Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Deep Agentic Search for Repository-Level Code Question Answering: An Empirical Study

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Indexed retrieval beats subagent exploration for repository code questions

desk verdict A careful, honest empirical study—Semantic beats Deep Agentic on SWE-QA accuracy and cost across four models—but the headline magnitude, especially the 41.8% coordination-failure story, is tied to one harness's defaults and needs artifacts and a sensitivity analysis before it generalizes. read the letter →

arxiv 2608.01507 v1 pith:ZVE64JCL submitted 2026-08-02 cs.SE cs.AIcs.IRcs.MA

Amirkia Rafiei Oskooei , Bora Ilci , Alperen Kayim , Mehmet Egemen Uzun , Berat Can , Kaan Emre Kara , Ozan Orhan , Mehmet S. Aktas This is my paper
classification cs.SEcs.AIcs.IRcs.MA
keywords codequestionansweringdeepagenticsearchsemanticcontextengineeringagentsfailuretaxonomySWE-QAbenchmarkretrieval-augmentedgeneration
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

The paper pits the two dominant designs for repository-level code question answering against each other: a semantic-search agent that retrieves code chunks from a vector index built in advance, and a deep agentic-search agent that delegates live exploration to an isolated sub-agent. Across four language models and 720 questions from fifteen Python repositories, semantic search answered 65.2% of questions correctly versus 46.2% for deep agentic search, and produced each correct answer at roughly half the cost. The gap was consistent in direction for every model and at every pass threshold. Coding every failed run shows why: deep agentic search introduced a new class of failure at the planner-to-sub-agent hand-off, which was usually silent and produced fluent but wrong answers. The paper argues that for read-only questions over an indexable repository, the newer, more elaborate design is not the better one.

What carries the argument

The load-bearing mechanism is the planner-to-sub-agent delegation hand-off in Deep Agentic search, contrasted with the pre-built vector index of Semantic search. In the deep design, a planner keeps its context clean by sending an isolated sub-agent to explore with terminal-style commands and return only a condensed result; the hand-off is where the paper locates the dominant, mostly silent failures. In the semantic design, the repository is chunked, embedded, and retrieved by a ReAct agent with three tools (structure lookup, retrieval, file reader), and its failures are high-effort retrieval or localization misses.

What would settle it

Re-run the same 720 questions with a deep-agent harness that shares the planner's context with the sub-agent, or that verifies the sub-agent's returned summary against the files it read; if the accuracy gap reverses or drops below the 3.8-percentage-point minimum observed across thresholds, the coordination-breakdown explanation is wrong. Alternatively, run a compute-matched version where Semantic gets the same token budget as Deep; if Semantic's advantage evaporates, the effect is budget, not paradigm.

Watch

Extended reading notes

Core claim

The central claim is that the delegation-based Deep Agentic search, now preferred in many command-line code assistants, is less accurate and more expensive than Semantic search for read-only repository question answering. Pooled across four models, Semantic passed 65.2% of SWE-QA questions against Deep's 46.2%, failed fewer (21.9% vs. 34.4%), and cost 0.32 dollars per correct answer against 0.74. The failure taxonomy attributes 41.8% of Deep's failures to coordination breakdown at the planner-to-sub-agent hand-off, a failure class a flat retrieval agent cannot have, and in 91% of those cases the agent still returned a fluent, confident, wrong answer. The paper reads this as evidence that the

Load-bearing premise

The two agent harnesses, used with their default tool contracts, instantiate the two paradigms representatively enough that the measured gap transfers to the commercial code assistants the paper discusses; a different sub-agent summarizer or step limit could change the size or sign of the gap.

Editorial extensions

If this is right

  • Code assistants that default to delegated exploration for read-only questions are sacrificing accuracy and money for context hygiene they may not need on single-shot questions.
  • Semantic search should be the default for stable, indexable repositories queried read-only, with delegated exploration reserved for regimes where an index is impractical.
  • The 41.8% coordination-breakdown share implies that improving hand-off fidelity, such as better sub-agent result summarization or verification of what the sub-agent actually returned, is the highest-leverage fix for deep agents.
  • The flat-to-negative relationship between effort and score suggests that adding more exploration or more tool calls will not close the gap; the fix has to be structural.
  • The threshold sweep and repository-level bootstrap indicate the ordering is robust to grading-cutoff choice, so the conclusion survives reasonable variations in evaluation.

