AI develops 27B model autonomously, Recursive Self-Improvement achieves industrial results
机器之心 · wechat · 2026-09-01
A joint team from SJTU enabled an Agent to lead the full-cycle development of a 27B industrial coding model, iCoder. From data evolution and SFT to reinforcement learning, the Agent handled experiment selection and strategy modification. The model ranks in the first tier on industrial coding benchmarks like RTL and GPU Kernels.
AI Self-Evolution Process
- Data: Agent unifies tasks into triples and filters them through strict verification.
- SFT: Targets gaps in the base model using verified teacher trajectories.
- OPSD: Writes back the ability to “correct with feedback” into parameters, bridging SFT and RL.
- RLVR: Uses hierarchical verifiable tasks and compiler/simulator results as rewards, training near the capability boundary.
Challenges & Solutions
The Agent successfully addressed issues like repetitive generation, reward hacking (e.g., calling reference frameworks), and evaluation biases caused by infrastructure failures. This demonstrates AI's ability to take over critical R&D tasks and encode experience into the next generation, representing a new paradigm for AI-led development.
More from coding & agent
- Why Document AI Integration Needs 6 SDKs: IDPForge Unifies the Pipeline — infrrd-ai · 2026-09-01
- Dev ships MCP server for sanctions screening and CVE lookups, pay-per-query via x402 — Glittering-Buyer1830 · 2026-09-01
- Agent-Driven Reverse Engineering of Peripherals Using Claude Opus 5 — bibryam · 2026-09-01
- New tool md² handles agent switching when usage limits are reached — JBO_76 · 2026-09-01
- TouchBar Revived by AI Agents: Real-Time Coding Status on Old Hardware — APPSO · 2026-09-01
- Vector DB Pitfalls: 8 Popular Databases Tested in Production — PrajwalTomar_ · 2026-09-01