🔍 Read the full analysis: Terminal-RSI: Separating AI Self-Improvement From Human Input on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A preprint from Tencent’s Hy Frontier lab and four universities reports that a 27-billion-parameter terminal agent improved on several coding and computer-use benchmarks through five training stages. The stages, rewards, and checkpoint choices were set by researchers, so the results do not establish unsupervised self-improvement; long-horizon task completion remained limited.
A 26-page preprint called Terminal-RSI reports that a command-line AI agent improved on several task benchmarks through a five-stage training process designed and managed by researchers. The work, attributed to Tencent’s Hy Frontier lab and four universities, does not show an AI rewriting and retraining itself without human involvement: people selected the training stages, rewards, learning rates and checkpoints.
Using the open Qwen3.8-27B model as its base, the researchers trained a terminal agent on a pool of 3,000 tasks. The sequence alternated between learning from verified successful attempts and reinforcement learning. On Terminal-Bench 4.0, the reported Mean@3 score rose from 0.00% to 19.84%. Mean@3 is the arithmetic mean of three evaluation runs, not the best result from three tries.
The paper also reports an equal-weight average across four Terminal-Bench suites rising from 20.68% to 50.58%. That calculation excludes the long-horizon test and SWR100, and gives each included suite the same weight regardless of its size. On the long-horizon benchmark, the final model completed an average of 5 of 46 tasks, compared with zero at the base stage. The source says these results are averages across three evaluation runs.
The training sequence was Base, RSR-1, RPR-1, RSR-2 and RPR-2. In Recursive Self-Rewrite, or RSR, the model’s verifier-approved attempts were converted into step-by-step plans, screened for leaked answers and run again; successful reruns became training examples. Recursive Policy-Refine, or RPR, used reinforcement learning to make higher-scoring behavior more likely. The paper says a critic network was carried between refinement stages and recalibrated before each one.
Infographic 01 · The results, stage by stage
Real gains on the benchmarks, and a small long-horizon result
A 27 billion parameter open model, trained in five stages that alternate learning from its own verified successes with reinforcement learning.
From solving nothing to about one task in five, over five training stages. Each score is the mean of three evaluation runs (Mean@3).
Base → final, per benchmark
Bar = final score; base in the label
Source: Yang et al., Terminal-RSI, arXiv 2610.22192 v3 (30 Oct 2026), Table 2 and Figure 5. All values as printed in the paper.
Powered by Thorsten Meyer AI
Why Verified Completion Matters
The reported gains matter because command-line agents can appear to make progress while failing to complete the requested work. Terminal-RSI’s central contribution, as described in the supplied account of the paper, is a reward design that separates success from failure. Under the earlier reward approach, failed attempts still received 50.09% of all reward in the measured set. The replacement kept every failure’s reward below every success’s.
That distinction addresses a practical weakness in agent training: rewarding partial progress can encourage activity that does not produce a working result. The paper defines success by whether a complete protected verifier passes, rather than whether the agent says it is done or passes only some checks. The source also reports 748 failed runs that declared completion within ten turns, illustrating why an independent check can matter to teams evaluating automated software work.
But benchmark improvements are not evidence that an agent can independently set its own objectives, training procedure or evaluation standards. Human researchers shaped each stage, and the gains vary by test. SWR100, described as a test of tasks kept out of training, rose from 3% to 6% and did not increase after the third stage. Terminal-Bench 2.1 likewise showed no gain in the final stage. Those limits matter when interpreting broad claims about self-improving AI.
Products Worth Considering
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p Camera – Blush
Apple iPad 11-inch with A16 Chip, 512GB, Wi-Fi 6, and Apple Pencil (USB-C) — Blue
As an affiliate, we earn on qualifying purchases.
How the Five Stages Worked
The source describes a terminal agent as a model that interacts with a computer through a command line: it reads files, enters commands and inspects results while attempting tasks such as repairing a build or fixing a data pipeline. These tasks require more than generating a plausible response; the machine must reach a state that meets the task’s requirements.
Terminal-RSI’s RSR stages trained on the model’s own attempts, but only after a verifier accepted them. The authors’ process then turned successful attempts into plans and checked those plans by executing them again. The two RPR stages applied reinforcement learning after those example-based stages. The source reports a control comparison: reinforcement learning started directly from the base model remained at 0.00% on Terminal-Bench 4.0 at the checkpoints recorded, while reinforcement learning after the first self-rewrite stage reached 16.67%.
The reported results are not uniformly strong across evaluations. Terminal-Bench 2.1 rose from 51.70% at base to 86.52% by RSR-2, then stayed at that score in the final stage. Terminal-Bench Hard reached 79%, while SWR100 ended at 6%. The source’s figures are specific to the benchmarks and evaluation setup reported; they do not establish performance across all real-world terminal tasks.
“It is a training pipeline, not an AI improving itself unsupervised.”
— Thorsten Meyer, ThorstenMeyerAI.com
Limits of the Reported Results
The supplied source identifies the work as a preprint but does not provide a publication date, the paper’s full author list or enough information to independently assess the study’s review status. It also does not include the full benchmark protocols, statistical uncertainty or details needed to judge how well these results generalize beyond the reported tests.
The available account describes the reward comparison and benchmark outcomes, but its supplied text ends before completing the explanation of the 4,352 recorded attempts. Further details about that measurement, the exact reward formula and the extent of human intervention in individual training runs are not available here. The reported results therefore support claims about a managed training pipeline and specific evaluations, not a general conclusion that AI systems can improve themselves without people.
Long-horizon performance remains a particular limitation in the reported numbers: the final model completed five of 46 tasks on average. It is also unclear from the supplied material how the model would perform on broader task distributions, under different verifiers or outside the researchers’ training and evaluation setup.
What Further Testing Should Show
The immediate next step for readers assessing the claim is to examine the full preprint, including its methods, appendices and limitations, and compare the reported scores with the benchmark protocols. The supplied source does not identify a scheduled follow-up study, peer-review decision or additional release, so none can be stated as confirmed.
Further evaluations would help clarify whether the approach transfers to more unseen tasks and whether improvements persist on long-horizon work, where the final model completed only 5 of 46 tasks on average. Additional reporting on the verifier, reward design, training resources and human decisions would also help distinguish model-generated training data from human-directed choices in the pipeline.
Key Questions
Did Terminal-RSI show an AI improving itself without human input?
No. The reported process used examples from the model’s own verified successes, but researchers chose and ran the stages and set the rewards, learning rates and checkpoint selection. The supplied account describes it as a human-managed training pipeline, not unsupervised self-improvement.
What benchmark result did the model report?
On Terminal-Bench 4.0, the reported score rose from 0.00% to 19.84%. The paper’s equal-weight average across four Terminal-Bench suites rose from 20.68% to 50.58%. The latter excludes the long-horizon benchmark and SWR100.
What does Mean@3 mean here?
According to the source, Mean@3 is the arithmetic mean of three independent evaluation runs. It is not the best score selected from three attempts.
How well did the model handle long tasks?
The final model completed an average of 5 of 46 long-horizon tasks, up from zero at the base stage. The result indicates progress in that test but also shows that most tasks were not completed.
Why did the researchers use a verifier?
A verifier checks the machine’s final state against protected tests instead of relying on the agent’s claim that it finished. The source says this matters because recorded failed attempts sometimes declared the task complete.
Source: ThorstenMeyerAI.com