Reading between the lines

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

  • The paper leaves open whether the balance shifts in long multi-task sessions, where an uncluttered orchestrator context has standalone value; the results do not rule out deep agentic search winning there.
  • A compute-matched or token-budget-matched comparison could separate the search-strategy effect from the cost of extra computation; the current design bundles them.
  • The failure taxonomy suggests a concrete testable extension: a hybrid router that starts with retrieval and only delegates exploration when retrieval's confidence is low might dominate both arms.
  • The silent-failure finding warns that user-visible confidence is not a reliability signal in deep agentic systems; any deployment should verify answers against cited code.
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

3 major / 5 minor

Summary. The paper empirically compares two repository-level code question answering architectures: Semantic Search, implemented as a LangChain ReAct agent with a pre-built vector index and the three SWE-QA benchmark tools, and Deep Agentic Search, implemented with the langchain-deep-agents harness using its default toolset (planning, virtual filesystem, sub-agent delegation, terminal-style search, and file reading). On 15 Python repositories and 720 questions per condition, across four models (Gemini 2.5 Flash, Gemini 2.5 Pro, Gemini 3 Flash, Qwen3-235B), the paper reports that Semantic achieves a higher Pass rate for every model (pooled 65.2% vs 46.2%), lower token and dollar cost, and a lower cost per correct answer. A failure taxonomy of all 1,621 failing runs attributes the largest share of Deep failures (41.8%) to planner-to-sub-agent coordination breakdowns, usually silent, and a further 13.5% to non-terminating loops. The paper interprets the result as showing that, for single-turn, read-only questions over an indexable repository, index-backed retrieval is the stronger and cheaper option, while carefully noting boundary conditions where Deep Agentic search remains necessary (non-indexable or rapidly changing repositories, long multi-task sessions).

Significance. If the measured comparison is taken at face value, this is a useful and timely empirical result. The study's internal validity is strong: all comparisons are paired within model on the same 720 questions; significance is assessed with exact McNemar and Wilcoxon tests with FDR correction; effect sizes are reported with bootstrap confidence intervals; the Pass-threshold sensitivity is swept from 50 to 90; a repository-level bootstrap is used; and the LLM judge is validated against a blind human panel with substantial agreement (Cohen's kappa 0.74 on the verdict, r=0.83 on the score). The failure taxonomy is a genuine contribution, with inter-annotator agreement reported. The direction of the accuracy difference is robust across four models and two vendors, and the paper is appropriately cautious in its discussion of boundary conditions. Its principal weakness is external validity: the Deep arm is instantiated by a single harness with default tool contracts, and the paper explicitly states that no sensitivity analysis across alternative deep-agent harnesses was run. Because the paper's explanatory story centers on a single failure mechanism (coordination breakdown) whose prevalence is

