This page may contain stale information. Last updated: 2026-07-05

Definition

HORIZON is a nvidia Research self-evolving agent framework that treats hardware design (especially RTL) as repository-level code evolution over isolated git worktrees.

Architecture

  1. Markdown harness: Goal, domain knowledge, evaluator, acceptance predicate
  2. Bootstrap agent: Compiles harness into project pack + git/runtime policy
  3. Hands-free loop: Edit worktree → run evaluator → commit on pass or log rejection
  4. Git as control plane: Commits mark checkpoints; diffs and logs form replayable search history

Evaluator (RTL Domain)

Compilation, simulation, coverage extraction, assertion/testbench checks — slot is domain-agnostic (unit tests, theorem provers, synthesis tools in other domains).

Benchmark Results (GPT-5.3)

100% pass rate on ChipBench, RTLLM-2.0, Verilog-Eval-v2, and nine CVDP categories. CVDP CID 012 improved from 47.8% → 100% through iterative evolution.

100% applies to evaluated benchmark suites only. One ChipBench miss was harness defect, not agent failure. Researchers explicitly state agentic hardware design is NOT solved — benchmarks are controlled proxies.

Sources