pith. sign in

arxiv: 2606.12203 · v1 · pith:3PKR5ME6new · submitted 2026-06-10 · 💻 cs.CL

Adaptive Multi-Resolution Procedural Knowledge Compression for Large Language Models

Changyue Wang , Weihang Su , Qingyao Ai , Yichen Tang , Runzhong Qiao , Xuancheng Li , Min Zhang , Yiqun Liu This is my paper

Pith reviewed 2026-06-27 09:43 UTC · model grok-4.3

classification 💻 cs.CL
keywords skill compressionprocedural knowledgelarge language modelssoft tokensmulti-resolutionadaptive compressionLLM inference efficiency
0
0 comments X

The pith

SKIM compresses procedural skills for LLMs to 30-60 percent of original token length while preserving performance.

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

Large language models rely on reusable natural language skills for autonomous workflows, but inserting full skill text repeatedly raises prefill costs. Most compression techniques target factual documents and do not maintain the logical dependencies required by procedural skills. SKIM provides an adaptive framework that produces different numbers of soft tokens according to each skill's complexity. The approach supports offline compression and meets the need for methods that scale across simple and complex skills. Experiments show the resulting compression retains task performance more effectively than prior techniques.

Core claim

SKIM is an adaptive multi-resolution soft token compression framework for procedural skills. Depending on the complexity of each skill, SKIM creates different numbers of soft tokens that not only improve the efficiency of LLM inference, but also preserve the effectiveness of skill usage. It satisfies three requirements: preserving logical dependencies among workflows and tool protocols, enabling lightweight offline compression for frequently updated skills, and adapting to varying complexities across skills.

What carries the argument

Adaptive multi-resolution soft token compression framework (SKIM) that produces varying numbers of soft tokens per skill based on its complexity.

If this is right

  • Repeated skill invocations incur lower prefill cost and latency in LLM contexts.
  • Task performance is retained more reliably than with compression methods built for factual text.
  • Community skills can be compressed offline without heavy computational overhead.
  • Compression ratios adjust automatically to the complexity of each individual skill.

Where Pith is reading between the lines

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

  • The same resolution-adaptive token mechanism could be tested on compressing multi-step agent plans or code snippets.
  • SKIM-style compression might combine with KV-cache optimizations to further reduce memory during long workflows.
  • Dynamic selection of resolution during inference could be explored if skill usage patterns change within a single session.

Load-bearing premise

Soft tokens created at different resolutions can preserve logical dependencies among workflows and tool protocols for skills of varying complexity.

What would settle it

A direct comparison showing that SKIM-compressed skills produce substantially lower success rates than full-text skills on tasks that chain multiple tool protocols.

Figures

Figures reproduced from arXiv: 2606.12203 by Changyue Wang, Min Zhang, Qingyao Ai, Runzhong Qiao, Weihang Su, Xuancheng Li, Yichen Tang, Yiqun Liu.

Figure 1
Figure 1. Figure 1: A ToolQA example. Factual information can [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SKIM. The upper panel illustrates the training process and offline resolution selection, where the compressor (consisting of slot tokens, an LLM-based compressor, and an MLP projector) learns to convert skills into soft-token prefixes aligned with the target LLM’s embedding space. The lower panel shows the inference architecture, where SKIM selects the soft-token prefix corresponding to the off… view at source ↗
Figure 3
Figure 3. Figure 3: Average token accuracy tradeoff across the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: BigCodeBench stress results under retrieved [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of final SKIM resolution choices for Qwen3-8B across the datasets in [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Untrained resolution budget ablation for [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Third stage LoRA ablation for Qwen3-8B on BigCodeBench, CHAMP, LogicBench, and Theo￾remQA. This ablation is run in the skill task alignment stage. Frozen trains the compressor side while keeping the target LLM weights fixed, and the rank variants use LoRA with alpha set to twice the rank. instead of deciding whether to load any skill in￾formation. Moreover, applying the exam without compressed candidates u… view at source ↗
Figure 9
Figure 9. Figure 9: Offline resolution selection ablations for Qwen3-8B on BigCodeBench, CHAMP, LogicBench, and [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
read the original abstract

Large language models (LLMs) are widely used to tackle complex tasks with autonomous workflows. Recently, reusable natural language skills have emerged as a popular paradigm to inject procedural knowledge into LLM applications. Since popular skills are often invoked repeatedly, placing their full text in every context significantly increases prefill cost and latency. While text compression techniques have the potential to solve this problem, most existing methods are designed to compress factual knowledge in documents instead of procedural knowledge, making them insufficient for skill compression. In this paper, we argue that an effective skill compression method should: 1) preserve logical dependencies among workflows and tool protocols, 2) enable lightweight, offline compression for frequently updated community skills, and 3) be adaptable to varying complexities across skills. To address this, we present SKIM (SKIll coMpression), an adaptive multi-resolution soft token compression framework for procedural skills. Depending on the complexity of each skill, SKIM creates different numbers of soft tokens that not only improve the efficiency of LLM inference, but also preserve the effectiveness of skill usage. Experiments indicate that SKIM compresses skills to 30 to 60 percent of their original token length while preserving task performance better than existing compression methods.We have released our code at https://github.com/bebr2/SKIM .