major comments (3)
  1. [Section 8; Appendix A.2; Section 6.4] The central explanatory claim—that 41.8% of Deep failures are planner-to-sub-agent coordination breakdowns and that these are usually silent—is measured using one specific deep-agent harness, langchain-deep-agents with its default toolset. Section 8 explicitly concedes that 'other harnesses implement delegation differently,' that sub-agent summarization and step limits affect behavior, and that no sensitivity analysis across alternative deep-agent harnesses was run. Since this failure mechanism is the single largest contributor to the Deep arm's disadvantage, the magnitude of the accuracy gap is not paradigm-invariant: a different sub-agent report contract, a different recursion/step limit, or a richer summarization protocol could plausibly change the 41.8% figure and the cost gap. This is load-bearing for the abstract's and conclusion's statements about 'Deep Agentic search' as a genera
  2. [Section 4.1; Section 5.1; Section 7] The two compared conditions differ not only in the search strategy (indexed retrieval vs live terminal exploration) but also in orchestration complexity: the Deep arm adds planning, a task list, sub-agent delegation, and a larger tool surface. Section 4.1 correctly notes that equalizing tool sets would collapse the conditions, but this makes the result a comparison of two cohesive 'paradigm bundles' rather than a single-factor test of search strategy. The abstract's statement that 'retrieval was the stronger and cheaper option' should therefore be qualified throughout as a statement about the ReAct-retrieval bundle versus the delegation-based deep-agent bundle. Without such qualification, or without an additional flat-agentic arm to separate the effect of delegation from the effect of on-demand exploration, readers may over-attribute the gap to the retrieval operation rather than to the
  3. [Section 6.3; Table 5] The first reported chi-square for mechanism-by-architecture (chi-square=508.8, Cramer's V=0.56) includes a structural zero: 'coordination breakdown' is impossible in the Semantic arm by construction. The paper acknowledges this and recomputes over the five shared mechanisms (chi-square=147.2, V=0.35), but the initial statistic is still presented as the headline association. I recommend reporting the five-mechanism chi-square as the primary statistic and moving the full six-mechanism figure to a supplementary role, since the structural-zero version violates the usual assumptions of the chi-square test and adds little information beyond the definitional fact that a flat agent has no delegation hand-off.
minor comments (5)
  1. [Abstract; Section 5.1] The text says Semantic 'answered 65.2% of questions correctly' and later uses 'Pass rate.' A Pass verdict is a rubric threshold (score >=70), not necessarily a fully correct answer. Consider using 'received a Pass verdict' or 'achieved a Pass rate' to avoid overstating what the metric measures.
  2. [Figure 10 caption; Section 5.1] The Figure 10 caption says 'Deep matched or exceeded Semantic on only one of the fifteen repositories,' while the text says 'Deep exceeded Semantic on only one of the fifteen repositories, django, by 3.0 points.' Please reconcile: if the repository was only exceeded (not matched), the caption should say 'exceeded.'
  3. [Section 4; reproducibility] No artifact link, repository, or DOI is provided for the harness code, configuration files, or execution traces. Given the paper's emphasis on inspectable defaults and reproducibility, I recommend releasing the code and traces (or at least the exact pinned versions of langchain-deep-agents and the agent prompts) as a supplement.
  4. [Table 3; Section 5.2] The column header 'Mean$/q' should be typeset as 'Mean $/q' or 'Mean cost per question' for clarity. Also, consider adding a note that the cost figures are derived from the price sheet in Section 4.2 and are not vendor invoices.
  5. [Section 6.3] The sentence 'The association is very strong, with a chi-square of 508.8 on five degrees of freedom' should note that this statistic is descriptive only, given the structural-zero issue and the non-independence of runs within repositories. The repository-level bootstrap is mentioned for the accuracy comparison; a similar bootstrap for the failure-mechanism comparison would strengthen the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the accuracy and cost comparison is grounded in an external benchmark, an independent validated judge, and published prices; no result reduces to a fitted input or to a self-citation chain.

full rationale

The paper's central claim—Semantic search outperforms Deep Agentic search on SWE-QA—is an empirical measurement, not a derivation. Both paradigms are instantiated with public LangChain harnesses over an external benchmark (SWE-QA), graded by Claude Sonnet 4.6, a judge from a different provider/family than all answering models, and validated against an independent human panel (Cohen's kappa 0.74 on verdict, Pearson 0.83 on score). Costs are computed from measured token counts times published per-token prices. No parameter is fitted to the outcome, and no accuracy or cost quantity is defined in terms of the conclusion. The failure taxonomy's 'coordination breakdown' category is definitionally possible only for Deep, but the authors explicitly acknowledge this and recompute the association excluding that mechanism (chi-square 147.2, Cramer's V 0.35 over the five shared mechanisms), so the explanatory claim does not reduce to the definition. The paper's own acknowledged limitation—using one default harness per paradigm and not running a sensitivity analysis across alternative deep-agent harnesses—is a construct/external-validity threat, not circularity, because nothing in that choice makes the measured gap equivalent to the input. Self-citations (e.g., Oskooei et al. 2026 on context rot and many-shot prompting) appear only as background motivation and are not load-bearing for the measured result; the comparison would stand even if those citations were removed. No equation or fitted parameter makes the outcome equivalent to the experimental setup, so the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The study is empirical rather than derived, so the ledger lists configuration choices and background assumptions rather than fitted coefficients. No model parameter is fitted to the outcome, but the comparison depends on retrieval configuration, termination bounds, the benchmark rubric, the judge, and the representativeness of the harnesses.

