UniRTL: Unifying Code and Graph for Robust RTL Representation Learning

Hongji Zhang; Lei Chen; Mingxuan Yuan; Qiang Xu; Yi Liu

arxiv: 2605.31040 · v1 · pith:KQRBOAUInew · submitted 2026-05-29 · 💻 cs.LG

UniRTL: Unifying Code and Graph for Robust RTL Representation Learning

Yi Liu , Hongji Zhang , Lei Chen , Mingxuan Yuan , Qiang Xu This is my paper

Pith reviewed 2026-06-28 23:46 UTC · model grok-4.3

classification 💻 cs.LG

keywords RTL representation learningmultimodal pretrainingcode and graph fusionhardware design automationperformance predictioncode retrievalcontrol data flow graphregister transfer level

0 comments

The pith

UniRTL learns unified RTL representations by jointly pretraining on code and its control data flow graph.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that single-modality methods for representing register transfer level designs fall short because code alone misses structural relations while graphs alone miss explicit semantics. It introduces a pretraining approach that aligns the two modalities at fine grain through mutual masking and builds the alignment in stages starting from text summaries and code. Experiments on performance prediction and code retrieval show consistent gains over prior single-modality baselines. If the claim holds, hardware design tools can rely on representations that capture both functional intent and structural behavior without manual feature engineering.

Core claim

UniRTL is a multimodal pretraining framework that learns unified RTL representations by jointly leveraging code and CDFG, achieving fine-grained alignment through mutual masked modeling and a hierarchical training strategy that first uses a pretrained graph-aware tokenizer and staged alignment of text and code before integrating the graph.

What carries the argument

UniRTL framework that performs mutual masked modeling between code and CDFG inside a staged training schedule that begins with text-code alignment and then adds graph integration.

If this is right

Representations trained on both modalities produce higher accuracy than code-only or graph-only baselines on circuit performance prediction.
The same representations improve retrieval of functionally similar RTL modules compared with prior single-modality methods.
The staged training schedule allows the model to absorb structural information without overwriting semantic signals learned earlier.
The resulting representations can serve as a shared foundation for multiple downstream hardware automation tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same alignment technique could be tested on other hardware description languages or on netlist-level graphs to check whether the benefit is specific to RTL.
If the joint representations reduce the amount of labeled data needed for new tasks, they could lower the barrier for applying machine learning to smaller design teams.
Extending the mutual masking objective to include timing or power reports might further enrich the learned features without changing the core architecture.
The hierarchical schedule offers a template for adding additional modalities such as simulation waveforms once the code-graph alignment is stable.

Load-bearing premise

The control data flow graph and the RTL code supply complementary information whose integration is required for a complete understanding of the design.

What would settle it

A controlled experiment in which an ablation that removes either the code modality or the graph modality matches or exceeds UniRTL accuracy on both performance prediction and code retrieval would falsify the necessity of the joint approach.

Figures

Figures reproduced from arXiv: 2605.31040 by Hongji Zhang, Lei Chen, Mingxuan Yuan, Qiang Xu, Yi Liu.

**Figure 1.** Figure 1: Example data point from our dataset, including RTL source code, and its corresponding functional summary and CDFG. For comparison, data flow (Guo et al., 2021) and register-level sub-circuit (Fang et al., 2025) are also shown, demonstrating the completeness of the constructed CDFG. strategy is limited: it merely identifies variable nodes in the code without capturing their full semantic relationships. More… view at source ↗

**Figure 2.** Figure 2: Overview of UniRTL. The framework achieves fine-grained cross-modal alignment via mutual masked modeling, and adopts a hierarchical training strategy: a graph-aware tokenizer is first pretrained, and text-code alignment is performed prior to graph incorporation. mlm (Feng et al., 2020) 1 as our base model, pretrained on the CodeSearchNet (Husain et al., 2019) code corpus using masked language modeling (Dev… view at source ↗

