Auto-Arch Tournament vs Structured Output Benchmark

Auto-Arch Tournament is an autonomous research system where an AI agent iteratively proposes, implements, and validates microarchitectural improvements to a RISC-V CPU. Starting from a standard 5-stage pipeline, the loop runs hypotheses in parallel, each going through formal verification (53 symbolic checks), cycle-accurate simulation, multi-seed FPGA place-and-route, and CoreMark CRC validation. Only hypotheses that beat the current champion get merged; everything else gets discarded. Starting from 301 iterations/second, the system hit 577 iter/s (+92%) across 73 attempts in 9.8 hours — producing a design 26% faster and 40% smaller in LUTs than the baseline. The insight the author drives home is that the real innovation isn't the AI agent — it's the verifier. The orchestrator is hardcoded to prevent agents from manipulating their own evaluation gates, a simple but critical design constraint that turns a creative process into a trustworthy one. Without a rigorous verification harness, agent-driven optimization becomes a confidence trick. This is early but fascinating proof that AI-driven hardware design loops can produce commercially meaningful gains. The repo uses Claude Code or Codex as the coding agent, SystemVerilog for the RTL, and standard open-source EDA tooling (Yosys, nextpnr, Verilator). It's a compelling template for anyone building agentic optimization loops where correctness matters.

Interfaze's Structured Output Benchmark (SOB) exposes a gap that has been quietly breaking production AI pipelines: models can produce syntactically valid JSON while getting the actual values wrong. SOB measures value accuracy across 21 models using 5,000 text passages, 209 OCR documents, and 115 meeting transcripts — scoring each on seven metrics including value accuracy, faithfulness (grounding vs. hallucination), type safety, and perfect-response rate. The benchmark reveals some sobering findings. Even top models like GPT-5.4 and Claude Sonnet 4.6 achieve ~83% on text but drop to 67% on images and only 23.7% on audio. No single model dominates all modalities — GPT-5.4, GLM-4.7, Qwen3.5-35B, and Gemini 2.5 Flash cluster within one point of each other on text. Perfect response rates (all seven metrics correct) rarely exceed 50% for even the best performers. For developers building data extraction pipelines, agents that read invoices, or any system where "correct JSON" means more than syntactically valid JSON, this is required reading. The dataset is on Hugging Face, the paper is on arXiv, and the playground lets you test your own model's structured output capability directly.