A 27-billion-parameter coding agent trained in a loop that reuses its own verified successes went from 51.70% to 86.52% on the Terminal-Bench 2.1 test, according to a paper posted to arXiv on 15 October 2026 by researchers listed under Tencent Hy Frontier and four universities.
The paper, Terminal-RSI, describes two cycles in which each improved model helps train the next. On the harder Terminal-Bench 4.0 the same agent went from 0.00% to 19.84%, the authors write.
Business Pill · SYNTHETIC DATA
A short explainer on synthetic data, when a model writes its own training examples. It teaches the general idea only and says nothing about the paper or any organisation in this story.
The key insight: As we read it, the paper’s claim is about the loop more than the model: a 27B base model that scored 51.70% on Terminal-Bench 2.1 reached 86.52% after two cycles of learning from its own verified runs, while a direct PPO run from the same base scored 48.31%, according to the paper.
The paper is “Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine,” by Junyao Yang, Zongxia Li, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Ruiwen Zhou, Haitao Mi, Shuicheng Yan and Leowei Liang. Seven of the nine list Tencent Hy Frontier; the others are listed under the National University of Singapore, and several Tencent authors also list NUS, the University of Maryland, the University of Georgia or Indiana University. Shuicheng Yan is the corresponding author.
The starting model is Qwen3.8-27B. The method alternates two steps: a rewrite step built on verified successful trajectories, and a policy-refinement step that inherits the previous stage’s critic, a valuemodel handed from one stage to the next, the paper says.
It runs that pair twice, for five stages in all: the base model, a first rewrite, a first refinement, a second rewrite and a second refinement. The final checkpoint is called Terminal-RSI-27B.
Training draws on a 3,000-task pool, 2,400 semantic-verifier tasks and 600 supplementary terminal tasks, according to the paper. Scores are the mean of three evaluation runs.
Yucheng Shi, one of the authors, described it on X: “We used an agent to help train a stronger successor. Then we used that successor to help train the next.”
Success of the base Qwen3.8-27B model and the final Terminal-RSI-27B checkpoint on five benchmarks, mean of three evaluation runs, from the paper’s results table (arXiv 2610.22192, 15 October 2026).
What Was Posted on X
We used an agent to help train a stronger successor. Then we used that successor to help train the next.
Our new paper, Terminal-RSI, studies this recursive self-improvement loop: learn from verified successes, improve through practice, and repeat.
Across two cycles, a 27B model’s Terminal-Bench 2.1 success rises from 51.7% to 86.5%, and Terminal-Bench 4.0 success rises from 0.0% to 19.8%, a new SOTA.
We’re interested in how far this kind of hill climbing can go, and what makes it stall.
Yucheng Shi, one of the paper’s authors, introducing Terminal-RSI on X: an agent helps train a stronger successor, which then helps train the next, with the post’s rounded Terminal-Bench figures and the arXiv link.
Banger paper from Tencent on recursive self-improvement for terminal agents.
(bookmark it)
This is a great read if you are tracking research on training terminal or coding agents with RL
Starting from Qwen3.8-27B, they alternate two phases over five stages. One phase fine-tunes on rewritten versions of trajectories that passed the verifier. The other runs PPO with a critic inherited from the previous stage.
The four-suite Terminal-Bench average goes from 20.68% to 50.58%, with 86.52% on Terminal-Bench 2.1 and 79.00% on Terminal-Bench Hard.
The reward audit is just as useful. Under a dense progress reward, failed runs collected 50.09% of the reward, and 48 failures passed every executed check but still missed the final test. Separating strict success from partial credit cut that share to 28.72%.
Elvis Saravia (DAIR.AI) recommending the paper on 17 Oct and summarising it in his own words: Qwen3.8-27B trained over five stages that alternate fine-tuning on verifier-passed, rewritten trajectories with PPO, and the paper’s Terminal-Bench figures (four-suite average 20.68% to 50.58%). He also highlights the paper’s reward audit, which he reports as failed runs collecting 50.09% of a dense progress reward, cut to 28.72% once strict success was separated from partial credit; those audit figures are his summary of the paper.
The Numbers, Stage by Stage
On Terminal-Bench 2.1 the scores went 51.70%, 70.10%, 77.53%, 86.52% and 86.52% across the five stages, so the last stage held the score rather than raising it, according to the paper’s results table.
On Terminal-Bench 4.0 the scores went 0.00%, 6.10%, 16.67%, 17.46% and 19.84%. On Terminal-Bench 3.0 the agent went from 0.00% to 16.95%, and on Terminal-Bench Hard from 31.00% to 79.00%, the authors write.
Averaged equally across those four suites, success rose from 20.68% to 50.58%, the paper says.
