Pith. sign in

REVIEW 4 major objections 5 minor 41 references

TD-Interpreter: Enhancing the Understanding of Timing Diagrams with Visual-Language Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read TD-Interpreter claims that fine-tuning a 7B multimodal LLM on synthetic timing-diagram QA data yields near-perfect interpretation of WaveDrom-style waveform pictures, outperforming untuned GPT-4o on the authors' benchmarks.

desk verdict Solid synthetic data pipeline for timing-diagram VQA, but the headline GPT-4o comparison is unquantified and the benchmark is self-generated, so the transfer claim is untested. read the letter →

arxiv 2507.16844 v1 pith:AAH334H7 submitted 2025-07-20 cs.LG

classification cs.LG
keywords timingdiagramsmultimodallargelanguagemodelsvisualquestionansweringLLaVAfine-tuningsyntheticdatagenerationdigitaldesignverificationclock-domaincrossingWaveDrom
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a lightweight, open multimodal language model can interpret timing diagrams, the waveform drawings that specify signal behavior in digital hardware, well enough to answer real design and verification questions about them. The authors argue that off-the-shelf models such as GPT-4o fail at this task because they produce generic textbook explanations and misread concrete waveforms, whereas a 7B LLaVA model fine-tuned on synthetic timing-diagram question-answer data can deliver case-specific analysis of protocol transactions, finite-state machines, and clock-domain crossings. If the claim holds, engineers could query third-party timing diagrams conversationally without sending proprietary IP to a commercial API. The reported evidence is an evaluation set generated by the same synthetic pipeline used for training, on which the fine-tuned model scores above 95 on BLEU and Rouge metrics.

What carries the argument

The load-bearing mechanism is a synthetic data-generation pipeline that converts Verilog modules into timing-diagram pictures and paired question-answer text. Concrete TDs come from parsing Verilog, generating a random testbench with EDAUtils' Vlogtbgen, simulating with iverilog, converting the VCD output to JSON with vcd2json, and rendering the JSON as a WaveDrom picture via wavedrom-cli; abstract TDs from datasheets and protocol manuals are supplied directly as JSON. Caption-based QA pairs are produced automatically from simulation values, and reasoning-based QA pairs are hand-designed around HDLBits problems and protocol specifications, with question types spanning specification, I/O, FSM, counter, static timing analysis, complex timing analysis, and clock-domain crossing. The model is LLaVA fine-tuned with LoRA (rank 8, 20M trainable parameters) for 100 epochs on roughly 10k sampled pictures, combining captioning and reasoning tasks during training.

What would settle it

Run TD-Interpreter on a held-out collection of timing diagrams that did not come from the paper's generator, for instance figures scanned from chip datasheets, asynchronous FIFO waveforms from a commercial simulator, and SPI traces from a logic analyzer, with questions and reference answers written independently by expert engineers; if Bleu/Rouge scores collapse or expert judges rate the answers generic or wrong, the central claim of general timing-diagram interpretation fails.

Watch

Extended reading notes

Core claim

The central claim is that visual-language learning on a sufficiently rich synthetic dataset turns a lightweight open multimodal LLM into a timing-diagram interpreter that outperforms untuned GPT-4o by a large margin. On the authors' benchmark, fine-tuned LLaVA reaches Bleu-4 = 95.9 and Rouge-1/2/l = 96.7/95.9/96.5, while untuned LLaVA reaches only 13.9/23.2/6.6/16.3; qualitative comparisons show GPT-4o giving generic or wrong answers on AHB wait states, asynchronous FIFO clock-domain-crossing relations, and SPI transaction details. The paper concludes from this that a 7B model can handle complex timing relations across multiple signals and clock cycles, synthesize FSMs from waveforms, and identify clock-domain-crossing issues when trained on waveform-aligned QA data.

Load-bearing premise

The load-bearing premise is that evaluation on diagrams produced by the same synthetic Verilog-to-WaveDrom pipeline and the same human-designed QA templates as the training data tells us how the model will behave on real third-party timing diagrams encountered in engineering practice.

