Two open-weight models beat GPT-6 Astra on agent benchmarks without changing a single weight. Here is what actually improved.

A paper submitted to arXiv on 14 September 2026 reports that two open-weight models, GLM-5.3 from Z.ai and Kimi-K3 from Moonshot AI, wrapped in a self-exploration loop that never updates a model weight, outscore GPT-6 Astra and Claude Opus 5 on two agent benchmarks. The paper is called "RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments," and a Discover AI video covering it drew the reasonable question of whether "recursive self-improvement" is the right name for what it does. We checked the paper's own text, its project page, its repo, and the field around it, supportive and competing sources both. The numbers hold. The naming is contestable. The useful part is neither of those things.

#The result, exactly as the paper states it

The results sentence, read directly from the paper's full text: "Using GLM-5.3 and Kimi-K3, it reaches 78.98 on OSWorld 2.0 and 84.82 on ALE, exceeding the reported GPT-6 Astra scores by 6.38 and 2.56 percentage points, respectively, and also scoring above Claude Opus 5 on both benchmarks." ALE is Agent's Last Exam. OSWorld 2.0 and OSWorld-v2 are the same benchmark, written both ways inside the paper itself; the original OSWorld is an established computer-use benchmark with a live site, so the v2 is a real successor and not an invented name. The abstract states the same finding more generally: the framework enables "Kimi-K3 and GLM-5.3 to outperform frontier closed-source models including GPT-6."

The authors are Sibo Zhu, Shicheng Fan, Xinyue Wang, Wenyi Wu, Kun Zhou (corresponding author and project lead), and Biwei Huang, affiliated with Aether AI, UC San Diego, and the University of Illinois Chicago; the paper's own author markers show four of them did the work as interns at Aether AI. Code is on GitHub, and the project page links both.

One thing that gets lost when the headline is "open-weight beats closed": the open-weight status of both models is real and independently confirmed, not just repeated from the video. A third-party piece from July 2026 covers Kimi K3 and GLM 5.2 explicitly as open-weight releases that "compete with the closed frontier," and GLM-5.3, released in August, is the same base model as 5.2 with post-training gains, per independent coverage. So the claim is not that a cheaper API beat a pricier one. It is that weights you can download, driven by a loop you can read, beat models you cannot inspect.

#What the loop does, and what it does not do

The mechanism, from the paper's own abstract and body: a "training-free multi-agent framework for recursive self-improvement through autonomous memory construction." Three agents. In the paper's words, "the curriculum agent decides what to explore next, the actor agent interacts with the environment and updates the memory, and the verifier agent grounds observed outcomes with environment feedback, allowing the system to progressively uncover and consolidate reusable causal structures." Exploration runs "broad-then-deep": parallel broad exploration to discover the environment's structure, then focused deep exploration of hard cases, hidden constraints and boundary conditions. What comes out is a memory of "reusable causal relationships between actions, conditions, and consequences." That memory is then frozen and reused. The model's parameters never change.

That last sentence is the whole result. The improvement from a frontier-lagging score to a frontier-beating one came from what the system knew about the environment, not from what the model could reason about in general. Read plainly, the mechanism separates two kinds of capability: general reasoning, which frontier models already have in abundance, versus environment-specific operational knowledge, which API does what, which action changes which state, where the edge cases are. The implication of the result is that the second category, not the first, is where a large share of real-world agent failure comes from. That is our reading of what the method and the numbers together imply, not a sentence lifted from the paper, and it is the distinction worth carrying regardless of any single benchmark score.

#Where the skepticism is fair

The video's presenter argues that this is not "strong" recursive self-improvement, because the system never modifies its own harness, its verification procedure, or its exploration protocol. It accumulates memory inside a fixed loop. He relabels it as bounded, anchored self-improvement through an accumulated memory representation. That is his editorial judgment, not a claim the paper makes about itself, and we hold it as such.

It is also well grounded. A separate academic paper, "Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops," submitted in July 2026 and revised in September by Mingguang Chen, Licheng Wang and Bo Qu, uses the same bounded-versus-strong taxonomy the presenter reaches for, independently of this video or this paper. The distinction maps onto published vocabulary. The competing view here is a competing definition of what RSIAgent's result should be called, not a competing claim that its numbers are wrong or its method does not work. Those are different kinds of disagreement, and only the second would change what an operator should do.

The presenter's second critique is sharper and, on our read of the paper, also fair. The paper describes its process as "a form of causal discovery" and uses "causal relationships" and "causal structures" throughout. Its Limitations section names four things: the added test-time compute cost of exploration, the dependence on exploration budgets, stopping policies and memory quality, the risk that a model-based verifier makes wrong judgments that propagate into memory, and that the experiments do not fully isolate each component's contribution. None of the four engages with the formal apparatus the phrase "causal discovery" usually implies in causal inference: no causal graphs, no intervention variables, no confounding, no counterfactual estimation anywhere in the fetched text. The language reaches further than the method formally supports, and the paper's own limitations section does not say so. That is a fair reading, not an invented one.

#What nobody has checked yet

We found no independent critique, replication, or extension of RSIAgent specifically. That is almost certainly because the paper was two days old at the time we looked, not because it has passed scrutiny. The repo was created on 13 September, one day before the arXiv submission, an ordering we could not explain from anything we fetched and are naming rather than assuming away.

What we did find is that the underlying premise is not this group's alone. FORGE, "Self-Evolving Agent Memory With No Weight Updates via Population Broadcast," was submitted in May 2026 by a fully separate group and asks the same question in its abstract: "Can LLM agents improve decision-making through self-generated memory without gradient updates?" Lilian Weng's July 2026 post on harness engineering names "file system as persistent memory" as a design pattern and gives self-improving harnesses their own section, two months before RSIAgent existed and therefore not citing it. Two independent groups working adjacent territory reached the same "memory over weights" premise before this result. That is real corroboration that the idea is not one lab's invention. It is not yet evidence that the field has converged, and no outside party has had a chance to stress-test this specific paper's numbers.

#What an operator does with this

If you are building or buying an agent that has to operate inside a specific system, the finding to carry out of this paper is the split it draws, not the leaderboard position. A capable model failing in your environment is often not failing at reasoning. It is failing at knowing which of your APIs does what, which action has which side effect, and where your edge cases live. RSIAgent's result is that filling in that second kind of knowledge, through exploration and a memory that persists, moved two open-weight models past the closed frontier on two benchmarks without any change to the models themselves. Whether or not "recursive self-improvement" is the right name for that, the lever it points at is one you already control: what your agent is allowed to learn about your environment before it is asked to act in it, and where that learning is kept.

Hold the rest as open. The numbers are the paper's own and have not been reproduced by anyone else. The causal language outruns the method. The repo timing is odd and unexplained. None of that erases the result. It sets the terms for believing it.