Agents for RTL
Putting agents to work on hardware — generating Verilog, then proving it wrong.
Hardware design is an unusually honest setting for agents: a simulator will tell you, without negotiation, whether the thing you generated works. Three projects build on that feedback signal, moving from generation to verification to learning from execution.
MAGE — multi-agent RTL generation
A multi-agent engine for automated RTL code generation, splitting the task across specialised agents rather than asking one model to emit a whole module in a single shot.
Published at DAC 2025 · paper · code
PRO-V-R1 — trainable verification agents
Generating RTL is only half the problem; someone has to write the testbench that catches the bugs. PRO-V-R1 is the first open-source trainable agentic framework for autonomous RTL verification, reaching 57.7% functional correctness and 34.0% on robust fault detection — competitive with proprietary models on several benchmarks.
LLM4Cov — learning under expensive execution
Execution-aware agents learn from tool feedback, but in hardware that feedback comes from industrial simulators: slow, costly, and non-differentiable. LLM4Cov is an offline agent-learning framework built around that constraint — execution-validated data curation, policy-aware agentic data synthesis, and worst-state-prioritized sampling.
A compact 4B model reaches a 69.2% coverage pass rate under agentic evaluation, beating its own teacher by 5.3 points and holding its own against models an order of magnitude larger.