Editorial extensions

If this is right

  • A 7B open multimodal model is sufficient for near-perfect answers on WaveDrom-style timing-diagram QA, so the approach is compatible with local deployment and IP-sensitive use.
  • The fine-tuned model reproduces ground-truth answers almost exactly on the evaluation set; the remaining deviations are characterized as small errors by the authors.
  • Engineers can ask about concrete waveform events, such as wait states, FSM transitions, CDC crossings, and SPI bit transfers, and receive answers tied to the pictured signals rather than textbook generalities.
  • The synthetic generation workflow can be reused to create training data for new modules and protocols by supplying Verilog sources or JSON descriptions and hand-authored reasoning questions.
  • Reported scores above 95 imply that, on in-distribution generated timing diagrams, the model's answers overlap the labeled answers almost completely.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The headline scores are in-distribution: the evaluation images and QA pairs are produced by the same Verilog-to-WaveDrom pipeline and labeling logic as the training data, so the numbers do not by themselves certify performance on arbitrary third-party timing diagrams.
  • The transfer question could be tested by collecting diagrams from real datasheets, commercial simulators, and logic analyzers, having expert engineers author and answer questions, and comparing TD-Interpreter's answers against both the experts and the reported 95-plus scores.
  • The same caption-plus-reasoning training recipe could plausibly transfer to other technical diagram families, such as statecharts, protocol sequence charts, or non-WaveDrom waveform styles, though the paper only demonstrates the WaveDrom format.
  • A human baseline would sharpen the comparison: the paper reports qualitative human feedback on four selected diagrams but does not measure expert accuracy or confidence on the full benchmark, so it is unknown whether the model exceeds typical engineer performance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces TD-Interpreter, a visual question-answering tool for timing diagrams, implemented by fine-tuning the 7B multimodal model LLaVA with LoRA. The authors report a human-centered study of 31 digital-design engineers, a synthetic data-generation pipeline that converts Verilog modules into WaveDrom timing-diagram images paired with caption-based and reasoning-based QA pairs, and an experimental evaluation in which fine-tuned LLaVA achieves BLEU-4=95.9 and Rouge-1/2/l=96.7/95.9/96.5, compared with untuned LLaVA scores of BLEU-4=13.9 and Rouge-1/2/l=23.2/6.6/16.3. Four qualitative examples are used to argue that TD-Interpreter outperforms GPT-4o, which the abstract summarizes as a large-margin advantage on the evaluated benchmarks.

Significance. The problem is relevant and the general approach is practical: a locally deployable, IP-protecting 7B MLLM for timing-diagram interpretation could benefit hardware design and verification teams. The paper has several strengths: it grounds the task design in an explicit human-needs survey, constructs a substantial synthetic data pipeline, uses an open-source base model, and candidly acknowledges the WaveDrom-only rendering style. If the performance claims were supported by a valid external evaluation and a quantitative baseline comparison, the contribution would be useful. At present, however, the headline claim is not established: the quantitative evaluation is in-distribution with respect to the data generator, and GPT-4o is only shown in anecdotal examples.

