Did Terminal-RSI Let An AI Agent Improve Itself With No Human Help?

🔍 Read the full analysis: Did Terminal-RSI Let An AI Agent Improve Itself With No Human Help? on ThorstenMeyerAI.com

A 26-page preprint called Terminal-RSI reports that a 27-billion-parameter terminal agent improved on several task benchmarks through five training stages, but it does not show the AI improving itself without human help. Researchers from Tencent’s Hy Frontier lab and four universities designed and ran the stages, set the rewards and training settings, and chose which model checkpoints to keep.

The final model, named Terminal-RSI-27B, scored 19.84% on Terminal-Bench 4.0, up from 0% for its base model. The paper also reports that its equal-weight average across four Terminal-Bench suites rose from 20.68% to 50.58%. Scores are Mean@3, an arithmetic average of three evaluation runs, rather than the best result from three attempts.

The training sequence alternated between learning from verified examples and reinforcement learning: Base, RSR-1, RPR-1, RSR-2 and RPR-2. In Recursive Self-Rewrite, the researchers retained attempts that passed a verifier, converted them into step-by-step plans, screened the plans for leaked answers and ran them again. Attempts that passed the second check became training examples. Recursive Policy-Refine then used reward scores to adjust the model’s behavior.

The gains were uneven. On a long-horizon benchmark, the final model completed an average of 5 of 46 tasks, compared with zero for the base model. On SWR100, a scientific-workflow test the source describes as kept out of training, the score rose from 3% to 6% but did not increase after the third stage. Terminal-Bench 2.1 also remained at 86.52% between the final two stages.

Why Verifying Completion Matters

The paper’s main practical point is about how agents are rewarded, not autonomous self-improvement. In command-line tasks, an agent can make partial progress and still fail the actual job, or declare success before completing it. The researchers say that in one recorded set, 748 failed runs declared completion within 10 turns. Their benchmark verifier checks the final machine state rather than accepting the agent’s own report.

Reward design affected what the system learned to prioritize. Under an earlier partial-credit reward, failed attempts still received 50.09% of all reward, according to the source’s account of the paper. The revised reward placed every failure below every success. That approach is intended to make verified task completion matter more than accumulating credit for partial work. It also highlights a continuing limitation: passing a test suite does not necessarily mean the agent built what a task requested.

For readers following claims about AI systems improving themselves, the distinction matters. The reported results show that a model can be trained iteratively using examples derived from its own verified attempts. They do not establish that it independently chose its goals, redesigned its training process or operated without human oversight. The benchmark gains are a reported research result, while broader claims of an AI “cracking” recursive self-improvement go beyond what the described process demonstrates.

How the Five Training Stages Worked

A terminal agent operates through a command line: it reads files, runs commands, examines outputs and tries to complete tasks such as repairing software or data workflows. Terminal-RSI starts from Qwen3.8-27B, an open model, and uses a pool of 3,000 tasks. The resulting pipeline combines supervised fine-tuning, which teaches the model from examples, with reinforcement learning, which adjusts behavior based on reward.

The researchers’ control result offers evidence that the order of training stages mattered in their experiments. The source reports that reinforcement learning applied directly to the base model scored 0% on Terminal-Bench 4.0 at every recorded checkpoint. When the self-rewrite stage came first, the subsequent reinforcement-learning stage scored 16.67% on that benchmark. This comparison supports the authors’ case for the sequence they tested, but it does not by itself show that the method will produce the same gains on other systems or tasks.

The paper defines success as passing the complete protected verifier, not an agent’s declaration or a subset of checks. That standard underpins the training examples and reported benchmark outcomes. Still, the reported measurements cover particular suites and a limited set of evaluation runs; the headline average across four Terminal-Bench suites excludes both the long-horizon benchmark and SWR100.

“It is a training pipeline, not an AI improving itself unsupervised.”

— Thorsten Meyer, ThorstenMeyerAI.com

Limits of the Reported Results

The results come from a preprint, and the supplied material does not establish that the work has been peer reviewed. The benchmark figures are reported by the paper and summarized by ThorstenMeyerAI.com; this article has no independent replication or evaluation data. The available source also does not provide enough detail to assess variation across the three runs beyond identifying the scores as Mean@3.

It remains unclear how well the approach generalizes beyond the listed benchmarks, how much performance depends on the selected tasks and reward definitions, and whether later stages would continue to improve results. The long-horizon outcome remains limited to an average of five completed tasks out of 46. The source also describes a final model that passed its tests without building what the task asked for, underscoring that verifier success can still be an incomplete measure of whether the intended work was done.

Most importantly, the process was not human-free: people set up and managed the training pipeline. The material does not show the model independently selecting training goals or operating an unsupervised loop that rewrites and retrains itself. Nor does it establish that such a system is impossible; it simply does not demonstrate one.

Further Testing and Replication

The next useful evidence would be independent replication of the reported scores, with fuller reporting on run-to-run variation and the choices made in each training stage. Tests on additional task suites could show whether the gains extend beyond the benchmarks used in the paper. Researchers would also need to examine whether agents satisfy the task’s actual intent, not just pass the checks used to score it.

The supplied material does not identify a peer-review decision, a release schedule or a planned follow-up study. Until those details emerge, the reported benchmark gains should be read as findings from this preprint, not as proof of autonomous self-improvement. What happens next will depend on publication and replication details that were not provided in the source.

Key Questions

Did Terminal-RSI improve itself without human help?

No. The report describes a human-designed and managed training pipeline. Researchers selected the stages, rewards, learning settings and checkpoints.

What benchmark gain did the paper report?

The final model scored 19.84% on Terminal-Bench 4.0, compared with 0% for the base model. The paper also reports an equal-weight average across four Terminal-Bench suites rising from 20.68% to 50.58%.

How did the model learn from its own attempts?

Researchers kept attempts that passed a protected verifier, turned them into step-by-step plans, screened and reran those plans, then used successful examples for training. Reinforcement-learning stages followed.

Did the final model solve long tasks reliably?

The reported average was 5 completed tasks out of 46, up from zero for the base model. That result indicates improvement in the test, but also shows that most tasks were not completed.

Has the Terminal-RSI work been independently confirmed?

The source describes Terminal-RSI as a preprint and provides no independent replication or peer-review status. The benchmark figures should be treated as results reported by the authors until further confirmation is available.

Source: ThorstenMeyerAI.com