pith. sign in

arxiv: 2607.01590 · v1 · pith:VKG3VNLInew · submitted 2026-07-02 · 💻 cs.AI · cs.SE

Hawk: Harnessing Hardware-Aware Knowledge for High-Performance NPU Kernel Generation

Junyi Wen , Ruiyan Zhuang , Yongjia Xu , Pengtu Li , Rui Zou , Hongyi Chen , Chingman Wan , Puxu Yang
show 2 more authors
Wuhui Chen Yanlin Wang
This is my paper

Pith reviewed 2026-07-03 14:58 UTC · model grok-4.3

classification 💻 cs.AI cs.SE
keywords NPU kernel generationhardware-aware knowledgeLLM code generationknowledge distillationneural processing unitskernel optimizationtraining-free frameworkexecution feedback
0
0 comments X

The pith

Hawk raises NPU kernel generation accuracy from 49.4 percent to 80 percent by distilling hardware constraints from execution results.

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

The paper introduces Hawk as a training-free framework that supplies large language models with hardware-specific knowledge to generate kernels for Neural Processing Units. Standard LLMs fail because they lack priors about NPU memory hierarchies and constraints, so copied code often crashes or slows down at runtime. Hawk counters this with three modules that build executable knowledge representations, retrieve hardware-aligned examples in two dimensions, and refine outputs by pruning errors using actual execution feedback. If correct, this removes the need for manual expert tuning when deploying models on specialized accelerators. The result matters because NPU kernel development remains a major industrial bottleneck that currently demands extensive human effort.

Core claim

Hawk is a training-free framework that harnesses hardware-aware knowledge through three core modules: the Run-Time Knowledge Synthesis Module, which couples error context with executable semantics via Triple-Part Executable Knowledge Representation; the Bottleneck-Aware Knowledge Retrieval Module, which projects queries into orthogonal syntactic and hardware-aligned semantic spaces using a 2D-Retrieval paradigm; and the Effect-Driven Knowledge Distillation Module, which performs LLM-driven semantic arbitration to prune errors and consolidate redundancies based on empirical execution feedback. On real-world NPU workloads this raises generation accuracy from 49.4 percent to 80.0 percent and de

What carries the argument

Hawk's three-module pipeline: Run-Time Knowledge Synthesis with Triple-Part Executable Knowledge Representation, Bottleneck-Aware Knowledge Retrieval via 2D-Retrieval, and Effect-Driven Knowledge Distillation driven by execution feedback.

If this is right

  • Kernel developers obtain correct and performant NPU code without hand-written hardware priors.
  • Generated kernels run up to 2.2 times faster on actual NPU hardware than prior automated methods.
  • The system continuously improves its knowledge base from runtime observations rather than static training data.
  • Accuracy gains hold across multiple real-world NPU workloads without additional model fine-tuning.

Where Pith is reading between the lines

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

  • Similar feedback-driven distillation could be applied to kernel generation for other accelerators such as GPUs or custom ASICs.
  • The approach offers a path to make LLMs more reliable for any low-level programming task where runtime constraints are hard to encode statically.
  • Integrating the three modules into existing code-generation pipelines might reduce the expert time needed to port models to new hardware targets.

Load-bearing premise

Execution feedback can guide the LLM to prune errors and consolidate knowledge without the model introducing fresh inaccuracies during semantic arbitration.

What would settle it

On a held-out set of NPU workloads the generated kernels achieve accuracy below 60 percent or no measurable speedup relative to the baselines used in the paper.

Figures

Figures reproduced from arXiv: 2607.01590 by Chingman Wan, Hongyi Chen, Junyi Wen, Pengtu Li, Puxu Yang, Ruiyan Zhuang, Rui Zou, Wuhui Chen, Yanlin Wang, Yongjia Xu.