major comments (4)
  1. [Abstract and Section VI-B] The abstract and conclusion claim that TD-Interpreter 'outperformed untuned GPT-4o by a large margin on the evaluated benchmarks,' but Section VI-B reports BLEU/Rouge scores only for untuned LLaVA versus fine-tuned LLaVA. GPT-4o appears only in four hand-picked qualitative examples, which are not a quantitative benchmark. Please provide GPT-4o (and preferably other strong MLLM baselines) evaluated on the same test set with the same metrics, including per-question-type results and statistical significance or confidence intervals.
  2. [Sections IV and VI-A] The evaluation benchmark is generated by the same pipeline used for training: Verilog modules, iverilog simulation, vcd2json conversion, wavedrom-cli rendering, and the author-designed QA logic. The test images are image-disjoint from training, so the results are not exact-image memorization, but the rendering style, question templates, answer phrasing, and signal-configuration distribution are all drawn from the same generator. The 95+ BLEU/Rouge scores therefore primarily measure how well the model reproduces the generator's template-level regularities, not how well it interprets real third-party timing diagrams. The paper needs an external test set—for example, manually annotated datasheet figures, real simulation traces, or WaveDrom images from a different source—to support the transfer claim.
  3. [Section VI-C] The limitation section acknowledges that only WaveDrom-style timing diagrams were considered, yet the motivating scenarios in Section I emphasize datasheet figures, third-party FIFO traces, and SPI waveforms, which may differ in visual style, notation, and query distribution. Since the main practical claim is usefulness on such real-world diagrams, the acknowledgment alone is insufficient. Please add a transfer experiment to non-WaveDrom renderings or to real-world timing diagrams, and report the resulting performance degradation, even if the external set is small.
  4. [Section VI-B, human feedback] The human 'feedback' reported for the four examples is anecdotal: it involves small numbers of participants, no fixed protocol, no inter-rater reliability measure, and no comparison against TD-Interpreter's accuracy on those items. If human assessment is intended to support the usefulness claim, it should be reported as a systematic study with quantitative measures (e.g., correctness rates, confidence ratings, and agreement statistics).
minor comments (5)
  1. [Section III-B] The sentence 'interestingly we found a diverse of of use cases' contains a duplicated 'of' and should read 'a diverse set of use cases.'
  2. [Section IV-B.1] The phrase 'when the first data stream of date finishes transmission' should read 'when the first data stream of data finishes transmission.'
  3. [Section VI-B.3] The heading 'Human feedbcak' contains a typo and should be 'Human feedback.'
  4. [Abstract and Section VI-B] The phrase 'untuned GPT-4o' is imprecise because GPT-4o is a closed model that is not fine-tuned in this work; 'out-of-the-box GPT-4o' or 'GPT-4o without fine-tuning' would be clearer.
  5. [References] Reference [25] appears to duplicate Reference [5]; if they are the same work, the duplicate citation should be removed or replaced with a distinct source.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the held-out evaluation is image-disjoint from training, and the reported improvements are empirical measurements rather than claims that reduce to their inputs by construction.

full rationale

I examined the paper's claimed derivation chain. The central empirical claim is that fine-tuned LLaVA achieves high BLEU/Rouge scores on an evaluation benchmark (Section VI-B) and qualitatively outperforms GPT-4o on selected examples. The evaluation benchmark is generated by the same Verilog-to-WaveDrom pipeline and QA-generation logic used to construct the training set (Sections IV and VI-A), which is an in-distribution evaluation limitation. However, the paper explicitly states that the evaluation TD pictures were never seen during training and differ due to random generation. There is no fitted parameter that is renamed as a prediction, no equation that defines an output in terms of the target claim, and no load-bearing self-citation or imported uniqueness argument. The lack of quantitative GPT-4o scores on the same benchmark and the untested transfer to real third-party timing diagrams are external-validity and evidence-quality concerns, not circularity. The related-work citation to TD-Magic [12] is self-citation by overlapping authors, but it is used only to differentiate scope and is not load-bearing for the present results. Accordingly, the paper's own derivation chain is self-contained: training data, held-out evaluation data, and the measured metrics form a standard supervised-learning loop without the target result being assumed in the inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. The key assumptions are about data representativeness, label validity, metric adequacy, and transfer from synthetic to real timing diagrams. The free parameters are training choices that are hand-set and not derived from first principles.

free parameters (4)
  • training_data_scale = ~10k samples (4942 caption-based + 5292 reasoning-based)
    Chosen empirically as 'desirable' for evaluation results (Section VI-A); no scaling law or ablation justifies this number.
  • LoRA_rank = 8
    Set to 8 in Section V-B without sensitivity analysis or justification for the choice.
  • learning_rate = 1e-4
    Reported in Section VI-A; no grid search, schedule, or ablation is provided.
  • training_epochs = 100
    Training ran for 17 hours over 100 epochs (Section VI-A); no early stopping or convergence criterion is reported.
assumptions (4)
  • domain assumption Synthetic wavedrom TDs with programmatic QA labels are representative of real-world timing diagrams from datasheets and third-party IP.
    The whole dataset and evaluation are built this way (Sections IV and VI-A); no real out-of-distribution test set is used quantitatively.
  • domain assumption DeepSeek-Coder-V2's captions, generated from module names and port lists, are valid ground truth for the visual content of a TD.
    Section IV-A-1 generates captions from module names and I/O ports rather than from the image; the model is trained to align images to these captions.
  • domain assumption BLEU and Rouge text-overlap scores are valid measures of VQA correctness for timing-diagram reasoning.
    Section VI-B uses them as the only quantitative metrics; semantically correct paraphrases could be penalized and textually similar wrong answers rewarded.
  • domain assumption LLaVA's pretrained visual-language alignment transfers to timing-diagram images after LoRA fine-tuning.
    The vision encoder is frozen and the adapter is trained on synthetic TDs only (Section V); transfer to real TD styles is assumed but not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TD-Interpreter: Enhancing the Understanding of Timing Diagrams with Visual-Language Learning." pith.science (2026). https://pith.science/paper/AAH334H7