**Figure 3.** Figure 3: Qualitative analysis of code–graph relationships learned by mutual masked modeling. (a) Nodes with the highest attention scores from code tokens. (b) Code tokens with the highest attention scores from graph nodes. Corresponding code token–graph node pairs are highlighted with the same background color. 4.5. Qualitative Analysis of Code–Graph Relationships During the graph incorporation stage, the mutual ma… view at source ↗

**Figure 4.** Figure 4: Representativeness analysis of the CDFG-convertible subset. We compare the full pretraining dataset and the subset of designs successfully converted into CDFGs using token-count distributions and UniRTL embedding distributions. A. Analysis of CDFG Conversion Failures and Subset Representativeness In Section 3.1, we note that not all collected RTL designs can be successfully converted into CDFGs. Many desig… view at source ↗

read the original abstract

Developing effective representations for register transfer level (RTL) designs is crucial for accelerating the hardware design workflow. Existing approaches, however, typically rely on a single data modality, either the RTL code or its associated graph-based representation, limiting the expressiveness and generalization ability of the learned representations. For RTL, the control data flow graph (CDFG) offers a comprehensive structural representation that preserves complete information, while the code modality explicitly encodes semantic and functional information. We argue that integrating these complementary modalities is essential for a thorough understanding of RTL designs. To this end, we propose UniRTL, a multimodal pretraining framework that learns unified RTL representations by jointly leveraging code and CDFG. UniRTL achieves fine-grained alignment between code and graph through mutual masked modeling and employs a hierarchical training strategy that incorporates a pretrained graph-aware tokenizer and staged alignment of text (i.e., functional summary) and code prior to graph integration. We evaluate UniRTL on two downstream tasks, performance prediction and code retrieval, under multiple settings. Experimental results show that UniRTL consistently outperforms prior methods, establishing it as a more robust and powerful foundation for advancing hardware design automation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

UniRTL's mutual masked modeling plus staged hierarchical training for code-CDFG alignment is the concrete new piece, but the abstract supplies no experiments, baselines or metrics so the outperformance claims cannot be checked.

read the letter

The main takeaway is that this paper puts forward a multimodal pretraining setup for RTL that aligns code with its control data flow graph using mutual masked modeling, then layers in a staged process that first uses a pretrained graph-aware tokenizer and aligns text summaries before bringing in the graph.

That combination and the staged pipeline are the actual additions beyond routine multimodal extensions.

The motivation is laid out clearly: CDFG supplies structure while code supplies semantics, so joint training should help. The hierarchical approach is a sensible implementation choice to avoid forcing coarse alignment too early.

The soft spot is straightforward. The abstract states that UniRTL outperforms prior methods on performance prediction and code retrieval but gives no datasets, no baseline descriptions, no metrics, and no numbers. Without those details it is impossible to tell whether the results hold or whether the graphs were extracted in a way that leaks information.

The central assumption that CDFG captures complete structural information is stated but not examined in the provided text.

This work is aimed at the small group doing machine learning for electronic design automation. A reader already following RTL representation papers would get the idea and could decide whether to adapt the alignment trick.

It is worth sending to peer review so the experimental section can be evaluated; the framing is coherent enough that a referee should see the full evidence.

Referee Report

1 major / 0 minor

Summary. The paper proposes UniRTL, a multimodal pretraining framework for RTL designs that jointly leverages code and control data flow graph (CDFG) modalities via mutual masked modeling for fine-grained alignment, a hierarchical training strategy with a pretrained graph-aware tokenizer, and staged alignment of text and code prior to graph integration. It evaluates the approach on performance prediction and code retrieval tasks under multiple settings and claims consistent outperformance over prior methods, positioning it as a robust foundation for hardware design automation.

Significance. If the empirical claims hold under rigorous validation, the work could meaningfully advance RTL representation learning by demonstrating the value of complementary code and structural modalities, potentially improving generalization in hardware design tasks.

major comments (1)

Abstract: the central claim of consistent outperformance on performance prediction and code retrieval is stated without any experimental details, baselines, metrics, datasets, or data-handling procedures, rendering it impossible to assess whether the data supports the claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for greater specificity in the abstract. We agree that providing key experimental context will strengthen the presentation and will revise the abstract accordingly in the next version.

read point-by-point responses

Referee: Abstract: the central claim of consistent outperformance on performance prediction and code retrieval is stated without any experimental details, baselines, metrics, datasets, or data-handling procedures, rendering it impossible to assess whether the data supports the claims.