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

0 major / 2 minor

Summary. The paper proposes SKIM, an adaptive multi-resolution soft token compression framework for procedural skills in LLMs. It creates varying numbers of soft tokens per skill based on complexity to meet three requirements: preserving logical dependencies among workflows and tool protocols, enabling lightweight offline compression, and adapting to skill complexities. The central empirical claim is that SKIM compresses skills to 30-60% of original token length while preserving task performance better than existing methods; code is released at https://github.com/bebr2/SKIM.

Significance. If the end-to-end task performance results hold across procedural skills of varying complexity, this addresses a practical bottleneck in LLM applications that repeatedly invoke reusable natural language skills, potentially lowering prefill costs and latency. The open release of code supports reproducibility, which strengthens the contribution for an empirical methods paper in this area.

minor comments (2)
  1. Abstract: the sentence 'Experiments indicate that SKIM compresses skills to 30 to 60 percent of their original token length while preserving task performance better than existing compression methods.We have released our code' is missing a space after the period.
  2. The abstract states the three requirements an effective method must meet but does not preview how the experiments directly test preservation of logical dependencies (requirement 1); the full paper should make this mapping explicit in the evaluation section.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary, recognition of the code release, and recommendation of minor revision. The referee's description of SKIM and its goals is accurate. No major comments appear in the report, so we have no points requiring rebuttal or clarification at this stage.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents SKIM as an empirical framework for skill compression rather than a mathematical derivation. The abstract and description contain no equations, fitted parameters, or load-bearing self-citations that reduce claims to inputs by construction. The three requirements are stated design goals, and results are reported from end-to-end experiments on task performance. No self-definitional, fitted-prediction, or uniqueness-imported steps are identifiable from the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Abstract supplies only high-level requirements and a headline result; no explicit free parameters, axioms, or invented entities are quantified.

free parameters (1)
  • number of soft tokens per skill
    Determined adaptively by skill complexity; exact mapping or fitting procedure not described.
axioms (1)
  • domain assumption Soft tokens at multiple resolutions can preserve logical dependencies among workflows and tool protocols
    Listed as the first required property of an effective skill compression method.
invented entities (1)
  • multi-resolution soft tokens no independent evidence
    purpose: Variable-length compression units that adapt to skill complexity
    Core new representation introduced by the framework

pith-pipeline@v0.9.1-grok · 5781 in / 1142 out tokens · 13914 ms · 2026-06-27T09:43:09.061974+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

68 extracted references · 11 canonical work pages

  1. [1]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  2. [2]

    2024 , eprint=

    Phi-4 Technical Report , author=. 2024 , eprint=

    2024

  3. [3]

    2026 , eprint=

    TokMem: One-Token Procedural Memory for Large Language Models , author=. 2026 , eprint=

    2026

  4. [4]

    2026 , eprint=

    A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications , author=. 2026 , eprint=

    2026

  5. [8]

    2026 , eprint=

    Decoupled Reasoning with Implicit Fact Tokens (DRIFT): A Dual-Model Framework for Efficient Long-Context Inference , author=. 2026 , eprint=

    2026

  6. [9]

    In-context Autoencoder for Context Compression in a Large Language Model , booktitle =

    Tao Ge and Jing Hu and Lei Wang and Xun Wang and Si. In-context Autoencoder for Context Compression in a Large Language Model , booktitle =. 2024 , url =

    2024

  7. [10]

    2025 , eprint=

    DeepSeek-OCR: Contexts Optical Compression , author=. 2025 , eprint=

    2025

  8. [11]

    2026 , eprint=

    Autoencoding-Free Context Compression for LLMs via Contextual Semantic Anchors , author=. 2026 , eprint=

    2026

  9. [12]

    2026 , eprint=

    COMI: Coarse-to-fine Context Compression via Marginal Information Gain , author=. 2026 , eprint=

    2026

  10. [13]

    Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region , pages=

    Mitigating entity-level hallucination in large language models , author=. Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region , pages=

    2024

  11. [14]

    Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region , pages=

    Dynamic and Parametric Retrieval Augmented Generation , author=. Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region , pages=

    2025

  12. [15]

    Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Parametric retrieval augmented generation , author=. Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  13. [16]

    2026 , eprint=

    Skill Retrieval Augmentation for Agentic AI , author=. 2026 , eprint=

    2026

  14. [17]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Dragin: Dynamic retrieval augmented generation based on the real-time information needs of large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  15. [18]

    CoRR , volume =

    Tri Nguyen and Mir Rosenberg and Xia Song and Jianfeng Gao and Saurabh Tiwary and Rangan Majumder and Li Deng , title =. CoRR , volume =. 2016 , url =

    2016

  16. [27]

    Applebaum and Zain Anwar and Maame Sarfo

    Nikhil Khandekar and Qiao Jin and Guangzhi Xiong and Soren Dunn and Serina S. Applebaum and Zain Anwar and Maame Sarfo. MedCalc-Bench: Evaluating Large Language Models for Medical Calculations , booktitle =. 2024 , url =

    2024

  17. [28]

    Yuchen Zhuang and Yue Yu and Kuan Wang and Haotian Sun and Chao Zhang , editor =. ToolQA:. Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

    2023

  18. [29]

    BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions , booktitle =

    Terry Yue Zhuo and Minh Chien Vu and Jenny Chim and Han Hu and Wenhao Yu and Ratnadira Widyasari and Imam Nur Bani Yusuf and Haolan Zhan and Junda He and Indraneil Paul and Simon Brunner and Chen Gong and James Hoang and Armel Randy Zebaze and Xiaoheng Hong and Wen. BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instruc...

    2025

  19. [30]

    Narasimhan and Yuan Cao , title =

    Shunyu Yao and Jeffrey Zhao and Dian Yu and Nan Du and Izhak Shafran and Karthik R. Narasimhan and Yuan Cao , title =. The Eleventh International Conference on Learning Representations,. 2023 , url =

    2023

  20. [31]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen. LoRA: Low-Rank Adaptation of Large Language Models , booktitle =. 2022 , url =

    2022

  21. [32]

    2026 , note =

    Luca Beurer-Kellner and Aleksei Kudrinskii and Marco Milanta and Kristian Bonde Nielsen and Hemang Sarkar and Liran Tal , title =. 2026 , note =

    2026

  22. [33]

    2025 , month = dec, url =

    2025

  23. [34]

    2026 , eprint=

    OpenAI GPT-5 System Card , author=. 2026 , eprint=

    2026

  24. [35]

    2018 , eprint=

    WikiHow: A Large Scale Text Summarization Dataset , author=. 2018 , eprint=

    2018

  25. [36]

    2025 , eprint=

    Context Cascade Compression: Exploring the Upper Limits of Text Compression , author=. 2025 , eprint=

    2025

  26. [37]

    2026 , eprint=

    SkVM: Revisiting Language VM for Skills across Heterogenous LLMs and Harnesses , author=. 2026 , eprint=

    2026

  27. [38]

    2026 , eprint=

    Red Skills or Blue Skills? A Dive Into Skills Published on ClawHub , author=. 2026 , eprint=

    2026

  28. [39]

    2026 , eprint=

    HarmfulSkillBench: How Do Harmful Skills Weaponize Your Agents? , author=. 2026 , eprint=

    2026

  29. [40]

    2026 , note =

    OpenClaw , title =. 2026 , note =

    2026

  30. [41]

    2026 , eprint=

    SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks , author=. 2026 , eprint=

    2026

  31. [42]

    2026 , eprint=

    SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents , author=. 2026 , eprint=

    2026

  32. [43]

    Hewett, Mojan Javaheripi, Piero Kauffmann, James R

    Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, and 8 others. 2024. https://arxiv.org/abs/2412.08905 Phi-4 technic...

    Pith/arXiv arXiv 2024

  33. [44]

    Le Chen, Erhu Feng, Yubin Xia, and Haibo Chen. 2026. https://arxiv.org/abs/2604.03088 Skvm: Revisiting language vm for skills across heterogenous llms and harnesses . Preprint, arXiv:2604.03088

    Pith/arXiv arXiv 2026

  34. [45]

    Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.489 Theoremqa: A theorem-driven question answering dataset . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023 , pages 7889-...

    work page doi:10.18653/v1/2023.emnlp-main.489 2023

  35. [46]

    Hongcheol Cho, Ryangkyung Kang, and Youngeun Kim. 2026. https://arxiv.org/abs/2605.05726 Skillret: A large-scale benchmark for skill retrieval in llm agents . Preprint, arXiv:2605.05726

