Wire flash
Claude Opus 5 nearly triples Qwen model's SWE-bench Verified score as automated AI researcher
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
A post on X reports that Claude Opus 5 nearly tripled a Qwen model's SWE-bench Verified score while operating as an automated AI researcher. The measurement was made possible by Evolvent_AI's newly released open-source RSIGym, which allows an AI agent to retrain a model and rewrite its harness inside a single budgeted environment. The post notes that frontier AI agents can substantially improve another AI model when training, serving, and testing are provided as ready-made services. Agents work from CPU-only containers and call remote services for LoRA fine-tuning, model serving, benchmarking, and sandboxes, all charged against a per-run budget. In the main test, six frontier agents started from Qwen3.5-35B-A3B-Base and a minimal harness, with $500 of services per benchmark. The post also advises that when improving an agent, one should read its failure logs and fix the harness first, as heavier training scored lower in 8 of 10 small tests.
Source report
Claude Opus 5 nearly tripled a Qwen model's SWE-bench Verified score while operating as an automated AI researcher.
The measurement was made possible by @Evolvent_AI's newly released open-source RSIGym, which allows an AI agent to retrain a model and rewrite its harness within a single budgeted environment.
Key Findings
- Frontier AI agents can substantially improve another AI model when training, serving, and testing are provided as ready-made services.
- Agents operate from CPU-only containers and call remote services for:
- LoRA fine-tuning
- Model serving
- Benchmarking
- Sandboxes
All services are charged against a per-run budget.
Main Test Results
In the primary test, six frontier agents started from Qwen3.5-35B-A3B-Base and a minimal harness, with $500 of services per benchmark.
Recommendation for Agent Improvement
If you are improving an agent, read its failure logs and fix the harness first: heavier training scored lower in 8 out of 10 small tests.
Source
rohanpaul_aiNeutral / independent