Figure 1
Figure 1. Figure 1: Ascend NPU architecture. turing deep memory and parallelization constraints. By fusing these dimensions, the module precisely isolates actionable, root-cause optimization strategies rather than merely fetching functionally analogous code snippets. Effect-Driven Knowledge Distillation (Addressing C3). To maintain a high-purity corpus, this module em￾ploys an LLM-driven semantic arbitration. Rather than rely… view at source ↗
Figure 2
Figure 2. Figure 2: Top-10 API reference fre￾quency across 2,690 kernel files. activation conv foreach hash index loss matmul norm optim pooling quant rnn vfusion activation conv foreach hash index loss matmul norm optim pooling quant rnn vfusion 0.0 0.2 0.4 0.6 0.8 1.0 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of prompting strategy on Ascend C generation. DS [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of Hawk. mated distillation process rigorously grounded in actual compilation and execution outcomes. IV. HAWK OVERVIEW To tackle the intricacies of NPU kernel generation, we introduce Hawk, a training-free framework built upon a foundational code agent that seamlessly har￾nesses hardware-aware knowledge into existing devel￾opment pipelines. Specifically, we implement Hawk as an advanced extension… view at source ↗
Figure 6
Figure 6. Figure 6: The overall performance of L1 kernels with GLM-5.1. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The overview of the Score trajectory throughout the L1 kernel generation iterations with GLM-5.1. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Normalized capability. 0.0 0.2 0.4 0.6 0.8 1.0 Acc 0.0 0.5 1.0 1.5 2.0 2.5 Speedup Better Vanilla LLM Few-shot CANNBot Hawk [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: The overall performance of L2 kernels with GLM-5.1. [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: The L1 kernel performance with DeepSeek-V4-Flash. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
read the original abstract

Developing high-performance kernels for Neural Processing Units (NPUs) is a critical industry bottleneck, requiring developers to manually navigate implicit hardware constraints and strict memory hierarchies. While large language models offer immense automation potential, they fail catastrophically on NPUs due to a fundamental lack of hardware-specific priors. Naively transplanting code snippets from similar NPU kernels may pass the compiler, but it consistently triggers runtime crashes and performance degradation by blindly violating underlying hardware constraints. To overcome this, we introduce Hawk, a training-free framework that harnesses hardware-aware knowledge through three core modules: (1) Run-Time Knowledge Synthesis Module, which employs a Triple-Part Executable Knowledge Representation to inherently couple the error context with executable semantics; (2) Bottleneck-Aware Knowledge Retrieval Module, which implements a 2D-Retrieval paradigm to project queries into orthogonal syntactic and hardware-aligned semantic spaces; and (3) Effect-Driven Knowledge Distillation Module, which leverages LLM-driven semantic arbitration to continuously distill the knowledge by pruning errors and consolidating redundancies based on the empirical execution feedback. Extensive evaluations on real-world NPU workloads demonstrate that Hawk elevates generation accuracy from 49.4% to 80.0%, while achieving up to a 2.2x execution speedup over state-of-the-art baselines.

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.

Referee Report

2 major / 1 minor

Summary. The paper introduces Hawk, a training-free framework for LLM-based generation of high-performance NPU kernels. It comprises three modules: (1) Run-Time Knowledge Synthesis using Triple-Part Executable Knowledge Representation to couple error context with executable semantics; (2) Bottleneck-Aware Knowledge Retrieval via 2D-Retrieval projecting queries into syntactic and hardware-aligned semantic spaces; and (3) Effect-Driven Knowledge Distillation employing LLM semantic arbitration on execution feedback to prune errors and consolidate knowledge. The central empirical claim is that Hawk raises kernel generation accuracy from 49.4% to 80.0% and delivers up to 2.2× execution speedup over SOTA baselines on real-world NPU workloads.

Significance. If the reported accuracy and speedup gains are reproducible and the arbitration step functions without systematic bias, the work would represent a meaningful advance in training-free, hardware-aware code generation for specialized accelerators. The explicit coupling of execution feedback with knowledge distillation and the 2D retrieval design are technically distinctive. However, the absence of detailed experimental protocols, baseline characterizations, and validation of the LLM arbitration mechanism limits the immediate assessability of these contributions.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (Effect-Driven Knowledge Distillation Module): The headline accuracy improvement (49.4% → 80.0%) and 2.2× speedup are attributed to LLM-driven semantic arbitration that prunes hardware violations using execution feedback. No mechanism is described to detect or correct LLM misinterpretations of NPU-specific constraints (memory hierarchy, implicit scheduling) when feedback is noisy; because the framework is training-free, any such errors propagate directly to the final metrics. This is load-bearing for the central claim.
  2. [Abstract and experimental evaluation section] Abstract and experimental evaluation section: The manuscript states concrete numerical gains (49.4% → 80.0% accuracy, 2.2× speedup) but supplies no baseline descriptions, workload/dataset details, number of trials, or error bars. Without these, it is impossible to determine whether the data support the attribution to the three modules rather than implementation artifacts or cherry-picked cases.
minor comments (1)
  1. [§2 and §3] Notation for the Triple-Part Executable Knowledge Representation and 2D-Retrieval paradigm is introduced without an accompanying diagram or pseudocode example, making the data flow between modules difficult to follow on first reading.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and commit to revisions that enhance the clarity and completeness of the work without altering its core contributions.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (Effect-Driven Knowledge Distillation Module): The headline accuracy improvement (49.4% → 80.0%) and 2.2× speedup are attributed to LLM-driven semantic arbitration that prunes hardware violations using execution feedback. No mechanism is described to detect or correct LLM misinterpretations of NPU-specific constraints (memory hierarchy, implicit scheduling) when feedback is noisy; because the framework is training-free, any such errors propagate directly to the final metrics. This is load-bearing for the central claim.