    Pith/arXiv arXiv 2026

  36. [47]

    Tao Ge, Jing Hu, Lei Wang, Xun Wang, Si - Qing Chen, and Furu Wei. 2024. https://openreview.net/forum?id=uREj4ZuGJE In-context autoencoder for context compression in a large language model . In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net

    2024

  37. [48]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lora: Low-rank adaptation of large language models . In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net

    2022

  38. [49]

    Haichuan Hu, Ye Shang, and Quanjun Zhang. 2026. https://arxiv.org/abs/2604.13064 Red skills or blue skills? a dive into skills published on clawhub . Preprint, arXiv:2604.13064

    Pith/arXiv arXiv 2026

  39. [50]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. https://doi.org/10.18653/v1/2024.acl-long.91 L ong LLML ingua: Accelerating and enhancing LLM s in long context scenarios via prompt compression . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    work page doi:10.18653/v1/2024.acl-long.91 2024

  40. [51]

    Yukun Jiang, Yage Zhang, Michael Backes, Xinyue Shen, and Yang Zhang. 2026. https://arxiv.org/abs/2604.15415 Harmfulskillbench: How do harmful skills weaponize your agents? Preprint, arXiv:2604.15415

    Pith/arXiv arXiv 2026

  41. [52]

    Applebaum, Zain Anwar, Maame Sarfo - Gyamfi, Conrad W

    Nikhil Khandekar, Qiao Jin, Guangzhi Xiong, Soren Dunn, Serina S. Applebaum, Zain Anwar, Maame Sarfo - Gyamfi, Conrad W. Safranek, Abid A Anwar, Andrew Zhang, Aidan Gilson, Maxwell B. Singer, Amisha D. Dave, Andrew Taylor, Aidong Zhang, Qingyu Chen, and Zhiyong Lu. 2024. http://papers.nips.cc/paper\_files/paper/2024/hash/99e81750f3fdfcaf9613db2dbf4bd623-A...

    2024

  42. [53]

    Mahnaz Koupaee and William Yang Wang. 2018. https://arxiv.org/abs/1810.09305 Wikihow: A large scale text summarization dataset . Preprint, arXiv:1810.09305

    Pith/arXiv arXiv 2018

  43. [54]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. https://doi.org/10.1145/3600006.3613165 Efficient memory management for large language model serving with pagedattention . In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP '23, page 611–626, New York, ...

    work page doi:10.1145/3600006.3613165 2023

  44. [55]

    Xuancheng Li, Haitao Li, Yujia Zhou, Qingyao Ai, and Yiqun Liu. 2025 a . https://doi.org/10.1145/3767695.3769499 Atacompressor: Adaptive task-aware compression for efficient long-context processing in llms . In Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region,...

    work page doi:10.1145/3767695.3769499 2025

  45. [56]

    Zongqian Li, Yixuan Su, and Nigel Collier. 2025 b . https://doi.org/10.18653/v1/2025.acl-long.1219 500x C ompressor: Generalized prompt compression for large language models . In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 25081--25091, Vienna, Austria. Association for Computationa...

    work page doi:10.18653/v1/2025.acl-long.1219 2025

  46. [57]

    Fanfan Liu and Haibo Qiu. 2025. https://arxiv.org/abs/2511.15244 Context cascade compression: Exploring the upper limits of text compression . Preprint, arXiv:2511.15244

    arXiv 2025

  47. [58]

    Yujun Mao, Yoon Kim, and Yilun Zhou. 2024. https://doi.org/10.18653/V1/2024.FINDINGS-ACL.785 CHAMP: A competition-level dataset for fine-grained analyses of llms' mathematical reasoning capabilities . In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 , Findings of ACL , pages ...

    work page doi:10.18653/v1/2024.findings-acl.785 2024

  48. [59]

    Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. https://arxiv.org/abs/1611.09268 MS MARCO: A human generated machine reading comprehension dataset . CoRR, abs/1611.09268

    Pith/arXiv arXiv 2016

  49. [60]

    OpenAI . 2025. https://openai.com/index/introducing-gpt-5-2/ Introducing GPT-5.2 . OpenAI blog post

    2025

  50. [61]

    OpenClaw. 2026 a . C law H ub --- clawhub.ai. https://clawhub.ai/. [Accessed 19-05-2026]

    2026

  51. [62]

    OpenClaw. 2026 b . O pen C law — P ersonal A I A ssistant --- openclaw.ai. https://openclaw.ai/. [Accessed 19-05-2026]

    2026

  52. [63]