Authors: We acknowledge that the current abstract is concise and omits specific details. In the revision we will add a brief clause noting the primary datasets (RTL benchmarks from OpenCores and industrial sources), representative baselines (CodeBERT, GraphCodeBERT, and prior RTL-specific models), core metrics (MAE/RMSE for performance prediction; Recall@1/5/10 and MRR for retrieval), and that all experiments follow standard train/validation/test splits with results reported in Section 4. This keeps the abstract within length limits while enabling readers to gauge the claims. Full experimental protocols remain in the main body. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents UniRTL as a multimodal pretraining framework that jointly leverages RTL code and CDFG via mutual masked modeling and staged alignment, with performance claims resting on empirical evaluations across downstream tasks (performance prediction, code retrieval). No load-bearing step reduces a prediction or result to its own inputs by construction, no self-definitional equivalence appears in the described methodology, and any self-citations are not invoked to justify uniqueness or force the central architecture. The derivation remains self-contained against external benchmarks and experimental results.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Ledger constructed from abstract only; no explicit free parameters, invented entities, or additional axioms beyond the core domain assumptions are stated.

axioms (2)

domain assumption CDFG offers a comprehensive structural representation that preserves complete information for RTL designs.
Invoked to justify the graph modality as complementary to code.
domain assumption Code modality explicitly encodes semantic and functional information.
Stated as the reason code is needed alongside the graph.

pith-pipeline@v0.9.1-grok · 5734 in / 1202 out tokens · 15916 ms · 2026-06-28T23:46:16.825636+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 9 canonical work pages · 8 internal anchors

[1]

Qwen Technical Report

Bai, J., Bai, S., Chu, Y ., Cui, Z., Dang, K., Deng, X., Fan, Y ., Ge, W., Han, Y ., Huang, F., et al. Qwen technical report.arXiv preprint arXiv:2309.16609,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

Speech- bert: An audio-and-text jointly learned language model for end-to-end spoken question answering

Chuang, Y .-S., Liu, C.-L., Lee, H.-y., and Lee, L.-s. Speech- bert: An audio-and-text jointly learned language model for end-to-end spoken question answering. InProc. In- terspeech 2020, pp. 4168–4172,

2020
[3]

Bert: Pre-training of deep bidirectional transformers for lan- guage understanding

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for lan- guage understanding. InProceedings of the 2019 confer- ence of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186,

2019
[4]

Codebert: A pre-trained model for programming and natural lan- guages

Feng, Z., Guo, D., Tang, D., Duan, N., Feng, X., Gong, M., Shou, L., Qin, B., Liu, T., Jiang, D., et al. Codebert: A pre-trained model for programming and natural lan- guages. InFindings of the Association for Computational Linguistics: EMNLP 2020, pp. 1536–1547,

2020
[5]

The Llama 3 Herd of Models

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

work page internal anchor Pith review Pith/arXiv arXiv
[6]

DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence

Guo, D., Zhu, Q., Yang, D., Xie, Z., Dong, K., Zhang, W., Chen, G., Bi, X., Wu, Y ., Li, Y ., et al. Deepseek-coder: When the large language model meets programming–the rise of code intelligence.arXiv preprint arXiv:2401.14196,

work page internal anchor Pith review Pith/arXiv arXiv
[7]

CodeSearchNet Challenge: Evaluating the State of Semantic Code Search

Husain, H., Wu, H.-H., Gazit, T., Allamanis, M., and Brockschmidt, M. Codesearchnet challenge: Evaluat- ing the state of semantic code search.arXiv preprint arXiv:1909.09436,

work page internal anchor Pith review Pith/arXiv arXiv 1909
[8]

