🔍 Read the full analysis: Did An AI Agent Learn To Improve Itself In Terminal-RSI? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A 26-page preprint reports that a 27-billion-parameter terminal agent improved on several benchmarks after five training stages that researchers designed and supervised. The results show progress from learning from verified successes and reinforcement learning, not an AI independently rewriting and retraining itself; long-horizon performance remained limited.
A research team behind the Terminal-RSI preprint reports that a 27-billion-parameter terminal agent raised its score on Terminal-Bench 4.0 from 0.00% to 19.84% through five training stages. The result is evidence of benchmark improvement under a researcher-designed process, not proof that an AI independently rewrote, retrained and improved itself without human direction.
The 26-page preprint comes from Tencent’s Hy Frontier lab and four universities. It describes a pipeline using an open model identified as Qwen3.8-27B, with a final checkpoint named Terminal-RSI-27B. The researchers trained on a pool of 3,000 tasks and evaluated the model across several terminal-agent benchmarks. The source describing the paper says each reported score is a mean across three evaluation runs.
The pipeline alternated two methods. In Recursive Self-Rewrite, or RSR, the model’s attempts were checked by a verifier; accepted attempts were turned into step-by-step plans, screened for leaked answers and run again. Successful examples were then used for supervised fine-tuning. In Recursive Policy-Refine, or RPR, the agent received reward signals and was trained to favor higher-scoring behavior. The sequence was Base, RSR-1, RPR-1, RSR-2 and RPR-2.
On Terminal-Bench 4.0, scores rose through those stages to 19.84%. The paper also reports an equal-weight average across four Terminal-Bench suites of 20.68% for the base model and 50.58% for the final model. Those figures cover only those four suites: the average excludes the long-horizon test and SWR100, and gives each included suite the same weight regardless of its size.
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 paper’s practical contribution is as much about how an agent is rewarded as about the resulting scores. Terminal tasks can be partly completed without being finished. A reward that gives credit for individual checks may encourage progress, but it can also reward activity that falls short of the requested outcome. The researchers report that, under an earlier reward design, failed attempts received 50.09% of all reward.
The revised reward is designed so that every failure receives less reward than every success. That distinction matters for systems used to edit code, repair builds or operate data pipelines: a convincing report from an agent is not the same as a working result. In the paper’s setup, hidden verifiers check the machine’s final state rather than relying on the agent to say it has succeeded.
The results also show the limits of the reported gains. On the long-horizon benchmark, the final model completed an average of 5 of 46 tasks, compared with zero for the base model. SWR100, the scientific-workflow test described as held out from training, rose from 3% to 6% and then did not increase in later stages. These findings support further work on training and verification, but do not establish broad autonomous improvement or reliable performance on extended tasks.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
Samsung 49-Inch Odyssey OLED G9 (G91SD) Dual QHD Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Five Stages Worked
Terminal agents work in a command-line environment: they inspect files, run commands, read outputs and try to carry out tasks such as fixing a broken build. Their performance can be difficult to judge if evaluation accepts the agent’s own claim that it is done. Terminal-RSI instead defines success through a complete protected verifier, according to the paper description, rather than an agent declaration or a subset of passed checks.
The researchers’ reported control provides a comparison for the staged approach. Plain reinforcement learning starting directly from the base model scored 0.00% on Terminal-Bench 4.0 at every checkpoint the authors recorded. After an RSR stage, the next reinforcement-learning stage reached 16.67%. That comparison suggests the verified examples produced by the initial rewrite stage helped the later training process, although it does not establish that the same pattern will hold across other models, tasks or training setups.
The source account also describes a failure mode seen in the paper’s recorded attempts: 748 failed runs declared the task complete within ten turns. The count helps explain why external checks matter, but it does not by itself say how often this behavior occurs in other systems or settings. The paper’s reported progress should be read against the benchmark definitions and training procedure, not as a general measure of terminal-agent reliability.
“It is a training pipeline, not an AI improving itself unsupervised.”
— Thorsten Meyer, describing the paper on ThorstenMeyerAI.com
Limits of the Reported Results
The reported results come from a preprint, and the supplied source material does not establish that they have been peer reviewed or independently reproduced. The available account gives benchmark results and selected details of the method, but does not provide enough information here to judge the full statistical uncertainty, the sensitivity of results to different training choices, or how performance would transfer to real-world deployments.
It is also unclear from the supplied material how the system would perform on tasks outside the named benchmarks, under different verification rules, or over longer sequences of work. The reported long-horizon score remains low in absolute terms, and the held-out SWR100 score stopped rising after the third stage. The paper therefore does not show that the agent can improve its own training process or choose its own goals without people setting the task, reward, stages and checkpoint-selection rules.
What Further Testing Must Show
The next evidence to watch for is independent replication of the benchmark results and fuller reporting of the training and evaluation setup. Tests across additional tasks, models and verifier designs would help establish whether the gains extend beyond this particular pipeline and benchmark mix. Longer-horizon evaluations would also show whether improvements in short terminal tasks translate into sustained task completion.
For now, Terminal-RSI is best read as a report on a structured training method: the researchers selected the stages, rewards, learning rates and checkpoints, while the model supplied examples that were verified and used in later training. Whether that process generalizes, and whether future systems can meaningfully automate more of the research decisions themselves, remains unproven by the results described here.
Key Questions
Did Terminal-RSI prove that an AI can improve itself without human involvement?
No. The reported process used human-designed and supervised training stages. Researchers set the rewards, learning rates and checkpoint decisions; the paper does not show an agent independently directing its own retraining.
What score did the final model reach on Terminal-Bench 4.0?
The paper reports 19.84% for the final model, up from 0.00% for the base model. The scores are described as means across three evaluation runs.
How did Terminal-RSI train the agent?
It alternated Recursive Self-Rewrite, which used verified successful attempts as training examples, with Recursive Policy-Refine, a reinforcement-learning stage. The sequence contained five checkpoints from the base model through two rounds of each method.
Did the agent become reliable on long tasks?
The reported long-horizon result was limited: the final model completed an average of 5 out of 46 tasks, up from zero for the base model. That result does not establish strong performance on extended tasks.
Why did the researchers use a verifier?
A verifier checks whether the task’s required final state was achieved, rather than trusting the agent’s own claim of success. The paper’s approach is intended to distinguish completed work from apparent progress.
Source: ThorstenMeyerAI.com
