acceptodds
Under review as a conference paper at ICLR 2027

GRODE-RTL: Source-Mapped Dependency Views for LLM-Based RTL Repair

Abstract

Large Language Models show great potential for automated Register-Transfer Level repair, yet they struggle with non-local faults where root causes and failing outputs reside far apart in source code. Treating RTL as a flat token sequence obscures hardware dataflows, whereas exposing complete synthesized graphs overwhelms the model with fine-grained metadata. To resolve this representation dilemma, we propose GRODE-RTL, an anchor-centric representation that extracts the single-clock-stage fan-in cone of a suspect register from the netlist, maps netlist nodes back to editable source lines, and organizes them into a compact hop-partitioned dependency view. By prioritizing direct drivers and folding distant hops, GRODE-RTL guides standard tool-using agents without exposing ground-truth hints or oracle feedback. We evaluate GRODE-RTL across 30 formally validated non-local faults using three LLMs (DeepSeek-V4-Flash, GLM-5.2, and MiniMax-M3) alongside a controlled six-configuration ablation study. GRODE-RTL improves pass@1 over source-level exploration by 16.7% on DeepSeek, by 26.7% on GLM-5.2, and by 20.0% on MiniMax-M3, with consistent gains on the hard-to-retrieve far-distance and indirect-anchor subsets of all three models. Our controls isolate structural relevance from generic context expansion, proving that selective dependency organization rather than full graph exposure is key to effective RTL repair.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Large Language Models, AI for Hardware, Graph Representation Learning, Program Representation, Agentic Code Editing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9zG8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRODE-RTL, a method for providing structured context to an LLM performing RTL repair. Given a suspicious signal, it extracts the signal’s fan-in cone from a synthesized netlist, maps the relevant logic back to source statements, and presents a compact dependency view to the agent. The evaluation covers 30 injected faults and three LLMs. The results show improvements over source-level exploration and the SeqFeed baseline, although most of the improvement in the ablation comes from selecting relevant source code.

Critical strengths
  1. The paper addresses a practical problem: code responsible for an incorrect signal can be difficult to find through source-level search, especially when related statements are far apart.

  2. Mapping circuit dependencies back to editable source locations makes the representation useful for repair. The approach can also be used with existing LLMs without additional training.

  3. The improvement over source-level exploration appears across all three evaluated models. The representation ablations and per-case outcomes provide useful information for understanding the results.

Critical weaknesses

I will summarize the main weaknesses and explain what I meant below.

W1. The technical contribution beyond established backward slicing and cone-of-influence analysis is not sufficiently clear.

W2. The evidence for the added value of dependency organization and compression is limited.

W3. The evaluation uses a small, restricted set of injected faults and assumes a useful anchor is already available.

W4. Some descriptions of the experimental protocol need clarification, particularly regarding the information and exploration budget available to each baseline.

Detailed comments:

  1. My main concern is what the method adds beyond using established program-analysis techniques to select context for an LLM. Starting from a suspicious signal and tracing backward through its dependencies is a familiar approach to narrowing the search space for debugging, error propagation analysis, taint analysis, etc. The paper also acknowledges that cone-of-influence reduction is standard.

  2. My understanding of the "representation", which was the main contribution of the paper, is the way the selected code and its dependencies are presented to the LLM. To demonstrate that such representation actually achieves better performance, it is not clear to me whether the dependency order, relationships, and other factors make sense in the first place. Why does providing such information to LLMs help?

  3. In my opinion, the method’s usefulness depends on obtaining an anchor whose dependency region contains the actual fault. The paper describes the anchor as the first upstream divergent sequential register, but provides limited detail on how this is determined. I would like to see a clearer description of anchor selection and some assessment of sensitivity to imperfect anchors.

  4. Table 3 seems to suggest that the benefits of the compression are quite limited.

Questions for authors
  1. Can you provide evidence (at least some intuition) that the proposed representation could affect the debugging success rate

  2. The evidence for the benefit of the compact representation is currently limited to an ablation on one model and 30 faults, where it improves over the cone-only configuration by one successful repair. Could the authors extend this comparison to the other two evaluated models and report results across multiple runs?