    Authors: We acknowledge that §3 does not explicitly describe a dedicated mechanism for detecting or correcting potential LLM misinterpretations of NPU constraints (such as memory hierarchy or implicit scheduling) when execution feedback is noisy. The current design relies on the Triple-Part Executable Knowledge Representation to couple error context directly with executable semantics, and on iterative effect-driven distillation using empirical feedback to prune errors. However, to strengthen the presentation, we will revise §3 to include an explicit discussion of how semantic arbitration is anchored in execution outcomes and to report empirical checks on arbitration consistency across feedback iterations. This addresses the load-bearing nature of the claim. revision: yes

  2. Referee: [Abstract and experimental evaluation section] Abstract and experimental evaluation section: The manuscript states concrete numerical gains (49.4% → 80.0% accuracy, 2.2× speedup) but supplies no baseline descriptions, workload/dataset details, number of trials, or error bars. Without these, it is impossible to determine whether the data support the attribution to the three modules rather than implementation artifacts or cherry-picked cases.

    Authors: We agree that the experimental evaluation section provides high-level results but omits the requested protocol details. In the revised manuscript we will add a dedicated subsection specifying the SOTA baselines and their configurations, the exact real-world NPU workloads and datasets, the number of independent trials, and error bars (or standard deviations) for all reported metrics. This will enable readers to evaluate the robustness of the gains and their attribution to Hawk's modules. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results from training-free modules

full rationale

The paper presents Hawk as a training-free three-module framework and reports accuracy (49.4% to 80.0%) and speedup (2.2x) as direct empirical outcomes of evaluations on real-world NPU workloads. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or description that would reduce the claimed results to the inputs by construction. The Effect-Driven Knowledge Distillation Module is described as using LLM arbitration on execution feedback, but this is presented as an operational step whose reliability is evaluated externally rather than defined into the metrics. The derivation chain is therefore self-contained against the reported benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 3 invented entities

The paper introduces three new modules as core invented components. The central assumption is that LLMs can leverage synthesized hardware knowledge effectively. No free parameters are mentioned.

axioms (1)
  • domain assumption Large language models can produce correct NPU kernels when supplied with appropriate hardware-aware executable knowledge
    The entire framework rests on LLMs being able to use the synthesized and distilled knowledge to generate working code.
invented entities (3)
  • Run-Time Knowledge Synthesis Module no independent evidence
    purpose: Creates Triple-Part Executable Knowledge Representation that couples error context with executable semantics
    New component introduced to address the lack of hardware priors in LLMs
  • Bottleneck-Aware Knowledge Retrieval Module no independent evidence
    purpose: Implements 2D-Retrieval projecting queries into syntactic and hardware semantic spaces
    New component introduced to locate relevant hardware knowledge
  • Effect-Driven Knowledge Distillation Module no independent evidence
    purpose: Uses LLM-driven semantic arbitration to prune errors and consolidate redundancies from execution feedback
    New component introduced to refine knowledge without training

pith-pipeline@v0.9.1-grok · 5790 in / 1583 out tokens · 38766 ms · 2026-07-03T14:58:52.991731+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 19 canonical work pages · 5 internal anchors

  1. [1]

    DeepSeek-V4: Towards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348, 2026