free parameters (6)
  • Retrieval top-k = k=10
    Number of chunks returned per search_rag call (Appendix B.3); chosen by the authors and not swept, affecting accuracy and cost of the Semantic condition.
  • Chunk size and overlap = 500 tokens / 50 tokens
    RecursiveCharacterTextSplitter configuration (Appendix B.1); affects retrieval quality, context composition, and fragmentation.
  • Deep agent recursion limit = 80 steps
    Agent graph recursion limit (Appendix C.1); runs reaching it are kept as failures, so this bound shapes the Deep failure distribution.
  • Wall-clock limit = 900 seconds
    Termination bound (Appendix C.1); can convert slow successful runs into failures, possibly favoring whichever paradigm answers faster.
  • Pass threshold = 70/100
    Benchmark-defined verdict cutoff (Appendix D.2); the paper sweeps 50 to 90, so the ordering is robust, but the magnitude of the gap depends on the cutoff.
  • Price sheet = Per-model USD rates, e.g., Gemini 2.5 Flash 0.30/2.50, Gemini 3 Flash 0.50/3.00
    Monetary cost and cost-per-correct-answer results derive from this published sheet (Section 4.2); token counts are also reported because prices change.
assumptions (5)
  • domain assumption The LangChain ReAct harness and the LangChain Deep Agents harness faithfully instantiate Semantic search and Deep Agentic search as deployed in practice.
    Section 4.1 and Appendices A-B define the tool contracts; Section 8 notes no sensitivity analysis across alternative harnesses.
  • domain assumption The Claude Sonnet 4.6 judge is a reliable proxy for human judgement on all 720 questions per condition.
    Validated on a stratified 800-answer sample with Cohen's kappa 0.74 and r=0.83 (Table 2); agreement is lower on Deep answers (kappa 0.69), so unseen questions could deviate.
  • domain assumption SWE-QA and its rubric validly operationalize repository-level code question answering.
    External benchmark (Peng et al. 2026a), 15 Python repositories; single benchmark limits external validity (Section 8).
  • domain assumption Training-data memorization does not create an asymmetric advantage for either paradigm.
    Section 8: paired design makes memorization raise both arms; the SWE-bench-Live subset check is descriptive only.
  • domain assumption Automatic coding of failure traces by an LLM approximates human coding.
    Section 6.1: validated on 400 failures with kappa 0.82 between consensus and automatic coding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Agentic Search for Repository-Level Code Question Answering: An Empirical Study." pith.science (2026). https://pith.science/paper/ZVE64JCL

@misc{pith2026260801507,
  author       = {Pith},
  title        = {Pith review of: Deep Agentic Search for Repository-Level Code Question Answering: An Empirical Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZVE64JCL}},
  note         = {Machine review of arXiv:2608.01507}
}
read the original abstract