@misc{pith2026250716844,
  author       = {Pith},
  title        = {Pith review of: TD-Interpreter: Enhancing the Understanding of Timing Diagrams with Visual-Language Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AAH334H7}},
  note         = {Machine review of arXiv:2507.16844}
}
read the original abstract

We introduce TD-Interpreter, a specialized ML tool that assists engineers in understanding complex timing diagrams (TDs), originating from a third party, during their design and verification process. TD-Interpreter is a visual question-answer environment which allows engineers to input a set of TDs and ask design and verification queries regarding these TDs. We implemented TD-Interpreter with multimodal learning by fine-tuning LLaVA, a lightweight 7B Multimodal Large Language Model (MLLM). To address limited training data availability, we developed a synthetic data generation workflow that aligns visual information with its textual interpretation. Our experimental evaluation demonstrates the usefulness of TD-Interpreter which outperformed untuned GPT-4o by a large margin on the evaluated benchmarks.

Figures

Figures reproduced from arXiv: 2507.16844 by the authors.

Figure 1
Figure 1. Conversations with TD-Interpreter to IP-privacy concerns, fine-tuning limitations, and insufficient training data. To address these problems, we explored cost-effective solutions for data construction, training, and local deployment, specially tailored for companies with limited resources and IP concerns. In particular: • We conduct a human-centered empirical study, by interviewing a group 31 experts in embedded sys… view at source ↗
Figure 2
Figure 2. Workflow of Generating TD Pictures Fine-tuning an MLLM for interpreting TDs, requires a dataset of annotated TDs. However, no such a dataset currently exists. Thus we developed a strategy for generating a large dataset of TDs paired with question-answer (QA) pairs for visual instruction tuning [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. TD of a 2-to-1 negative multiplexer An example of our training strategy, is the outcome obtained for the TD shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Desired TDs for the Serial Receiver Protocol [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 4
Figure 4. Figure 4: Question Types Distribution 2) Statistics: Through the outlined caption-based data-generation method, a total of 221,983 TDs were generated. For these TDs, we were able to generate 9,915,694 QA pairs, of which 4,306,008 are of the form "What is the value of signal s at…
Figure 7
Figure 7. Figure 7: Question Distribution 4) Statistics: Table I showcases all the design tasks we considered. Overall, we designed 310 different types of questions [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Model Architecture of TD-Interpreter number of RGB channels in the TD picture. The vision encoder g(·) first extracts the visual features from the input image as hv = g(Xv). Subsequently, the language model predicts the next token of the target textual response Xa from…
Figure 9
Figure 9. Figure 9: AHB Burst Write Transfer • Q1 (CTA): Why takes writing to address 0x3C longer than other transfers? • A1 (TD-Interpreter): «The write to address 0x3C does not take longer. Instead it’s the write to address 0x38 which takes longer. One of the subordinates adds a wait st…
Figure 10
Figure 10. Figure 10: Fancy Timer 2) Fancy Timer: The TD in [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Complete SPI protocol (slave has no internal clock) [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 24 canonical work pages

  1. [1]

    Adc129s022,

    “Adc129s022,” https://www.ti.com/lit/ds/symlink/adc128s022.pdf

  2. [2]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” 2023. [Online]. Available: https://arxiv.org/abs/2304.08485

  3. [3]

    Verilogeval: Evaluating large language models for verilog code generation,

    M. Liu, N. Pinckney, B. Khailany, and H. Ren, “Verilogeval: Evaluating large language models for verilog code generation,” inProc. of ICCAD 2023: IEEE/ACM International Conference on Computer Aided Design. IEEE, 2023, pp. 1–8

  4. [4]

    Chipnemo: Domain- adapted llms for chip design,

    M. Liu, T.-D. Ene, R. Kirby, C. Cheng, N. Pinckney, R. Liang, J. Alben, H. Anand, S. Banerjee, I. Bayraktarogluet al., “Chipnemo: Domain- adapted llms for chip design,”arXiv preprint arXiv:2311.00176, 2023

