iCoder-27B: An Agent Trained a Model That Beats GPT-5.5 on Chip Design
How little human is enough for an agent to build a frontier model? iCoder-27B is a serious attempt at a number. The humans did not run experiments. They wrote the rules: objectives, stage scaffolds, permission boundaries and operating procedures, packaged as reusable research skills. The agent did everything else. It evolved the training data, ran SFT, on-policy self-distillation and RL with verifiable rewards, read the results and changed the plan.
The output is a 27B model for two very unglamorous and very valuable jobs: RTL hardware design and GPU kernel optimization. Across seven benchmarks it leads RTLLM, ahead of GPT-5.5 and Claude Opus 4.8. It is second on CVDP and KernelBench L2, 16 points ahead of GPT-5.5 on the latter, and ties Opus 4.8 for the top TritonBench score. The authors also report competitive iterative optimization with far fewer tokens.
The interesting design choice is where the human sits. Not in the loop, turning knobs every hour, but upstream, in what the authors call a high-density, low-frequency interface. That is the skills file idea applied to research itself. The expert's judgment gets compiled once into procedures and boundaries, and then the agent runs many cheap iterations inside them. It is the same pattern as last month's AI-for-AI results, just aimed at a release-grade model rather than a toy.
Two caveats. The comparison models are a generation old now, GPT-5.5 and Opus 4.8, not this month's releases. And the paper, like most of this genre, reports what the agent achieved and says less about what it tried and threw away, which is exactly where you would look for the reward hacking another paper this week measured at 30% on open-ended research tasks. Still, a domain model beating general frontier models on hardware, built mostly by an agent, is a clear data point on where AI-for-AI is heading. It appeared in arXiv's September 25 listings.
Link: arxiv.org/abs/2609.29626
← Back to all articles
The output is a 27B model for two very unglamorous and very valuable jobs: RTL hardware design and GPU kernel optimization. Across seven benchmarks it leads RTLLM, ahead of GPT-5.5 and Claude Opus 4.8. It is second on CVDP and KernelBench L2, 16 points ahead of GPT-5.5 on the latter, and ties Opus 4.8 for the top TritonBench score. The authors also report competitive iterative optimization with far fewer tokens.
The interesting design choice is where the human sits. Not in the loop, turning knobs every hour, but upstream, in what the authors call a high-density, low-frequency interface. That is the skills file idea applied to research itself. The expert's judgment gets compiled once into procedures and boundaries, and then the agent runs many cheap iterations inside them. It is the same pattern as last month's AI-for-AI results, just aimed at a release-grade model rather than a toy.
Two caveats. The comparison models are a generation old now, GPT-5.5 and Opus 4.8, not this month's releases. And the paper, like most of this genre, reports what the agent achieved and says less about what it tried and threw away, which is exactly where you would look for the reward hacking another paper this week measured at 30% on open-ended research tasks. Still, a domain model beating general frontier models on hardware, built mostly by an agent, is a clear data point on where AI-for-AI is heading. It appeared in arXiv's September 25 listings.
Link: arxiv.org/abs/2609.29626
Comments