Blogs
The Room Where It Happened: What DAC 2026's Research Panels and Special Sessions Left Behind
Nov 16, 2026 · Frank Schirrmeister
Disclosure: I serve as Research Panels Chair of DAC 2027. This article is written in that capacity and reflects my personal views, not those of my employer, Synopsys.
You can still catch up
Before Long Beach I borrowed from Hamilton to describe the six research panels and eight research special sessions of DAC 2026. In the musical, Aaron Burr gets a whole song, "The Room Where It Happens," about a deal struck over a private dinner and about how much he wants to be on the inside. I called those fourteen sessions DAC's version of that room. Being there while DAC happens is still the best seat. Unlike Burr, though, the rest of us have other options.
More than three months later, I went looking for what those sessions left in writing. If you could not make it, how much of the conversation can you still catch up on today?
The short answer: more than I expected. One panel got two thoughtful write-ups, two special sessions left three papers between them, and the questions the other panels posed are still worth your time. This post is the reading list. It also explains why I care, because proposals for the DAC 2027 research panels and research special sessions are due Nov 18, 2026.
The panel with two write-ups
"Who Verifies the Verifier? Trust and Accountability in the Age of Autonomous EDA Agents" was moderated by Houman Homayoun, with Charles Alpert, Monica Farkash, Qiaoyan Yu, Ahmad-Reza Sadeghi and Jeyavijayan Rajendran on the panel. In July I singled out its abstract for the Ouroboros Effect: the worry that agents trained on the output of earlier agents slowly poison themselves.
Two public accounts tell us what came out of it.
Simon Davidmann's review in EE Times, published September 10, reports two positions without naming the speakers. From the AMD seat: what agents produce needs trust criteria of its own, and AI-generated IP is already on a path to tape-out. From the Cadence seat: agents may work right up to sign-off, but sign-off itself stays protected. Much EDA revenue already comes from tools that check what other tools construct, so independent checkers gain value as agents spread.
Simon's own summary of the pattern he trusts is one I expect to hear quoted for a while: "AI proposes while physics, formal methods, or simulation decides."
SemiHub's paper review of August 9 files the panel under a problem it calls verification oracles. Its reading is that the panel was about accountability more than performance. Who approves what an agent produced, and who answers when it is wrong?
The two accounts land in the same place. Generating designs is getting cheap. Judging them is where the work moves, and by some panel participant's argument, the business too.
Five questions still open for you
The other five panels left their questions, their rosters and a few proud posts from panelists and attendees. None of these debates was settled in a single session, and none has been closed in print. Each one is still going, and still yours to join:
New Minds, New Models: Who Will Drive AI-Led Design Innovation? Chip companies own the design data and EDA vendors own the toolchains. Who drives the next breakthrough?
Is the EDA Industry Fundamentally Ill-Prepared for True 3D Heterogeneous Integration? Can today's tools be adapted, or do they need a rewrite?
Who Holds the Upper Hand: Electrical or Photonic Links in Scaled-out Chiplet Systems? Do photonic links become essential as systems scale out, or do electrical links keep their lead?
Bridging Quantum, AI, and EDA: The Next Frontier of Design Automation. What does quantum simulation paired with AI change in design automation over the next decade?
GenAI for Aerospace and Defense Silicon: Design Authority or Design Assist? The title is the question.
So consider this an invitation. If you attended any of the five, or simply have a view on one of the questions, send me your comments and impressions. I would be glad to read them, to link written accounts, and to see the best arguments turn into panels for 2027.
What the special sessions left behind
The eight research special sessions have a built-in advantage here, because a special session comes with the option of a paper. Two of them used it, and you can read the results today.
A roadmap that names the bottleneck. The session "LLMs for Silicon: From Code Generation to Autonomous EDA and Layout Closure" had Andrew Kahng of UC San Diego on the program, with a roadmap toward an automated EDA R&D engineer. The paper, written with Amur Ghose and Bodhisatta Pramanik, sorts agentic EDA into three tiers by what the agent acts on: the flow, the tool's own code, or the design artifact.
Its conjecture is the part I keep coming back to. Per the authors, progress is now limited more by the systems stack around the model than by the model itself. They then list six missing capabilities, from verifying agent-written code to machine-readable tool interfaces. That is a research agenda, written down where anyone can pick it up.
A result that admits a miss. "When Generative AI Meets Analog Design: A Journey From Agent to Expert?" included PANDA, from Yibo Lin's group at Peking University. The paper describes an LLM that coordinates the path from design intent to analog layout. For its OTA example, gain and phase margin held after layout. Unity-gain bandwidth did not: 10.84 MHz before layout, 4.19 MHz after, against a target above 8 MHz.
I trust a paper more when it prints that row of the table. It also tells the next team exactly where to dig.
A survey that gets you up to speed. Ulf Schlichtmann of TU Munich was on the same LLM session with a talk on front-end design. About two weeks before the conference, he, Kangwei Xu and Bing Li posted a review on arXiv that traces the move from local assistance toward agentic execution in front-end tasks.
All three are a click away. Being in the room gave you early access.
And one from my old neighborhood. On the Engineering Track side, "AI and Digital Techniques Powering the Future of Analog Design" is the one special session for which I found a report of what a speaker said. Mike Gianfagna's SemiWiki piece of August 6 covers Bijan Kiani's talk and a booth demo.
The numbers it relays are still worth having. Per the article, Kiani cited Siemens and Wilson Research data in which analog blocks went from causing 10 percent of respins in 2005 to 40 percent in 2025, while digital fell from 40 to 15. And the framing rhymes with the panel: SPICE stays the sign-off authority, and the model decides where to point it.
Three take-aways
Read side by side, the write-ups and the papers agree more than their authors probably know.
Something deterministic gets the last word. The panel's Cadence position keeps sign-off protected. Simon's rule has physics, formal methods or simulation decide. PANDA's bandwidth miss was exposed by post-layout extraction, the deterministic step at the end of its flow. The Kiani talk keeps SPICE as the authority. That is four sources in three formats with one conclusion.
The bottleneck is moving from generating to judging. SemiHub says so about the panel. The Kahng paper says so in its own way: several of its six missing capabilities concern evaluation, from a ladder of checks that trades fidelity against cost to verification of code that agents write.
The work is in the system around the model. That is the Kahng conjecture, and its last missing capability is machine-readable interfaces. Simon's review argues separately that an English PDF is not a sufficient specification for AI-mediated design. Both point at interfaces built for agents to read.
None of the three take-aways was a session title. They show up when you lay the written record of several sessions side by side, which anyone can now do.
Leave something behind
That is my request to everyone preparing a proposal this week.
If you are proposing a research special session to Farinaz Koushanfar, who chairs them for DAC 2027, the call gives you the tool. Speakers can submit one combined manuscript, individual manuscripts, or none, each up to four pages plus a page of references, with technical review. The paper is optional. I would not treat it as optional.
If you are proposing a research panel to me, there is no manuscript, so the record depends on people. Tell me in the proposal who will write up the discussion, and where. It is not a field on the form. I will notice anyway.
Both calls close Nov 18, 2026.
And finally
A session lasts two hours at most. What gets written about it lasts longer and reaches people who never picked up a badge. That is good news for everyone who could not travel, and a reason to write more of it down in 2027.
So pick your question or your theme, call your people and send it in. DAC 2027 runs July 11 to 14 in San Jose.