    A. Xu, B. Lin, B. Xue, B. Wang, B. Xu, B. Wu, B. Zhang, C. Lin, C. Dong, C. Linget al., “Deepseek-v4: Towards highly efficient million-token context intelligence,”arXiv preprint arXiv:2606.19348, 2026

    work page arXiv 2026

  2. [2]

    Qwen3.5-Omni Technical Report

    Q. Team, “Qwen3. 5-omni technical report,”arXiv preprint arXiv:2604.15804, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    GLM-5: from Vibe Coding to Agentic Engineering

    A. Zeng, X. Lv, Z. Hou, Z. Du, Q. Zheng, B. Chen, D. Yin, C. Ge, C. Huang, C. Xieet al., “Glm-5: from vibe coding to agentic engineering,”arXiv preprint arXiv:2602.15763, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  4. [4]

    GPT-4 Technical Report

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat et al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    AscendKernelGen: A Systematic Study of LLM-Based Kernel Generation for Neural Processing Units

    X. Cao, J. Zhai, P. Li, Z. Hu, C. Yan, B. Mu, G. Fang, B. She, J. Li, Y . Suet al., “Ascendkernelgen: A systematic study of llm-based kernel generation for neural processing units,”arXiv preprint arXiv:2601.07160, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    Towards cold- start drafting and continual refining: A value-driven memory approach with application to npu kernel synthesis,

    Y . Zheng, Z. Li, S. Zhang, H. Wang, J. Sheng, J. Wang, J. Yan, W. Zhang, Y . Wen, B. Tanget al., “Towards cold- start drafting and continual refining: A value-driven memory approach with application to npu kernel synthesis,”arXiv preprint arXiv:2603.10846, 2026

    work page arXiv 2026

  7. [7]

    As- cendcraft: Automatic ascend npu kernel generation via dsl-guided transcompilation,

    Z. Wen, S. Shao, Z. Li, Y . Ge, T. Xu, Y . Lin, and T. Zhang, “As- cendcraft: Automatic ascend npu kernel generation via dsl-guided transcompilation,”arXiv preprint arXiv:2601.22760, 2026

    work page arXiv 2026

  8. [8]

    Akg: automatic kernel generation for neural processing units using polyhedral transformations,

    J. Zhao, B. Li, W. Nie, Z. Geng, R. Zhang, X. Gao, B. Cheng, C. Wu, Y . Cheng, Z. Liet al., “Akg: automatic kernel generation for neural processing units using polyhedral transformations,” inProceedings of the 42nd ACM SIGPLAN International Con- ference on Programming Language Design and Implementation, 2021, pp. 1233–1248

    2021

  9. [9]

    To- wards better answers: Automated stack overflow post updating,

    Y . Mai, Z. Gao, H. Wang, T. Bi, X. Hu, X. Xia, and J. Sun, “To- wards better answers: Automated stack overflow post updating,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE, 2025, pp. 591–603

    2025

  10. [10]

    Morest: Model-based restful api testing with execution feedback,

    Y . Liu, Y . Li, G. Deng, Y . Liu, R. Wan, R. Wu, D. Ji, S. Xu, and M. Bao, “Morest: Model-based restful api testing with execution feedback,” inProceedings of the 44th International Conference on Software Engineering, 2022, pp. 1406–1417

    2022

  11. [11]

    Kernelllm: Making kernel development more accessible,

    Z. V . Fisches, S. Paliskara, S. Guo, A. Zhang, J. Spisak, C. Cummins, H. Leather, G. Synnaeve, J. Isaacson, A. Markosyan, and M. Saroufim, “Kernelllm: Making kernel development more accessible,” 6 2025. [Online]. Available: https://huggingface.co/facebook/KernelLLM

    2025

  12. [12]

    Reinforcement tuning open source llms for kernel generation

    A. Garg, J. Heo, and M. Mou, “Reinforcement tuning open source llms for kernel generation.” [Online]. Available: https://api.semanticscholar.org/CorpusID:281103975

  13. [13]

    Tritonrl: Training llms to think and code triton without cheating,

    J. Woo, S. Zhu, A. Nie, Z. Jia, Y . Wang, and Y . Park, “Tritonrl: Training llms to think and code triton without cheating,”arXiv preprint arXiv:2510.17891, 2025

    work page arXiv 2025

  14. [14]

    React: Ir-level patch presence test for binary,

    Q. Zhan, X. Hu, X. Xia, and S. Li, “React: Ir-level patch presence test for binary,” inProceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering, 2024, pp. 381–392

