Proving Ground: Agentic AI for Design Verification Challenge
Registration- Competition Dates:
- Ten-day online window, Saturday, November 7, 2026, 09:00 to Monday, November 16, 2026, 23:59 Hong Kong Time (UTC+8), the first day of ICCD 2026. Practice designs available from October 19, 2026.
- Location:
- Online, worldwide. Results announced November 17, 2026. Awards at the ICCD 2026 closing session, Hong Kong SAR, China, November 18, 2026 (winners who cannot attend are recognized online).
- Registration Deadline:
- November 10, 2026. Registration opens October 5, 2026.
Overview
The Proving Ground: Agentic AI for Design Verification Challenge at ICCD 2026 invites participants worldwide to put AI agents to work on the hardest part of chip design: proving that a change to a design is correct. Over a ten-day window that ends on the first day of ICCD 2026, participants receive three real register-transfer level (RTL) designs, released one at a time, each with a specification, a public testbench and a change request. Each baseline also contains a small number of specification deviations that the public testbench does not catch.
Using any AI agent, language model, coding assistant or tool of their choice, participants find and fix the deviations, implement the change request, and deliver the modified RTL together with an evidence package: the tests and assertions that show the design is correct.
Every submission is run through the Flow, the organizers' automated backend: a hidden regression suite and hidden formal properties check correctness, mutation analysis measures how much of the design the participant's own tests and assertions actually cover, and a standardized synthesis and place-and-route flow reports power, performance and area (PPA) against the baseline. Results appear on a live leaderboard throughout the window, and participants may submit as many improved versions as they like within the limits below.
Submission Format
A submission is one archive containing the modified RTL in the released directory layout, a tests/ directory with the participant's testbenches (Verilator or cocotb), an assertions/ directory with SystemVerilog assertions bound to the design, and a one-page report.md describing what was found, what was changed and how it was verified. The released flow image runs the same public checks locally, so a submission can be tested before it is uploaded.
Scoring
Each submission receives three measurements from the Flow.
Correctness (C, 0 to 1) is the fraction of hidden regression tests and hidden properties that pass, including those for the change request.
Verification quality (V, 0 to 1) is the mutation score: the Flow injects forty single-fault mutants into the submitted RTL (operator swaps, constant changes, negated conditions, dropped resets and enables, off-by-one bounds) and counts how many the participant's tests and assertions detect.
PPA (P, 0 to 1) is 0.5 for a design equal to the baseline after place and route, higher for lower area, better timing and lower power, lower for regressions, with a cap of 0.3 for any design that fails its timing constraint.
The submission score is 100 × C × (0.5 V + 0.5 P).
A participant's design score is the best of their submissions on that design, and the total is the sum over the three designs. A submission that leaves the baseline untouched scores roughly 30; a fully correct, well-verified and improved design scores above 80.
Feedback and Leaderboard
Every submission returns its C, V and P within about fifteen minutes, with the list of hidden checks that failed (by name, not by test content), the mutants that survived (by location), and the PPA report at every stage of the flow: synthesis, placement, clock tree, routing. The leaderboard updates live with each participant's total and best per design.
One submission per participant per design every thirty minutes, up to thirty per design over the window. Office hours with the organizers are held on Zoom twice a week during the window; a questions board is open on the platform throughout.
Schedule
Design 1 (a datapath block) is released November 7; Design 2 (a control block with a bus interface) on November 10; Design 3 (a small pipelined core) on November 13.
All three remain open until the window closes on November 16 at 23:59 Hong Kong Time. The Flow finishes queued submissions within the hour; the leaderboard is final on the morning of November 17.
Awards and Recognition
The top three totals at the close of the window receive cash awards and formal recognition at the ICCD 2026 closing session:
- 1st Place:$1,000 (USD)
- 2nd Place:$750 (USD)
- 3rd Place:$500 (USD)
The Explorer award and a separate prize go to the participant with the largest number of distinct valid submissions (distinct in the Flow's netlist comparison, correctness above 0.9), recognizing the widest verified exploration of the design space. All participants with a total above 150 receive certificates.
Eligibility
Open to all participants globally. Individuals, no teams; one account and one leaderboard entry per person. Any AI agents, language models, coding assistants, EDA tools and scripts may be used, on the participant's own machines and accounts; the submission is judged, not the method. No prior verification experience is required: the specification, the public testbench and the flow image are enough to start. Organizers and their students are not eligible.
Registration
Registration is required to participate. Participants must register before November 10, 2026, via the competition portal [registration link attached below]. Participants are instructed to sign up, add the required registration details and, once logged in to the competition page, enroll in "ICCD 2026 Proving Ground: Agentic AI for Design Verification Challenge". Registered participants receive the practice designs, the flow image and a worked example submission on October 19, 2026. Design 1 is released to registered participants on November 7 at 09:00 Hong Kong Time.
Organizers
Kevin Immanuel Gubbi,
Mohammadnavid Tarighat,
Houman Homayoun,
Soheil Salehi,
Farinaz Koushanfar
kgubbi@ucdavis.edu