    Vicky Zhao, Lili Qiu, and Dongmei Zhang

    Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Menglin Xia, Xufang Luo, Jue Zhang, Qingwei Lin, Victor R \"u hle, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Dongmei Zhang. 2024. https://doi.org/10.18653/v1/2024.findings-acl.57 LLML ingua-2: Data distillation for efficient and faithful task-agnostic prompt compression . In Findings of the Associatio...

    work page doi:10.18653/v1/2024.findings-acl.57 2024

  53. [64]

    Mihir Parmar, Nisarg Patel, Neeraj Varshney, Mutsumi Nakamura, Man Luo, Santosh Mashetty, Arindam Mitra, and Chitta Baral. 2024. https://doi.org/10.18653/v1/2024.acl-long.739 L ogic B ench: Towards systematic evaluation of logical reasoning ability of large language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Li...

    work page doi:10.18653/v1/2024.acl-long.739 2024

  54. [65]

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. 2020. https://doi.org/10.1145/3394486.3406703 Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters . In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '20, page 3505–3506, New York, NY, U...

    work page doi:10.1145/3394486.3406703 2020

  55. [66]

    Weihang Su, Qian Dong, Qingyao Ai, and Yiqun Liu. 2025 a . Dynamic and parametric retrieval augmented generation. In Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, pages 453--458

    2025

  56. [67]

    Weihang Su, Jianming Long, Qingyao Ai, Yichen Tang, Changyue Wang, Yiteng Tu, and Yiqun Liu. 2026. https://arxiv.org/abs/2604.24594 Skill retrieval augmentation for agentic ai . Preprint, arXiv:2604.24594

    Pith/arXiv arXiv 2026

  57. [68]

    Weihang Su, Yichen Tang, Qingyao Ai, Changyue Wang, Zhijing Wu, and Yiqun Liu. 2024 a . Mitigating entity-level hallucination in large language models. In Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, pages 23--31

    2024

  58. [69]

    Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024 b . Dragin: Dynamic retrieval augmented generation based on the real-time information needs of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12991--13013

    2024

  59. [70]

    Weihang Su, Yichen Tang, Qingyao Ai, Junxi Yan, Changyue Wang, Hongning Wang, Ziyi Ye, Yujia Zhou, and Yiqun Liu. 2025 b . Parametric retrieval augmented generation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1240--1250

    2025

  60. [71]

    Zijun Wu, Yongchang Hao, and Lili Mou. 2026. https://arxiv.org/abs/2510.00444 Tokmem: One-token procedural memory for large language models . Preprint, arXiv:2510.00444

    arXiv 2026

  61. [72]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. https://doi.org/10.1145/3626772.3657878 C-pack: Packed resources for general chinese embeddings . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, page 641–649, New York, NY, USA. Assoc...

    work page doi:10.1145/3626772.3657878 2024

  62. [73]

    Wenxuan Xie, Yujia Wang, Xin Tan, Chaochao Lu, Xia Hu, and Xuhong Wang. 2026. https://arxiv.org/abs/2602.10021 Decoupled reasoning with implicit fact tokens (drift): A dual-model framework for efficient long-context inference . Preprint, arXiv:2602.10021

    arXiv 2026

  63. [74]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    Pith/arXiv arXiv 2025

  64. [75]

    Cohen, Ruslan Salakhutdinov, and Christopher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/V1/D18-1259 Hotpotqa: A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October...

    work page doi:10.18653/v1/d18-1259 2018

  65. [76]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. https://openreview.net/forum?id=WE\_vluYUL-X React: Synergizing reasoning and acting in language models . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net

    2023

  66. [77]

    Yingli Zhou, Wang Shu, Yaodong Su, Wenchuan Du, Yixiang Fang, and Xuemin Lin. 2026. https://arxiv.org/abs/2605.07358 A comprehensive survey on agent skills: Taxonomy, techniques, and applications . Preprint, arXiv:2605.07358

    Pith/arXiv arXiv 2026

  67. [78]

    Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/9cb2a7495900f8b602cb10159246a016-Abstract-Datasets\_and\_Benchmarks.html Toolqa: A dataset for LLM question answering with external tools . In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information ...

    2023

  68. [79]

    Terry Yue Zhuo, Minh Chien Vu, Jenny Chim, Han Hu, Wenhao Yu, Ratnadira Widyasari, Imam Nur Bani Yusuf, Haolan Zhan, Junda He, Indraneil Paul, Simon Brunner, Chen Gong, James Hoang, Armel Randy Zebaze, Xiaoheng Hong, Wen - Ding Li, Jean Kaddour, Ming Xu, Zhihan Zhang, and 2 others. 2025. https://openreview.net/forum?id=YrycTjllL0 Bigcodebench: Benchmarkin...

    2025