Code agents spend much of their effort simply locating the right code inside a repository. Two approaches dominate current practice. In Semantic Search, the agent retrieves code blocks from a vector index built from the repository in advance. In Deep Agentic Search (also known as grep-search by subagent), a planning agent delegates the exploration to a separate subagent that works in an isolated context window and returns only a condensed result. The second design, which is considered good context engineering practice, exists to protect the main agent from context pollution (also known as context rot), the loss of accuracy that occurs as unrelated material accumulates in the context window. Recent code agents (such as Claude Code, Codex, Antigravity, etc) have adopted it quickly, but there is little evidence on whether it produces better answers. We compare the two approaches on SWE-QA, a benchmark for repository-level code question answering. Semantic search answered 65.2% of questions correctly against 46.2% for deep agentic search, and it produced each correct answer at less than half the cost. To explain the gap, we then coded every failed run into a taxonomy of failure modes. The taxonomy shows that deep agentic search did not remove failures but introduced a new class of them: the single largest share of its failures, 41.8%, occurred at the hand-off between the planner and its sub-agent, and these were usually silent, ending in a fluent and confident answer that was wrong. Deep agentic search addresses a real problem and is now the preferred design in many code agents. However, our results show that the protection it offers may not be free, and that for read-only questions over a repository that can be indexed, retrieval was the stronger and cheaper option.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 49 canonical work pages

  1. [1]

    SWE - QA : Can Language Models Answer Repository-level Code Questions?

    Peng, Weihan and Shi, Yuling and Wang, Yuhang and Zhang, Xinyun and Shen, Beijun and Gu, Xiaodong. SWE - QA : Can Language Models Answer Repository-level Code Questions?. Findings of the A ssociation for C omputational L inguistics: ACL 2026. 2026. doi:10.18653/v1/2026.findings-acl.402

    work page doi:10.18653/v1/2026.findings-acl.402 2026

  2. [2]

    Peng, Wang and others , title =

    work page

  3. [3]

    Richard and Koch, Gary G

    Landis, J. Richard and Koch, Gary G. , title =. Biometrics , volume =. 1977 , doi =

    1977

  4. [4]

    2025 , eprint =

    Jayant Havare and Saurav Chaudhary and Ganesh Ramakrishnan and Kaushik Maharajan and Srikanth Tamilselvam , title =. 2025 , eprint =

    work page 2025

  5. [5]

    2026 , eprint =

    Shubham Ugare and Satish Chandra , title =. 2026 , eprint =

    work page 2026

  6. [6]

    2026 , eprint =

    Deli Huang and Cunguang Wang and Hongyin Tang and Zhe Tang and Linsen Guo and Dongyu Ru and Ruoshi Yuan and Ziyue Zhu and Xiaoyu Li and Ziwen Wang and Chen Zhang and Anchun Gui and Wen Zan and Jiaqi Zhang and Xuezhi Cao and Jingang Wang and Xunliang Cai and Yixin Cao , title =. 2026 , eprint =

    work page 2026

  7. [7]

    2026 , eprint =

    Yoseph Berhanu Alebachew and Hunter Leary and Swanand Vaishampayan and Chris Brown , title =. 2026 , eprint =

    work page 2026

  8. [8]

    2025 , eprint =

    Liang Wang and Haonan Chen and Nan Yang and Xiaolong Huang and Zhicheng Dou and Furu Wei , title =. 2025 , eprint =

    work page 2025

Show all 57 references
  1. [9]

    2026 , eprint =

    Ankit Gupta and Aditya Prasad and Rameswar Panda , title =. 2026 , eprint =

    2026

  2. [10]

    2026 , eprint =

    Yuling Shi and Chaoxiang Xie and Zhensu Sun and Yeheng Chen and Chenxu Zhang and Longfei Yun and Chengcheng Wan and Hongyu Zhang and David Lo and Xiaodong Gu , title =. 2026 , eprint =

    2026

  3. [11]

    2026 , eprint =

    Jun Zhang and JianYing Qu and Hanwen Du and Zhongkai Sun and Yehua Yang and Qiao Zhao , title =. 2026 , eprint =

    2026

  4. [12]

    Ahmed and M

    M. Ahmed and M. Dorrah and A. Ashraf and Y. Adel and A. Elatrozy and B. E. Mohamed and W. Gomaa , title =. 2024 6th Novel Intelligent and Leading Emerging Sciences Conference (NILES) , pages =. 2024 , doi =

    2024

  5. [13]

    Companion Proceedings of the ACM Web Conference 2025 (WWW '25 Companion) , year =