Nv-embed: Improved techniques for training llms as generalist embedding models

Lee, C., Roy, R., Xu, M., Raiman, J., Shoeybi, M., Catan- zaro, B., and Ping, W. Nv-embed: Improved techniques for training llms as generalist embedding models. In International Conference on Learning Representations, volume 2025, pp. 79310–79333,

2025
[9]

Craftrtl: High- quality synthetic data generation for verilog code models with correct-by-construction non-textual representations and targeted code repair

Liu, M., Tsai, Y .-D., Zhou, W., and Ren, H. Craftrtl: High- quality synthetic data generation for verilog code models with correct-by-construction non-textual representations and targeted code repair. InInternational Conference on Learning Representations, volume 2025, pp. 90377– 90422, 2025a. Liu, S., Fang, W., Lu, Y ., Wang, J., Zhang, Q., Zhang, H., a...

2025
[10]

Beyond Tokens: Enhancing RTL Quality Estimation via Structural Graph Learning

Liu, Y ., Xu, C., Zhou, Y ., Li, Z., and Xu, Q. Deeprtl: Bridg- ing verilog understanding and generation with a unified representation model. InThe Thirteenth International Conference on Learning Representations, 2025b. Liu, Y ., Zhang, H., Wang, Y ., Tsaras, D., Chen, L., Yuan, M., and Xu, Q. Beyond tokens: Enhancing rtl quality estimation via structural...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[11]

Mteb: Massive text embedding benchmark

Muennighoff, N., Tazi, N., Magne, L., and Reimers, N. Mteb: Massive text embedding benchmark. InProceed- ings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pp. 2014– 2037,

2014
[12]

Generative representational in- struction tuning

Muennighoff, N., Su, H., Wang, L., Yang, N., Wei, F., Yu, T., Singh, A., and Kiela, D. Generative representational in- struction tuning. InInternational Conference on Learning Representations, volume 2025, pp. 45544–45613,

2025
[13]

Text and Code Embeddings by Contrastive Pre-Training

Neelakantan, A., Xu, T., Puri, R., Radford, A., Han, J. M., Tworek, J., Yuan, Q., Tezak, N., Kim, J. W., Hallacy, C., et al. Text and code embeddings by contrastive pre- training.arXiv preprint arXiv:2201.10005,

work page internal anchor Pith review Pith/arXiv arXiv
[14]

Oord, A. v. d., Li, Y ., and Vinyals, O. Representation learn- ing with contrastive predictive coding.arXiv preprint arXiv:1807.03748,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

Code Llama: Open Foundation Models for Code

Roziere, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X. E., Adi, Y ., Liu, J., Sauvestre, R., Remez, T., et al. Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950,

work page internal anchor Pith review Pith/arXiv arXiv
[16]

Saleh, R. A. and Saleh, A. Statistical properties of the log-cosh loss function used in machine learning.arXiv preprint arXiv:2208.04564,

work page arXiv
[17]

We compare the full pretraining dataset and the subset of designs successfully converted into CDFGs using token-count distributions and UniRTL embedding distributions

12 UniRTL: Unifying Code and Graph for Robust RTL Representation Learning 0 250 500 750 1000 1250 1500 1750 2000 # Tokens (cl100k_base) 0.000 0.001 0.002 0.003 0.004Density (a) Token Distribution Graph-Pretrain Data Graph-Pretrain KDE All Data All Data KDE −20 −15 −10 −5 0 5 10 Principal Component 1 −4 −2 0 2 4 6 8 10 Principal Component 2 (b) Embedding D...

2000
[18]

This transformation does not affect the practical utility of the predictor, as we are more concerned with the relative quality of different designs

and StructRTL (Liu et al., 2025c) to apply a logarithm transformation to these values, making the target distribution more suitable for model learning. This transformation does not affect the practical utility of the predictor, as we are more concerned with the relative quality of different designs. For training, we adopt the log-cosh loss (Saleh & Saleh,...