Reviewer_7p5DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to present relevant RTL context to an LLM repairing a fault distant from a suspicious signal. GRODE-RTL synthesizes the buggy design with Yosys, extracts a fan-in cone bounded by sequential state, maps it back to editable source statements, and presents a compact view organized by dependency depth and typed edges. It caps statements per hop and folds distant hops.

Evaluation covers 30 injected faults from 23 designs under a controlled-anchor protocol. Pass rates improve substantially over source exploration on all three backbones (by roughly 17 to 27 percentage points). The ablation suggests that relevant cone statements account for most of the gain, while compact organization adds one net successful case over the plain cone. Success is measured by a 20-step bounded SAT check against the reference design.

Critical strengths
  1. The representation usefully connects hardware dependencies to actionable edits. Mapping a synthesized fan-in cone back to source statements addresses a practical obstacle, since knowing a suspicious signal does not directly identify which lines to inspect. The state boundary and source locations give the method a concrete role in an agent's repair workflow without model training (Sections 3.2-3.5).

  2. The reported gains are also consistent across three backbones, with useful controls. GRODE-RTL outperforms Explorer and SeqFeed on each model (Table 1), including the far-distance and indirect-anchor subsets, while using substantially fewer tokens than SeqFeed. The cone, random-context, and full-graph controls help assess where the gains come from.

Critical weaknesses
  1. The central issue is that the evidence does not yet establish the value of the distinctive representation choices. In Table 3, GRODE-RTL solves only one more case than the plain cone overall, and one more in each harder subset. Compact organization also uses somewhat more total tokens and view characters than the cone (its compression savings are relative to the full graph). With one sample per case and no uncertainty estimates, the incremental benefit remains unclear. The largest gain also needs a clearer account of timeouts. Counting Table 4, timeouts are frequent on GLM, and Explorer timeouts account for more than half of GRODE-RTL's paired wins there. Since the paper acknowledges relay timeouts, the results need to distinguish infrastructure noise from agent behavior. The current evidence supports relevant cone context more strongly than a robust benefit from compact organization, conditional on the selected faults lying within the anchor's cone.

  2. Several protocol and reporting inconsistencies also make the comparisons difficult to interpret. The stated 12-turn cap (Section 4.2) conflicts with the example submitting at turn 13 (Appendix A.6). The shared-anchor description is also unclear, since the base prompt denies a preselected suspect signal while the structural configurations receive an anchor-returning tool (Appendix A.5). Whether Explorer received equivalent localization evidence materially affects the comparison. The MiniMax no-edit counts differ between Tables 1 and 4, and the treatment of timeouts in the conditional success rates is unclear. These may reflect different definitions, but patch status at timeout is needed to reconcile them. The case study also claims a targeted read around faulty line 629, while the listed trace exhausts its exploration budget before obtaining the graph and shows no such read. Finally, a 20-step bounded SAT pass establishes agreement within that horizon under the encoded assumptions. The semantic audit mentioned in Figure 4 is not described, leaving the scope of the successful-repair claims unclear.

Questions for authors
  1. Does SeqFeed include both SeQuery and SeGraph, and did every condition, including Explorer, receive the same anchor? SeGraph already traces static RTL dependencies, so its characterization and the baseline implementation need clarification. I do not treat the overlap with this recent work as a reason for rejection.

  2. Please reconcile the turn limit, the claimed read of line 629, and the no-edit counts. How are timeouts before and after a patch recorded, and what denominators define conditional repair success?

Reviewer_AP7oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRODE-RTL proposes a source-mapped dependency representation for LLM-based RTL repair. Starting from a suspicious signal, it extracts a single-stage fan-in cone from a Yosys netlist, maps the relevant logic back to RTL source lines, and organizes the resulting statements by dependency depth and relation type. A compact view prioritizes direct drivers and summarizes more distant dependencies to guide source inspection and local edits.

Critical strengths
  1. This paper presents a practical connection between circuit structure and source-level repair. This is a useful systems contribution because it connects existing hardware analysis tools to the actual decisions a repair agent must make.

  2. The cross-model results and the ablations provide useful initial evidence about which context is helpful.

Critical weaknesses
  1. The evaluation does not establish the reliability of the gains or the incremental benefit of the compact representation. This paper evaluates only 30 injected faults from 23 designs, with one run per case and configuration at temperature 1.0. No confidence intervals or significance tests are reported, although the supplementary evaluator supports repeated attempts and paired comparisons. With this sample size, a single outcome changes pass@1 by approximately 3.3 percentage points.

