred one.
2
Benchmark consolidation, not more benchmarks. At least five overlapping full-lifecycle research-automation benchmarks shipped in a ten-week window this spring, and this research pass could not even resolve how ResearchClawBench differs from ARC-Bench. A researcher who reads all five carefully and publishes a genuine comparison โ where they actually disagree, not just that they exist โ would be filling a real gap rather than adding a sixth benchmark.
3
Independent replication of self-reported results. AutoResearchClaw's 54.7%-over-AI-Scientist-v2 claim, and the PostTrainBench "half the human uplift" figure, are both self-reported by teams with a stake in the result (the first on their own benchmark, the second reported only via a secondary summary). Replicating either independently โ or showing they don't hold up โ is more valuable right now than proposing a new system.
4
The genuine-RSI-vs-bounded-refinement distinction, made operational. The July survey's taxonomy and September's "Last AI Built by Humans" are both circling the same question from different angles: what would it actually take to distinguish real open-ended RSI from bounded self-refinement with good PR? Nobody in this literature has yet proposed a measurable test for that distinction. That's arguably the single most valuable open theoretical problem visible in the 2026 output.
Practitioner Focus: What's Actually Usable Right Now
For teams deciding what to adopt rather than what to research, the 2026 evidence supports a narrower conclusion than the headlines suggest:
1
Bounded self-refinement is production-ready for narrow, well-scoped loops. Anthropic's own alignment-researcher pipeline is the clearest evidence: automating a specific, well-defined evaluation task (checking for known categories of alignment failure) with a fixed outer loop and human oversight of the loop itself, not the agent inside it.
2
Full autonomous-research pipelines are not there yet for unsupervised use. Every AI-automating-AI-research system in this update โ AutoResearchClaw, AutoLab, ResearchClawBench, and the rest โ is still benchmark-stage work, evaluated against other AI systems' output rather than deployed to replace a human research workflow end to end. PostTrainBench's own headline finding (AI-run post-training at roughly half of human uplift, itself an unverified figure pending direct confirmation) is consistent with "useful assistant, not yet a replacement."
3
Treat lab self-reported adoption metrics as directional, not precise. "Claude leads 26% of R&D" and "30,000 concurrent agents" make for a compelling headline but were not independently confirmed here. If you're making a budget or staffing decision based on how far ahead frontier labs already are, get the primary-source number, not the aggregator's paraphrase.
4
The risk conversation is now running in parallel with adoption, not behind it. METR's cross-lab misalignment pilot on internally-used AI agents launched the same season AlphaEvolve reached general availability. Practitioners adopting agent-driven internal tooling in 2026 are doing so alongside active, lab-backed risk assessment work on exactly that pattern โ worth tracking METR's output specifically if your organization is building similar internal agent workflows.
Ready to work hands-on? A Practitioner's Primer on RSI covers what to read first, real open-source repos to run, how to design a new experiment, compute and funding resources, and where to publish.
โ ๏ธ Confidence Notes and Gaps
This article was built entirely from web search results, not full-text paper reads, so every entry above is at most abstract-level even where marked HIGH confidence. Before citing specific numbers from this piece in your own work, go read the primary source directly โ especially arXiv:2607.07663, arXiv:2606.26294, arXiv:2605.20025, and Anthropic's automated-alignment-researcher page, which carry the most load-bearing claims here. Items explicitly flagged LOW or MEDIUM (DARWIN's details, the specific Anthropic R&D percentages, "A-Evolve-Training," and "GPT-5.6 Sol") should be treated as unconfirmed leads, not facts, until independently verified. As with the base survey, YouTube could not be directly fetched in the environment this article was researched in, so no video links are claimed for any 2026 item โ every entry above either has no video or wasn't checked for one.
๐ Full Reference List
Recursive Self-Improvement: A Literature Survey (base article this piece extends)
Digital Red Queen โ arXiv:2601.03335
DARWIN โ arXiv:2602.05848
International AI Safety Report 2026 โ arXiv:2602.21012
METR timelines note โ Feb 10, 2026
METR: Measuring AI R&D Automation โ arXiv:2603.03992
PostTrainBench โ arXiv:2603.08640
AutoResearchBench โ arXiv:2604.25256
AutoLLMResearch โ arXiv:2605.11518
AutoResearchClaw & ARC-Bench โ arXiv:2605.20025
AutoLab โ arXiv:2606.05080
"Act As a Real Researcher" โ arXiv:2606.07462
ResearchClawBench โ arXiv:2606.07591
Red Queen Gรถdel Machine โ arXiv:2606.26294
METR Frontier Risk Report โ May 19, 2026
RSI field-mapping survey โ arXiv:2607.07663
Recursive Criticality of AI Self-Improvement โ arXiv:2609.00137
The Last AI Built by Humans โ arXiv:2609.11873