Tencent Lets a 27B Model Rewrite Itself, First Curing It of Pretending to Succeed

Tencent has released a 27B model called Terminal-RSI, which can rewrite, retrain, and refine itself, with no human watching in real time during the process. The original post is from X user Superman, with the image below.

The really noteworthy part is not the big term 'recursive self-improvement,' but how it cures an old bad habit of agents: declaring success halfway through the job.

Agents doing terminal tasks often behave like this: they make some local fixes, pass the semantic check, and then when final environment verification is run, they fail. The original post calls this 'lazy.' More precisely, it has learned to get partial credit but has not learned to cross the finish line.

Terminal-RSI's approach is a two-stage loop: RSR, recursive self-rewrite, in which the model fine-tunes on its own successful trajectories; RPR, recursive policy refinement, then runs PPO, with the critic inherited directly from the previous stage. The training sequence is like hopscotch: Base to RSR-1, then RPR-1, then RSR-2, RPR-2.

The numbers are eye-catching. Based on Qwen3.8-27B, the average success rate across four Terminal-Bench suites is pulled from 20.68% to 50.58%. On the hardest Terminal-Bench 4.0, the base model is 0.00%, and after Terminal-RSI it reaches 19.84%; the original post says this is already comparable to reference models such as Gemini 3.8 Flash.

The secret is in the reward. It does not use dense rewards to coax the model into 'trying'; instead it switches to success-prioritized reward, mathematically separating strict success from partial credit. If the code does not pass the final terminal gate, it cannot get the large reward. This design is more substantial than the slogan 'the model will self-evolve.' The training signal has moved from looking right to actually passing.

But don't overstate it yet. Some people under the original post question whether the evidence is sufficient. X user The SI Maximalist says automated retraining is exciting, but the recursive claim needs proof that the model's improvement in this round can help it generate the next round of improvement; simply running an unattended training pipeline does not automatically establish that. X user 大伟｜AI × Web3 also points out that being able to self-rewrite and retrain is not the same as cracking recursive self-improvement; the key is whether capability gains persist across rounds, whether costs spiral out of control, and whether the evaluation has been gamed.

19.84% on the hardest suite is not high either. X user Robin says success-prioritized reward still has a long way to go. X user Anatol Wegner is more direct, saying this looks more like teaching the model to benchhack.

In the same week, there was another piece of news: the Bloomington City Council voted unanimously to ban recursive self-improving entities that can greatly surpass humans in all cognitive tasks from entering the city limits, citing existential risk and loss of control. No one made clear how the enforcement target would be defined.

The most valuable part of Terminal-RSI is that it pushes the agent's training objective one step toward final verification. As for whether it means 'China has cracked recursive self-improvement,' first look at cross-round gains, cost, and evaluation contamination. Before a model can rewrite itself, it should first learn not to pretend to succeed.

Original post link: https://x.com/thesupermannx/status/2119341077975249222