    Ruida Hu and Chao Peng and Jingyi Ren and Bo Jiang and Xiangxin Meng and Qinyun Wu and Pengfei Gao and Xinchen Wang and Cuiyun Gao , title =. Companion Proceedings of the ACM Web Conference 2025 (WWW '25 Companion) , year =. 2412.14764 , archivePrefix =

    2025 arXiv

  6. [14]

    2025 , eprint =

    Jian Yang and Wei Zhang and Yizhi Li and Shawn Guo and Haowen Wang and Aishan Liu and Ge Zhang and Zili Wang and Zhoujun Li and Xianglong Liu and Weifeng Lv , title =. 2025 , eprint =

    2025

  7. [15]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Xiangyang Li and Kuicai Dong and Yi Quan Lee and Wei Xia and Hao Zhang and Xinyi Dai and Yasheng Wang and Ruiming Tang , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2025 , doi =

    2025

  8. [16]

    Lyu , title =

    Shuzheng Gao and Chaozheng Wang and Shuqing Li and Yun Peng and Michael R. Lyu , title =. Proceedings of the Third International Workshop on Large Language Models for Code (LLM4Code '26) , year =

  9. [17]

    Carreon-Rascon and Xuanzhao Dong and Feng Luo , title =

    Jingjing Wang and Xiwen Chen and Wenhui Zhu and Huayu Li and Zhengxiao He and Feiyang Cai and Ana S. Carreon-Rascon and Xuanzhao Dong and Feng Luo , title =. 2026 , eprint =

    2026

  10. [18]

    YunSeok Choi and CheolWon Na and Jee-Hyong Lee , title =. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages =. 2025 , doi =

    2025

  11. [19]

    2025 , eprint =

    Jialiang Chen and Kaifa Zhao and Jie Liu and Chao Peng and Jierui Liu and Hang Zhu and Pengfei Gao and Ping Yang and Shuiguang Deng , title =. 2025 , eprint =

    2025

  12. [20]

    IEEE Transactions on Software Engineering , year =

    Jing Gong and Yanghui Wu and Linxi Liang and Yanlin Wang and Jiachi Chen and Mingwei Liu and Zibin Zheng , title =. IEEE Transactions on Software Engineering , year =. 2406.11589 , archivePrefix =

  13. [21]

    Proceedings of the ACM Web Conference 2026 (WWW '26) , year =

    Xiaoxi Li and Wenxiang Jiao and Jiarui Jin and Guanting Dong and Jiajie Jin and Yinuo Wang and Hao Wang and Yutao Zhu and Ji-Rong Wen and Yuan Lu and Zhicheng Dou , title =. Proceedings of the ACM Web Conference 2026 (WWW '26) , year =. 2510.21618 , archivePrefix =

    2026

  14. [22]

    2025 , eprint =

    Guochang Li and Yuchen Liu and Zhen Qin and Yunkun Wang and Jianping Zhong and Chen Zhi and Binhua Li and Fei Huang and Yongbin Li and Shuiguang Deng , title =. 2025 , eprint =

    2025

  15. [23]

    2026 , eprint =

    Xinchen Wang and Ruida Hu and Cuiyun Gao and Pengfei Gao and Chao Peng , title =. 2026 , eprint =

    2026

  16. [24]

    Lopes and Hil

    Eduardo Dos S. Lopes and Hil. Exploring. 2025 LI Latin American Computer Conference (CLEI) , year =

    2025

  17. [25]

    2026 , eprint =

    Zhonghang Li and Zongwei Li and Yuxuan Chen and Han Shi and Jiawei Li and Jierun Chen and Haoli Bai and Chao Huang , title =. 2026 , eprint =

    2026

  18. [26]

    2026 , eprint =

    Shaoqiu Zhang and Maoquan Wang and Yuling Shi and Yuhang Wang and Xiaodong Gu and Yongqiang Yao and Tori Gong and Sheng Chen and Rao Fu and Anisha Agarwal and Spandan Garg and Gabriel Ryan and Colin Merkel and Yufan Huang and Shengyu Fu , title =. 2026 , eprint =