2022
[19]

Determine whether the given pair of RTL code snippets is functionally equivalent

and NV-Embed-v2 (Lee et al., 2025), we prepend the instruction “Determine whether the given pair of RTL code snippets is functionally equivalent.” to their model-specific templates to adapt their embeddings to this task. For GraphCodeBERT (Guo et al.,

2025
[20]

We evaluate models using five metrics: Average Precision (AP), Accuracy, F1, Precision, and Recall, with AP serving as the main metric

with hard negatives for downstream fine-tuning on this task: Lfec =− 1 M MX i=1 log   exp cos(fθ(Ri),fθ(Ei)) τ PM j=1 exp cos(fθ(Ri),fθ(Ej ))) τ +PM j=1 Phj k=1 cos(fθ(Ri),fθ(Uj [k])) τ   (13) where M is the batch size, fθ is the embedding function, τ is the temperature hyperparameter, and hj = min(length(Uj),max hard negatives), is the number of hard...

2023

[1] [1]

Qwen Technical Report

Bai, J., Bai, S., Chu, Y ., Cui, Z., Dang, K., Deng, X., Fan, Y ., Ge, W., Han, Y ., Huang, F., et al. Qwen technical report.arXiv preprint arXiv:2309.16609,

work page internal anchor Pith review Pith/arXiv arXiv

[2] [2]

Speech- bert: An audio-and-text jointly learned language model for end-to-end spoken question answering

Chuang, Y .-S., Liu, C.-L., Lee, H.-y., and Lee, L.-s. Speech- bert: An audio-and-text jointly learned language model for end-to-end spoken question answering. InProc. In- terspeech 2020, pp. 4168–4172,

2020

[3] [3]

Bert: Pre-training of deep bidirectional transformers for lan- guage understanding

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for lan- guage understanding. InProceedings of the 2019 confer- ence of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186,

2019

[4] [4]

Codebert: A pre-trained model for programming and natural lan- guages

Feng, Z., Guo, D., Tang, D., Duan, N., Feng, X., Gong, M., Shou, L., Qin, B., Liu, T., Jiang, D., et al. Codebert: A pre-trained model for programming and natural lan- guages. InFindings of the Association for Computational Linguistics: EMNLP 2020, pp. 1536–1547,

2020

[5] [5]

The Llama 3 Herd of Models

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

work page internal anchor Pith review Pith/arXiv arXiv

[6] [6]

DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence

Guo, D., Zhu, Q., Yang, D., Xie, Z., Dong, K., Zhang, W., Chen, G., Bi, X., Wu, Y ., Li, Y ., et al. Deepseek-coder: When the large language model meets programming–the rise of code intelligence.arXiv preprint arXiv:2401.14196,

work page internal anchor Pith review Pith/arXiv arXiv

[7] [7]

CodeSearchNet Challenge: Evaluating the State of Semantic Code Search

Husain, H., Wu, H.-H., Gazit, T., Allamanis, M., and Brockschmidt, M. Codesearchnet challenge: Evaluat- ing the state of semantic code search.arXiv preprint arXiv:1909.09436,

work page internal anchor Pith review Pith/arXiv arXiv 1909

[8] [8]

Nv-embed: Improved techniques for training llms as generalist embedding models

Lee, C., Roy, R., Xu, M., Raiman, J., Shoeybi, M., Catan- zaro, B., and Ping, W. Nv-embed: Improved techniques for training llms as generalist embedding models. In International Conference on Learning Representations, volume 2025, pp. 79310–79333,

2025

[9] [9]

Craftrtl: High- quality synthetic data generation for verilog code models with correct-by-construction non-textual representations and targeted code repair

Liu, M., Tsai, Y .-D., Zhou, W., and Ren, H. Craftrtl: High- quality synthetic data generation for verilog code models with correct-by-construction non-textual representations and targeted code repair. InInternational Conference on Learning Representations, volume 2025, pp. 90377– 90422, 2025a. Liu, S., Fang, W., Lu, Y ., Wang, J., Zhang, Q., Zhang, H., a...