On a long-horizon benchmark, LHTB, strict completions went from 0 of 46 tasks to 5 of 46. On SWR100, a set with no task overlap with the training pool, success doubled from 3.00% to 6.00% and did not rise after the first refinement stage, the authors write.
A separate direct PPO run from the base model, which the paper reports as a historical aggregate, scored 48.31% on Terminal-Bench 2.1, below the base model’s 51.70%. The paper says the first refinement stage outperformed direct PPO by up to +29.2% on Terminal-Bench 2.1 and +16.7% on Terminal-Bench 4.0 at step 20.
How It Compares With Frontier Models
The paper compares its final model with figures for frontier models in a reference table. On Terminal-Bench 2.1 its 86.52% is above the listed 85.40% for GPT-6 Sol and 85.00% for Kimi K3, according to the paper.
On Terminal-Bench 4.0 its 19.84% is above the listed 19.70% for Gemini 3.8 Flash and 14.60% for DeepSeek V4 Pro. The authors write that GPT-6 Sol is higher on Terminal-Bench 4.0 and Gemini 3.8 Flash higher on Terminal-Bench 2.1.
The paper adds: “External scores are reference values; their primary-source and protocol reconciliation remains incomplete.” Shi’s post calls the Terminal-Bench 4.0 result “a new SOTA”; the paper’s own comparison lists GPT-6 Sol higher on that benchmark.
Shi wrote that the team is “interested in how far this kind of hill climbing can go, and what makes it stall.”
The Structural Read
As we read it, the verifier is the scarce input: the training pool is 3,000 tasks, 2,400 of them semantic-verifier tasks, and each rewrite stage builds on verified successful trajectories, according to the paper.
As we read it, the gains shrink as the loop runs: on Terminal-Bench 2.1 the last stage held the score at 86.52%, and by our arithmetic the equal-weight mean rose 13.57 points in the first stage but 1.83 in the last.
As we read it, the frontier comparison is narrower than the post: by the paper’s own table, GPT-6 Sol is higher on Terminal-Bench 4.0 and Gemini 3.8 Flash on Terminal-Bench 2.1, and the authors call the external scores reference values.
Yang et al., arXiv 2610.22192
“each retained stage consolidates discovered behaviors and progressively advances the agent from partial local repairs to strict terminal verification”
Three Implications
TEAMS TRAINING AGENTS The paper says starting from verified rewrites gave refinement an anchor: the first refinement stage outperformed direct PPO by up to +29.2% on Terminal-Bench 2.1 at step 20.
BENCHMARK READERS The paper says its evaluation repeats measure inference variability rather than independent training replication, so the stage-by-stage gains come from one training sequence.
MODEL BUYERS The final 27B checkpoint sits close to the frontier figures the paper lists on two terminal benchmarks, which the authors say are reference values not yet reconciled with primary sources.
The framework’s core claim is that “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.” One of its four moats is “The Verifier Moat — built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”
As we read it, Terminal-RSI is a small instance of that verifier moat: the paper’s gains come from tasks with protected verifiers that decide which runs count as successes. The framework predates the paper and does not discuss it.
Equal-weight mean success across Terminal-Bench 4.0, 3.0, 2.1 and Hard at each retained stage: 20.68%, 34.25%, 44.50%, 48.75% and 50.58%, according to the paper (arXiv 2610.22192, 15 October 2026).
What Is Not Established
The paper is a preprint, and its authors list the limits. “Three evaluations do not replace independent training seeds or establish statistical superiority,” they write, and “Five stages of one 27B backbone establish neither scaling laws nor universal monotonicity.”
The authors also write that broader repository and pretraining overlap “remain incompletely audited,” and that the stage-specific learning rates, rewards and initialisation change together, so the paper does not isolate which part caused which gain.
Business Engineer Framework
The Four Intelligence Moats
A Business Engineer framework on where intelligence accumulates in each AI paradigm, including the verifier moat built on reward signals.
A paper from researchers listed under Tencent Hy Frontier and four universities reports that a 27B agent trained in two cycles of rewriting its own verified successes and further practice went from 51.70% to 86.52% on Terminal-Bench 2.1 and from 0.00% to 19.84% on Terminal-Bench 4.0.
The authors compare those scores with frontier models but write that the external figures are reference values whose reconciliation is incomplete.
A note on sourcing. We read the arXiv preprint 2610.22192 in full, version 1, dated 15 October 2026; author names and affiliations are as listed on the paper. All figures are the paper’s own, from its results table and text; the frontier-model scores are the paper’s reference values, which its authors say are not yet fully reconciled with primary sources. We found the paper through Yucheng Shi’s post on X.
The paper links a code repository and a model-weights page; we did not test either. The TreeFlex and agent-harness references link to our own earlier reporting.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