    2026

  19. [27]

    2026 , eprint =

    Sahil Sen and Akhil Kasturi and Elias Lumer and Anmol Gulati and Vamse Kumar Subbiah , title =. 2026 , eprint =

    2026

  20. [28]

    2026 , eprint =

    Yuntong Hu and Tongli Su and Liang Zhao and Bowen Zhu and Hasibul Haque , title =. 2026 , eprint =

    2026

  21. [29]

    Analysis of Images, Social Networks and Texts (AIST 2024) , publisher =

    Georgy Andryushchenko and Vladimir Ivanov and Vladimir Makharev and Elizaveta Tukhtina and Aidar Valeev , title =. Analysis of Images, Social Networks and Texts (AIST 2024) , publisher =. 2024 , doi =

    2024

  22. [30]

    2026 , eprint =

    Niko Usai and Dario Montagnini and Kristian Ilianov Iliev and Raffaele Camanzo , title =. 2026 , eprint =

    2026

  23. [31]

    2025 , eprint =

    Stefano Rando and Luca Romani and Alessio Sampieri and Yuta Kyuragi and Luca Franco and Fabio Galasso and Tatsunori Hashimoto and John Yang , title =. 2025 , eprint =

    2025

  24. [32]

    Proceedings of the IEEE/ACM International Conference on Automated Software Engineering (ASE) , year =

    Yuling Shi and Yichun Qian and Hongyu Zhang and Beijun Shen and Xiaodong Gu , title =. Proceedings of the IEEE/ACM International Conference on Automated Software Engineering (ASE) , year =

  25. [33]

    2026 , eprint =

    Jiazhan Feng and Zhan Qin and Cuiyun Gao and Ruiqi Wang and Chaozheng Wang and Yingwei Ma and Xiaoyuan Xie , title =. 2026 , eprint =

    2026

  26. [34]

    2025 , eprint =

    Pratik Shah and Rajat Ghosh and Aryan Singhal and Debojyoti Dutta , title =. 2025 , eprint =

    2025

  27. [35]

    International Conference on Learning Representations (ICLR) , year =

    Shunyu Yao and Jeffrey Zhao and Dian Yu and Nan Du and Izhak Shafran and Karthik Narasimhan and Yuan Cao , title =. International Conference on Learning Representations (ICLR) , year =. 2210.03629 , archivePrefix =

    arXiv

  28. [36]

    2025 , eprint =

    Shadikur Rahman and Aroosa Hameed and Gautam Srivastava and Syed Muhammad Danish , title =. 2025 , eprint =

    2025

  29. [37]

    2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA) , publisher =

    Mehmet Oguz Kocadere and Yusuf Emir Comert and Sudenaz Yazici and Samil Giray Karacay and Ibrahim Baran Yildiz and Tugba Gurgen Erdogan , title =. 2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA) , publisher =. 2026 , doi =

    2026

  30. [38]

    2023 , eprint =

    Yunfan Gao and Yun Xiong and Xinyu Gao and Kangxiang Jia and Jinliu Pan and Yuxi Bi and Yi Dai and Jiawei Sun and Meng Wang and Haofen Wang , title =. 2023 , eprint =

    2023

  31. [39]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2005.11401 , archivePrefix =

    2005 arXiv

  32. [40]

    2026 , eprint =

    Kishan Maharaj and Nandakishore Menon and Ashita Saxena and Srikanth Tamilselvam , title =. 2026 , eprint =

    2026

  33. [41]

    Proceedings of the 3rd International Workshop on Large Language Models for Code (LLM4Code '26) , year =

    Ferida Mohammed and Fatma Ayad and Petros Maniatis and Satish Chandra and Elizabeth Dinella , title =. Proceedings of the 3rd International Workshop on Large Language Models for Code (LLM4Code '26) , year =. 2601.16456 , archivePrefix =

    arXiv

  34. [42]