2025

[10] [10]

Beyond Tokens: Enhancing RTL Quality Estimation via Structural Graph Learning

Liu, Y ., Xu, C., Zhou, Y ., Li, Z., and Xu, Q. Deeprtl: Bridg- ing verilog understanding and generation with a unified representation model. InThe Thirteenth International Conference on Learning Representations, 2025b. Liu, Y ., Zhang, H., Wang, Y ., Tsaras, D., Chen, L., Yuan, M., and Xu, Q. Beyond tokens: Enhancing rtl quality estimation via structural...

work page internal anchor Pith review Pith/arXiv arXiv 2025

[11] [11]

Mteb: Massive text embedding benchmark

Muennighoff, N., Tazi, N., Magne, L., and Reimers, N. Mteb: Massive text embedding benchmark. InProceed- ings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pp. 2014– 2037,

2014

[12] [12]

Generative representational in- struction tuning

Muennighoff, N., Su, H., Wang, L., Yang, N., Wei, F., Yu, T., Singh, A., and Kiela, D. Generative representational in- struction tuning. InInternational Conference on Learning Representations, volume 2025, pp. 45544–45613,

2025

[13] [13]

Text and Code Embeddings by Contrastive Pre-Training

Neelakantan, A., Xu, T., Puri, R., Radford, A., Han, J. M., Tworek, J., Yuan, Q., Tezak, N., Kim, J. W., Hallacy, C., et al. Text and code embeddings by contrastive pre- training.arXiv preprint arXiv:2201.10005,

work page internal anchor Pith review Pith/arXiv arXiv

[14] [14]

Oord, A. v. d., Li, Y ., and Vinyals, O. Representation learn- ing with contrastive predictive coding.arXiv preprint arXiv:1807.03748,

work page internal anchor Pith review Pith/arXiv arXiv

[15] [15]

Code Llama: Open Foundation Models for Code

Roziere, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X. E., Adi, Y ., Liu, J., Sauvestre, R., Remez, T., et al. Code llama: Open foundation models for code.arXiv preprint arXiv:2308.12950,

work page internal anchor Pith review Pith/arXiv arXiv

[16] [16]

Saleh, R. A. and Saleh, A. Statistical properties of the log-cosh loss function used in machine learning.arXiv preprint arXiv:2208.04564,

work page arXiv

[17] [17]

We compare the full pretraining dataset and the subset of designs successfully converted into CDFGs using token-count distributions and UniRTL embedding distributions

12 UniRTL: Unifying Code and Graph for Robust RTL Representation Learning 0 250 500 750 1000 1250 1500 1750 2000 # Tokens (cl100k_base) 0.000 0.001 0.002 0.003 0.004Density (a) Token Distribution Graph-Pretrain Data Graph-Pretrain KDE All Data All Data KDE −20 −15 −10 −5 0 5 10 Principal Component 1 −4 −2 0 2 4 6 8 10 Principal Component 2 (b) Embedding D...

2000

[18] [18]

This transformation does not affect the practical utility of the predictor, as we are more concerned with the relative quality of different designs

and StructRTL (Liu et al., 2025c) to apply a logarithm transformation to these values, making the target distribution more suitable for model learning. This transformation does not affect the practical utility of the predictor, as we are more concerned with the relative quality of different designs. For training, we adopt the log-cosh loss (Saleh & Saleh,...

2022

[19] [19]

Determine whether the given pair of RTL code snippets is functionally equivalent

and NV-Embed-v2 (Lee et al., 2025), we prepend the instruction “Determine whether the given pair of RTL code snippets is functionally equivalent.” to their model-specific templates to adapt their embeddings to this task. For GraphCodeBERT (Guo et al.,

2025

[20] [20]

We evaluate models using five metrics: Average Precision (AP), Accuracy, F1, Precision, and Recall, with AP serving as the main metric

with hard negatives for downstream fine-tuning on this task: Lfec =− 1 M MX i=1 log   exp cos(fθ(Ri),fθ(Ei)) τ PM j=1 exp cos(fθ(Ri),fθ(Ej ))) τ +PM j=1 Phj k=1 cos(fθ(Ri),fθ(Uj [k])) τ   (13) where M is the batch size, fθ is the embedding function, τ is the temperature hyperparameter, and hj = min(length(Uj),max hard negatives), is the number of hard...

2023