    2024

  15. [15]

    Type- alias analysis: Enabling llvm ir with accurate types,

    J. Zhou, Z. Pan, W. Shen, X. Wang, K. Lu, and Z. Qian, “Type- alias analysis: Enabling llvm ir with accurate types,”Proceedings of the ACM on Software Engineering, vol. 2, no. ISSTA, pp. 2203–2226, 2025

    2025

  16. [16]

    Robust vul- nerability detection across compilations: Llvm-ir vs. assembly with transformer model,

    R. Shir, P. Surve, Y . Elovici, and A. Shabtai, “Robust vul- nerability detection across compilations: Llvm-ir vs. assembly with transformer model,”Proceedings of the ACM on Software Engineering, vol. 2, no. ISSTA, pp. 618–639, 2025

    2025

  17. [17]

    The incredible shrinking context... in a decompiler near you,

    S. Lagouvardos, Y . Bollanos, N. Grech, and Y . Smaragdakis, “The incredible shrinking context... in a decompiler near you,” Proceedings of the ACM on Software Engineering, vol. 2, no. ISSTA, pp. 1350–1373, 2025

    2025

  18. [18]

    Few-shot natural language to first-order logic translation via code generation,

    J. Liu, “Few-shot natural language to first-order logic translation via code generation,” inProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2025, pp. 10 939–10 960

    2025

  19. [19]

    Why in-context learning models are good few-shot learners?

    S. Wu, Y . Wang, and Q. Yao, “Why in-context learning models are good few-shot learners?” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  20. [20]

    A theory of emergent in-context learning as implicit structure induction

    M. Hahn and N. Goyal, “A theory of emergent in-context learning as implicit structure induction,”arXiv preprint arXiv:2303.07971, 2023

    work page arXiv 2023

  21. [21]

    On edge- fog-cloud collaboration and reaping its benefits: a heterogeneous multi-tier edge computing architecture,

    N. Fernando, S. Shrestha, S. W. Loke, and K. Lee, “On edge- fog-cloud collaboration and reaping its benefits: a heterogeneous multi-tier edge computing architecture,”Future Internet, vol. 17, no. 1, p. 22, 2025

    2025

  22. [22]

    Heterogeneous multi-computing-units based lt coded computation in wireless networks,

    B. Fang and X. Qiao, “Heterogeneous multi-computing-units based lt coded computation in wireless networks,” in2026 IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2026, pp. 1–6

    2026

  23. [23]

    Skybyte: architecting an efficient memory-semantic cxl-based ssd with os and hardware co-design,

    H. Zhang, Y . Xue, Y . E. Zhou, S. Li, and J. Huang, “Skybyte: architecting an efficient memory-semantic cxl-based ssd with os and hardware co-design,” in2025 IEEE International Symposium on High Performance Computer Architecture (HPCA). IEEE, 2025, pp. 577–593

    2025

  24. [24]

    Faster and stronger: Unleashing data processing potential through hardware heterogeneity,

    C. Wang, Y . Luo, W. Du, K. Wang, N. Gu, and J. Yu, “Faster and stronger: Unleashing data processing potential through hardware heterogeneity,”IEEE Internet of Things Journal, vol. 12, no. 10, pp. 14 559–14 576, 2025

    2025

  25. [25]

    Cuda agent: Large-scale agentic rl for high-performance cuda kernel generation,

    W. Dai, H. Wu, Q. Yu, H.-a. Gao, J. Li, C. Jiang, W. Lou, Y . Song, H. Yu, J. Chenet al., “Cuda agent: Large-scale agentic rl for high-performance cuda kernel generation,”arXiv preprint arXiv:2602.24286, 2026

    work page arXiv 2026

  26. [26]

    Kevin: Multi-turn rl for generating cuda kernels,

    C. Baronio, P. Marsella, B. Pan, S. Guo, and S. Alberti, “Kevin: Multi-turn rl for generating cuda kernels,”arXiv preprint arXiv:2507.11948, 2025

    work page arXiv 2025

  27. [27]

    Cuda-l1: Improving cuda optimization via contrastive reinforcement learning,

    X. Li, X. Sun, A. Wang, J. Li, and C. Shum, “Cuda-l1: Improving cuda optimization via contrastive reinforcement learning,”arXiv preprint arXiv:2507.14111, 2025

    work page arXiv 2025

  28. [28]

