cs.AROct 7, 2026
SaveWhen Algorithmic Exploration Becomes Cheap: A Case Study of Agentic Research in EDA
Organizations: College of Integrated Circuits and Micro-Nano Electronics Fudan University, Shanghai, China
Abstract
As EDA researchers, we conducted eight deliberate trials of agentic algorithm exploration, selecting several topics outside our areas of depth. One faculty member and seven students participated, including students without publication experience. With limited intervention in the algorithms, agents developed mathematical constructions, analyzed existing tools, and implemented improvements; some efforts fell short of their practical goals. We also used AI to collect, classify, and analyze 8,420 papers from four EDA conferences and two journals over 2022-2026. Among 2,380 primary-core papers, we classified 97.7% from titles and abstracts as computationally closed, including work on new formulations. Together, these observations suggest that much of EDA offers an executable environment for increasingly accessible algorithm research. We see an opportunity for tool developers to investigate ideas they previously lacked time to pursue. We also ask how EDA should validate and reward research when results become easier to produce than to examine, and what papers and venue labels will continue to tell us about a contribution.
Figures & tables
Figure 1. A deliberate set of EDA research trials. We selected unfamiliar questions and delegated much of the algorithmic exploration. The outcomes are considered through their contribution to tools and transferable understanding. A faculty member and students choose unfamiliar EDA questions. An executable research loop connects to an improved EDA tool and a reusable research result.
Figure 2. Pressure is applied to acceptance, not to a prescribed algorithmic route. a , Premises sustain autonomous search; independent evaluation determines whether a candidate meets the gate. Falling short returns work to exploration. b , Requirements tighten from basic theory to useful implementation, while engineering feedback can reopen the mathematical question. The prompts and evaluation mechanisms vary by task. Branching research paths approach a fixed gate controlled by an independent evaluator. One route stops, another passes, and a rejected candidate loops back to exploration. A second panel depicts theory, construction, and engineering with feedback returning to theory.
Figure 3. Two routes into unfamiliar research. a , Rooted spanning trees trade total length against root-to-terminal distance; highlighted paths illustrate the tradeoff (schematic edges, rectilinear lengths). b , Tracing an existing tool’s operators led to an exact reduction from 40 actions to 31 ( 2 ) . The question and the codebase supplied different starting knowledge. Two schematic trees contrast shared edges with shorter root paths. A second panel shows source-code analysis leading to a reduced menu of logic transformations.
Figure 4. Three obstacles between an intermediate result and practical benefit: model mismatch in remapping, full-solver cost in clock-tree realizability, and interactions among placement components. The clock-tree panel combines two comparisons: stronger certificates reduced calls, but a monolithic baseline was faster. Output-preserving logic screening and rewriting offered a more direct route to runtime gains. Three illustrated comparisons show a mathematical objective versus a physical circuit, a tree certificate versus a full-model solver, and separate mathematical components versus an integrated placement. Each locates a different obstacle to practical improvement.
Figure 5. Early progress and longer follow-up across the trials. a , Rooted spanning-tree branch windows, including pauses; the diamond marks the first height-partition construction at 4.8 hours. b , Recorded project windows; diamonds mark the rewriting direction choice (day 3), the first complete contraction tool study (day 5), and LUT C++ campaign entry (day 7). Placement starts at its final-score phase. Endpoints are listed in the companion process data. Rooted spanning-tree branches have different outcomes after windows of 5.9 to 23.1 elapsed hours. Six longer projects span 23 to 62 calendar days; three diamonds show milestones early in their respective windows.
Figure 6. A new formulation need not open the research loop. An existing objective and a newly specified constraint can both be investigated computationally; acquiring new physical information requires another source of feedback. The drawings illustrate these distinctions. Counts are from the primary-core EDA corpus. Two document-to-computer feedback loops contrast an existing objective with an added computational rule. A third panel shows a physical probe providing new measurements to a computer. The first two categories comprise 97.7 percent of primary-core papers.
Figure 7. Computational closure is widespread across core EDA. Counts use primary EDA topics, with Late Breaking Results excluded and accepted Early Access articles retained. Blue and muted red-brown together account for 2,325 of 2,380 papers. The red-brown category includes new objectives and constraints that remain computationally testable. Track labels are shortened from the DAC taxonomy. Stacked horizontal bars show the three workflow classes in nine core EDA tracks. Physical design has 450 papers, the largest count, including 173 new computational formulations.
Figure 8. An uneven change in recurring author portfolios. The same 45 last-author names appear in DAC, DATE, and ASP-DAC in all three years. a , Paper counts and distinct first–last pairs. b , Established formulations (blue), new computational formulations (red-brown), and external intervention (light blue). Explicitly LLM-related papers and Late Breaking Results are excluded. A fixed cohort of 45 last authors has 100, 104, and 135 papers in 2024, 2025, and 2026; the accompanying bars separate the three workflow classes.
Figure 9. The industrial opportunity includes improvements within existing tools. Cheaper exploration makes more investigations into scheduling, data structures, and optimization worth attempting against production workloads. Small gains can be reused across many designs. Three algorithm-development workstations feed improvements into a shared industrial EDA tool with production feedback.
Figure 10. Public-data observations for subsequent EDA cohorts. Track coverage, recurring methods and benchmarks, collaboration, and diffusion together; link preprints to their later venue versions. Four panels illustrate problem coverage, repeated method-benchmark pairs, first-last author collaboration, and the spread of research combinations between groups.