  5. [5]

    Benchmarking large language models for automated verilog rtl code generation,

    S. Thakur, B. Ahmad, Z. Fan, H. Pearce, B. Tan, R. Karri, B. Dolan- Gavitt, and S. Garg, “Benchmarking large language models for automated verilog rtl code generation,” inProc. of DATE 2023: Design, Automation & Test in Europe Conference & Exhibition. IEEE, 2023, pp. 1–6

  6. [6]

    Verigen: A large language model for verilog code generation,

    S. Thakur, B. Ahmad, H. Pearce, B. Tan, B. Dolan-Gavitt, R. Karri, and S. Garg, “Verigen: A large language model for verilog code generation,” ACM Transactions on Design Automation of Electronic Systems, vol. 29, no. 3, pp. 1–31, 2024

  7. [7]

    Llm4hwdesign,

    “Llm4hwdesign,” https://nvlabs.github.io/LLM4HWDesign/

  8. [8]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

Show all 41 references
  1. [9]

    Codegen: An open large language model for code with multi-turn program synthesis,

    E. Nijkamp, B. Pang, H. Hayashi, L. Tu, H. Wang, Y . Zhou, S. Savarese, and C. Xiong, “Codegen: An open large language model for code with multi-turn program synthesis,”arXiv preprint arXiv:2203.13474, 2022

  2. [10]

    Gpt4aigchip: Towards next-generation ai accelerator design automation via large language models,

    Y . Fu, Y . Zhang, Z. Yu, S. Li, Z. Ye, C. Li, C. Wan, and Y . C. Lin, “Gpt4aigchip: Towards next-generation ai accelerator design automation via large language models,” inProc. of ICCAD 2023: IEEE/ACM International Conference on Computer Aided Design. IEEE, 2023, pp. 1–9

  3. [11]

    Natural language is not enough: Benchmarking multi-modal generative ai for verilog generation,

    K. Chang, Z. Chen, Y . Zhou, W. Zhu, H. Xu, C. Li, M. Wang, S. Liang, H. Li, Y . Hanet al., “Natural language is not enough: Benchmarking multi-modal generative ai for verilog generation,”arXiv preprint arXiv:2407.08473, 2024

  4. [12]

    Td-magic: From pictures of timing diagrams to formal specifications,

    J. He, D. Ni ˇckovi´c, E. Bartocci, and R. Grosu, “Td-magic: From pictures of timing diagrams to formal specifications,” inProc. of DAC 2023: 60th ACM/IEEE Design Automation Conference. IEEE, 2023, pp. 1–6

  5. [13]

    Boston, MA: Springer US, 2002, ch

    Advanced ASIC Chip Synthesis Using Synopsys Design Compiler™ Physical Compiler™ and PrimeTime. Boston, MA: Springer US, 2002, ch. Static Timing Analysis, pp. 261–303. [Online]. Available: https://doi.org/10.1007/0-306-47507-3_13

  6. [14]

    Monitoring temporal properties of continuous signals,

    O. Maler and D. Nickovic, “Monitoring temporal properties of continuous signals,” inInternational symposium on formal techniques in real-time and fault-tolerant systems. Springer, 2004, pp. 152–166

  7. [15]

    Timing diagrams: Formalization and algorithmic verification,

    F. Kathi, “Timing diagrams: Formalization and algorithmic verification,” J. Log. Lang. Inf., vol. 8, no. 3, pp. 323–361, 1999

  8. [16]

    Efficient decompositional model checking for regular timing diagrams,

    N. Amla, E. A. Emerson, and K. S. Namjoshi, “Efficient decompositional model checking for regular timing diagrams,” inProc. of CHARME, ser. LNCS, vol. 1703. Springer, 1999, pp. 67–81

  9. [17]

    Survey on mining signal temporal logic specifications,

    E. Bartocci, C. Mateis, E. Nesterini, and D. Nickovic, “Survey on mining signal temporal logic specifications,”Information and Computation, vol. 289, p. 104957, 2022

  10. [18]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Luet al., “Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 202...

  11. [19]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” in9th International Conference on Learn...