    Kernelblaster: Continual cross-task cuda optimization via memory-augmented in-context reinforcement learning,

    K. S. Dong, S. Modi, D. Nikiforov, S. Damani, E. Lin, S. K. S. Hari, and C. Kozyrakis, “Kernelblaster: Continual cross-task cuda optimization via memory-augmented in-context reinforcement learning,”arXiv preprint arXiv:2602.14293, 2026

    work page arXiv 2026

  29. [29]

    Stark: Strategic team of agents for refining kernels,

    J. Dong, Y . Yang, T. Liu, Y . Wang, F. Qi, V . Tarokh, K. Ran- gadurai, and S. Yang, “Stark: Strategic team of agents for refining kernels,”arXiv preprint arXiv:2510.16996, 2025

    work page arXiv 2025

  30. [30]

    Cud- aforge: An agent framework with hardware feedback for cuda kernel optimization,

    Z. Zhang, R. Wang, S. Li, Y . Luo, M. Hong, and C. Ding, “Cud- aforge: An agent framework with hardware feedback for cuda kernel optimization,”arXiv preprint arXiv:2511.01884, 2025

    work page arXiv 2025

  31. [31]

    Avo: agentic variation operators for au- tonomous evolutionary search,

    T. Chen, Z. Ye, B. Xu, Z. Ye, T. Liu, A. Hassani, T. Chen, A. Kerr, H. Wu, Y . Xuet al., “Avo: agentic variation operators for au- tonomous evolutionary search,”arXiv preprint arXiv:2603.24517, 2026

    work page arXiv 2026

  32. [32]

    Multik- ernelbench: A multi-platform benchmark for kernel generation,

    Z. Wen, Y . Zhang, Z. Li, Z. Liu, L. Xie, and T. Zhang, “Multik- ernelbench: A multi-platform benchmark for kernel generation,” arXiv e-prints, pp. arXiv–2507, 2025

    2025

  33. [33]

    (2026) ops-nn

    Huawei. (2026) ops-nn. [Online]. Available: https://gitcode.com/cann/ops-nn

    2026

  34. [34]

    (2026) Cannbot

    ——. (2026) Cannbot. [Online]. Available: https://gitcode.com/cann/cannbot-skills

    2026

  35. [35]

    Injecting the bm25 score as text improves bert-based re-rankers,

    A. Askari, A. Abolghasemi, G. Pasi, W. Kraaij, and S. Verberne, “Injecting the bm25 score as text improves bert-based re-rankers,” inEuropean Conference on Information Retrieval. Springer, 2023, pp. 66–83

    2023

  36. [36]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models

    Y . Zhang, M. Li, D. Long, X. Zhang, H. Lin, B. Yang, P. Xie, A. Yang, D. Liu, J. Linet al., “Qwen3 embedding: Advancing text embedding and reranking through foundation models,”arXiv preprint arXiv:2506.05176, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    Reciprocal rank fusion outperforms condorcet and individual rank learning methods,

    G. V . Cormack, C. L. Clarke, and S. Buettcher, “Reciprocal rank fusion outperforms condorcet and individual rank learning methods,” inProceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, 2009, pp. 758–759

    2009

  38. [38]

    Rag-fusion: a new take on retrieval-augmented generation.arXiv preprint arXiv:2402.03367, 2024

    Z. Rackauckas, “Rag-fusion: a new take on retrieval-augmented generation,”arXiv preprint arXiv:2402.03367, 2024

    work page arXiv 2024

  39. [39]

    Mmmorrf: Multimodal multilingual modularized reciprocal rank fusion,

    S. Samuel, D. DeGenaro, J. Guallar-Blasco, K. Sanders, S. Eis- ape, A. Reddy, A. Martin, A. Yates, E. Yang, C. Carpenteret al., “Mmmorrf: Multimodal multilingual modularized reciprocal rank fusion,” inProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Re- trieval, 2025, pp. 4004–4009

    2025

  40. [40]

    Does few-shot learning help llm performance in code synthesis?

    D. Xu, T. Xie, B. Xia, H. Li, Y . Bai, Y . Sun, and W. Wang, “Does few-shot learning help llm performance in code synthesis?”arXiv preprint arXiv:2412.02906, 2024

    work page arXiv 2024