Evolvent AI Open-sourcing RSIGym and releasing AI recursive self-improvement evaluation experiment results

AT A GLANCE

Evolvent AI RSIGym has been open-sourced. Experiments show that some single-round optimizations are effective, but there is not yet evidence of multi-round recursive self-improvement.

Article

RSIGym open-sourced: enabling AI research into how to improve AI, targeting recursive self-improvement

Dongcha Beating AI breaking news:AI Evolvent, which develops AI has open-sourced RSIGym, with the goal of studying AI recursive self-improvement (RSI). Put simply, this enables AI to identify system weaknesses, propose modifications, test their effects, and then attempt to use the improved system for the next round of research.

RSIGym packages training, inference, evaluation, and sandboxing as ready-to-use services. Research Agents can generate training data, adjust fine-tuning parameters, train models, and then revise their approach based on evaluation results. It can also modify Harness, namely the program that controls how models call tools and execute tasks. Researchers can have an Agent modify only training data, only Harness, or optimize both simultaneously. The platform centrally manages experiment budgets, permissions, and execution logs.

The team also introduced RSI-Index to test AI Agent's ability to improve another AI system. 6 research Agents each started from the same Qwen3.5 base model and a simple Harness and optimized performance on 5 types of tasks, including programming and mathematics. Opus 5 ranked first overall. Its modified system improved from SWE on the 100-problem subset of -bench Verified 17.67%to 50.33%; on the mathematics test AIME it improved from 31.67%to 97.78%。

Another experiment modified only Harness, without changing model weights. Opus 5 used execution logs from DeepSeek Harness to fix issues such as stopping work after output truncation and prematurely declaring tasks complete. With Qwen3.6 model weights completely unchanged, the Terminal-Bench 2.0 score improved from 30.34%to 40.82%。

The team also attempted to have Qwen3.8-27B improve itself, serving both as the research Agent and the model being improved. On 6 test problems, it found a fine-tuning approach that appeared effective, but across the full 37 problems, its score instead fell slightly from 56.66%to 56.26%. In addition, each experiment involved only a single research run, and the Agent could access evaluation problems during development. There is currently no evidence that an improved AI can continue enhancing its own R&D capabilities to achieve multi-round recursive self-improvement.

Original link https://m.theblockbeats.info/flash/371144

Key points

Evolvent AI has open-sourced RSIGym, with the goal of studying AI's recursive self-improvement.
RSIGym provides training, inference, evaluation, and sandboxing services, and centrally manages experiment budgets, permissions, and execution logs.
Research Agents can modify training data, fine-tuning parameters, and Harness, and can also optimize training data and Harness。
The team introduced RSI-Index to test AI Agent's ability to improve another AI system.
On the SWE-problem subset of -bench Verified 100, the system modified by Opus 5 improved from 17.67%to 50.33%；AIME The score improved from 31.67%to 97.78%。
When modifying only Harness and with Qwen3.6 model weights unchanged, the Terminal-Bench 2.0 score improved from 30.34%to 40.82%。
Qwen3.8-27B The self-improvement approach appeared effective on 6 test problems, but across the full 37 problems, the score fell from 56.66%to 56.26%。
Each experiment involved only a single research run, and the Agent could access evaluation problems during development.
There is currently no evidence that an improved AI can continue enhancing its own R&D capabilities to achieve multi-round recursive self-improvement.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

RSIGym is like a facility for AI researchers to use:AI They can propose improvements, train or adjust the system, and then use evaluation results to decide what to do next. Here, the Harness is the program that controls how the model calls tools and performs tasks. Existing experiments show that the AI can improve scores on some tasks by modifying the training scheme or Harness improve a single evaluation score, but this does not mean it can already continuously and iteratively become better at research and development.

Why it matters to readers

If such methods are reliable,AI could contribute to training-data creation, parameter adjustment, tool-calling workflow optimization, and performance evaluation, reducing the human input required for some AI R&D work. Current evidence supports improvements only in certain single-round experiments and cannot demonstrate that recursive self-improvement has been achieved.

Beginners can understand RSIGym as a platform that lets an AI try to modify another AI system. The focus is not only on whether scores increase, but also on whether testing is independent, results can be reproduced, and improvements can persist across multiple rounds.

Risks and unknowns

The materials do not provide official project announcements, a code repository, papers, or complete evaluation configurations; the open-source status and experimental details still require verification against the original materials.
Each experiment group was run only once, so the stability and reproducibility of the results are unknown.

Evolvent AI Open-sourcing RSIGym and releasing AI recursive self-improvement evaluation experiment results

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center