← News

ResearchSources: arXiv · 20 Nov 2026

GRODE-RTL maps register dependencies to source lines to guide LLM hardware repair

A new arXiv paper proposes an anchor-centric code representation that helps large language models fix non-local faults in Register-Transfer Level designs, reporting pass@1 gains across three models.

In plain words

Chip designers describe hardware in RTL code, and researchers are testing whether large language models can automatically fix bugs in it. The paper's authors say models struggle when a bug's cause and its visible failure sit far apart in the code. Their method extracts only the circuit logic feeding a suspect register and maps it back to editable source lines, so the model sees a focused view instead of the whole design.

What was announced

The authors argue that treating RTL as a flat token sequence hides hardware dataflows, while exposing complete synthesized graphs buries the model in fine-grained metadata. GRODE-RTL addresses this by extracting the single-clock-stage fan-in cone of a suspect register from the netlist, mapping netlist nodes back to editable source lines, and organizing the result into a compact hop-partitioned dependency view. Prioritizing direct drivers and folding distant hops guides standard tool-using agents without ground-truth hints or oracle feedback, according to the paper.

Evaluation covered 30 formally validated non-local faults with three LLMs: DeepSeek-V4-Flash, GLM-5.2, and MiniMax-M3. Reported pass@1 improvements over source-level exploration were 16.7% on DeepSeek, 26.7% on GLM-5.2, and 20.0% on MiniMax-M3, with gains on far-distance and indirect-anchor subsets. A six-configuration ablation study, the authors state, isolates structural relevance from generic context expansion. These figures are the authors' own and have not been independently confirmed.

What it does not show

  • Evidence is limited to the paper's abstract; full methodology and results are not available for verification.
  • The 30-fault benchmark and three-model results are author-reported and not independently confirmed.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.