  2. The evaluation relies on a suitable anchor, while the scope and consistency of that localization assumption remain insufficiently established. The author explicitly restricts the method to a single-stage fan-in cone and evaluates repair conditional on a prelocalized anchor. A fault in an earlier pipeline stage is outside this view unless localization first selects an appropriate upstream register. This is a reasonable scope choice, but the paper does not evaluate incorrect or ambiguous anchors, localization failures, or root causes outside the selected cone. Consequently, the results establish less about end-to-end RTL debugging than some of the broader claims suggest.

Questions for authors
  1. Did Explorer, SeqFeed, and GRODE-RTL receive the same anchor identity and the same reference/current behavioral evidence in the experiments reported in Table 1?

  2. Are repeated-run results or complete per-case ablation outcomes available? In particular, does the compact representation consistently outperform cone-only across independent runs, and what uncertainty surrounds that difference?

  3. Were any candidate faults excluded because localization failed or because the fault lay outside the selected anchor’s cone? How were such cases accounted for, and is there any evaluation showing how repair success changes when the supplied anchor is imperfect?

Comments on the paper

2 comments

Dear Reviewer,

Thank you again for your time and constructive feedback on our manuscript. As the ICLR rebuttal period is approaching its end in 3 days, we wanted to kindly follow up to ensure our previous responses have adequately addressed your concerns.

If you have any further questions or require any additional experiments, please feel free to let us know at your earliest convenience so that we have sufficient time to conduct them before the deadline. We truly appreciate your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Summary of Rebuttal Responses, Additional Experiments, and Revisions

Summary of Responses and Revisions

We thank all reviewers for their time and constructive feedback. Below is a brief overview of our responses and revisions across 990 additional evaluation episodes, with references to the corresponding reviewer questions. All experimental updates and clarifications have been incorporated into the revised manuscript.

Repeated evaluations strengthen the evidence for structural repair context. We added repeated-run experiments across the original backbones and report mean repair rates with standard deviations. GRODE-RTL consistently outperforms Explorer in these runs. The close performance of cone-only and GRODE-RTL on several backbones supports the value of their shared cone-extraction and source-mapping components. (Reviewer AP7o, Q2; Reviewer 9zG8, Q2.)

Additional backbones clarify the benefit of compact organization. We extended the compact-versus-cone comparison to Qwen, Kimi, and Gemini. Compact records additional successful repairs on each newly evaluated backbone, with larger observed gains on Gemini and MiniMax. These results provide complementary evidence that organizing dependencies can help beyond selecting relevant source context, while showing that the size of this benefit varies across models. (Reviewer AP7o, Q2; Reviewer 9zG8, Q1 and Q2.)

Baseline information and the role of anchor selection are clarified. We distinguish the initial register exhibiting a verification mismatch from the analysis anchor selected within the structural pipeline. Explorer receives the initial failure evidence but is not additionally supplied with GRODE-RTL’s selected analysis anchor. We also clarify that SeqFeed includes both SeQuery and SeGraph and already supports structural dependency analysis. Our distinction concerns how circuit analysis is integrated into the repair workflow and presented as editable source context. (Reviewer AP7o, Q1; Reviewer 7p5D, Q1.)

The localization assumption and benchmark scope are made explicit. We explain the difference between an observed mismatch location and a suitable upstream analysis anchor, and clarify that the benchmark evaluates faults within the associated anchor’s cone.We clarify that the reported repair results are conditioned on the selected analysis anchor, separating the effectiveness of the repair representation from the quality of upstream localization. (Reviewer AP7o, Q3.)

Protocol and outcome reporting are reconciled. We clarify turn accounting, correct the case-study description to match the recorded tool calls, and explain the distinction between terminal no-edit labels and unchanged final source. We also specify how timeout episodes enter the reported outcome categories and conditional repair denominators. These clarifications separate candidate correctness, source changes, and submission status. (Reviewer 7p5D, Q2.)

We sincerely appreciate the reviewers’ suggestions. They helped us strengthen the experimental evidence and more clearly distinguish the contribution of the core source-mapped structural context from the additional benefit of compact dependency organization.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.