  12. [20]

    Gemini: a family of highly capable multimodal models,

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millicanet al., “Gemini: a family of highly capable multimodal models,”arXiv preprint arXiv:2312.11805, 2023

  13. [21]

    Verilog testbench generator

    EDAUtils, “Verilog testbench generator.” [Online]. Available: https: //www.edautils.com/VlogTBGen.html

  14. [22]

    Icarus verilog compilation system

    Stephen Williams, “Icarus verilog compilation system.” [Online]. Available: https://github.com/steveicarus/iverilog

  15. [23]

    vcd2json

    Github user: Namake Nanamaru, “vcd2json.” [Online]. Available: https://github.com/nanamake/vcd2json

  16. [24]

    wavedrom-cli

    WaveDrom, “wavedrom-cli.” [Online]. Available: https://github.com/ wavedrom/cli

  17. [25]

    Benchmarking large language models for automated verilog rtl code generation,

    S. Thakur, B. Ahmad, Z. Fan, H. Pearce, B. Tan, R. Karri, B. Dolan-Gavitt, and S. Garg, “Benchmarking large language models for automated verilog rtl code generation,” 2022. [Online]. Available: https://arxiv.org/abs/2212.11140

  18. [26]

    Deepseek-coder-v2: Breaking the barrier of closed-source models in code intelligence,

    DeepSeek-AI, Q. Zhu, D. Guo, Z. Shao, D. Yang, P. Wang, R. Xu, Y . Wu, Y . Li, H. Gao, S. Ma, W. Zeng, X. Bi, Z. Gu, H. Xu, D. Dai, K. Dong, L. Zhang, Y . Piao, Z. Gou, Z. Xie, Z. Hao, B. Wang, J. Song, D. Chen, X. Xie, K. Guan, Y . You, A. Liu, Q. Du, W. Gao, X. Lu, Q. Chen, ...

  19. [27]

    Generated knowledge prompting for commonsense reasoning,

    J. Liu, A. Liu, X. Lu, S. Welleck, P. West, R. L. Bras, Y . Choi, and H. Hajishirzi, “Generated knowledge prompting for commonsense reasoning,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, ...

  20. [28]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” inProceedings of the 36th International Conference on Neural Information Processing Systems, ser. NIPS ’22. ...

  21. [29]

    Hdlbits,

    “Hdlbits,” https://hdlbits.01xz.net/wiki/Problem_sets

  22. [30]

    serialdp,

    “serialdp,” https://hdlbits.01xz.net/wiki/Fsm_serialdp. [31]AMBA AXI and ACE Protocol Specification, Arm, October 2011. [32]AMBA AHB Protocol Specification, Arm, September 2021. [33]AMBA APB Protocol Specification, Arm, April 2021

  23. [34]

    serialdata,

    “serialdata,” https://hdlbits.01xz.net/wiki/Fsm_serialdata

  24. [35]

    motor timer,

    “motor timer,” https://hdlbits.01xz.net/wiki/Exams/2013_q2bfsm

  25. [36]

    complete fsm,

    “complete fsm,” https://hdlbits.01xz.net/wiki/Exams/review2015_fsm

  26. [37]

    fancy timer,

    “fancy timer,” https://hdlbits.01xz.net/wiki/Exams/review2015_ fancytimer

  27. [38]

    “hdlc,” https://hdlbits.01xz.net/wiki/Fsm_hdlc

  28. [39]

    wcounter,

    “wcounter,” https://hdlbits.01xz.net/wiki/Exams/2014_q3fsm

  29. [40]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

  30. [41]

    Layer normalization,

    J. L. Ba, “Layer normalization,”arXiv preprint arXiv:1607.06450, 2016

  31. [42]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,

    W.-L. Chiang, Z. Li, Z. Lin, Y . Sheng, Z. Wu, H. Zhang, L. Zheng, S. Zhuang, Y . Zhuang, J. E. Gonzalezet al., “Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality,”See https://vicuna. lmsys. org (accessed 14 April 2023), vol. 2, no. 3, p. 6, 2023

  32. [43]

    Root mean square layer normalization,

    B. Zhang and R. Sennrich, “Root mean square layer normalization,” Advances in Neural Information Processing Systems, vol. 32, 2019

  33. [44]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,”arXiv preprint arXiv:2106.09685, 2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.