    Findings of the Association for Computational Linguistics: ACL 2026 , year =

    Jing Zhang and Lianghong Guo and Yanlin Wang and Mingwei Liu and Jiachi Chen and Yuchi Ma and Ensheng Shi and Terry Yue Zhuo and Hongyu Zhang and Zibin Zheng , title =. Findings of the Association for Computational Linguistics: ACL 2026 , year =. 2512.08867 , archivePrefix =

    2026

  35. [43]

    2026 , eprint =

    Mohit Raghavendra and Soham Dan and Miguel Romero Calvo and Yannis Yiming He and Johannes Baptist Mols and Gautam Anand and Cole McCollum and Edgar Arakelyan and Vijay Bharadwaj and Andrew Park and Jeff Da and MohammadHossein Rezaei and Bing Liu and Brad Kenstler and Yunzhong ...

    2026

  36. [44]

    2026 , eprint =

    Shaoqiu Zhang and Yuhang Wang and Jialiang Liang and Yuling Shi and Wenhao Zeng and Maoquan Wang and Shilin He and Ningyuan Xu and Siyu Ye and Kai Cai and Xiaodong Gu , title =. 2026 , eprint =

    2026

  37. [45]

    2026 , eprint =

    Yuhang Wang and Yuling Shi and Mo Yang and Rongrui Zhang and Shilin He and Heng Lian and Yuting Chen and Siyu Ye and Kai Cai and Xiaodong Gu , title =. 2026 , eprint =

    2026

  38. [46]

    Annual Meeting of the Association for Computational Linguistics (ACL 2026) , year =

    Songcheng Cai and Zhiheng Lyu and Yuansheng Ni and Xiangchao Chen and Baichuan Zhou and Shenzhe Zhu and Yi Lu and Haozhe Wang and Chi Ruan and Benjamin Schneider and Weixu Zhang and Xiang Li and Andy Zheng and Yuyu Zhang and Ping Nie and Wenhu Chen , title =. Annual Meeting of...

    2026

  39. [47]

    International Conference on Learning Representations (ICLR) , year =

    Lingjie Jiang and Shaohan Huang and Xun Wu and Yixia Li and Dongdong Zhang and Furu Wei , title =. International Conference on Learning Representations (ICLR) , year =. 2508.09945 , archivePrefix =

    arXiv

  40. [48]

    2025 , howpublished =

    Effective Context Engineering for. 2025 , howpublished =

    2025

  41. [49]

    2025 , howpublished =

    Context Engineering for. 2025 , howpublished =

    2025

  42. [50]

    Don't Build Multi-Agents , year =

  43. [51]

    Context Engineering for Agents , year =

  44. [52]

    , title =

    Oskooei, Amirkia Rafiei and Yukcu, Selcan and Bozoglan, Mehmet Cevheri and Aktas, Mehmet S. , title =. Computational Science and Its Applications,. 2025 , publisher =

    2025

  45. [53]

    Selcan and Bozoglan, Mehmet Cevheri and Aktas, Mehmet S

    Rafiei Oskooei, Amirkia and Yukcu, S. Selcan and Bozoglan, Mehmet Cevheri and Aktas, Mehmet S. , title =. Proceedings of the 3rd International Workshop on Large Language Models for Code (LLM4Code) , pages =

  46. [54]

    , title =

    Oskooei, Amirkia Rafiei and Cosdan, Kaan Baturalp and Isiktas, Husamettin and Aktas, Mehmet S. , title =. Proceedings of the 1st Workshop on Code Translation, Transformation, and Modernization (ReCode '26) , pages =. 2026 , publisher =

    2026

  47. [55]

    2026 , eprint =

    Liu, Jiacheng and Zhao, Xiaohan and Shang, Xinyi and Shen, Zhiqiang , title =. 2026 , eprint =

    2026

  48. [56]

    Decoding the Configuration of

    dos Santos, H. Decoding the Configuration of. Proceedings of the 2026 International Workshop on Agentic Engineering (AGENT@ICSE '26) , pages =. 2026 , publisher =

    2026

  49. [57]

    and Nyst, Engel and Malhotra, Rohit and Zhou, Xuhui and Chen, Valerie and Brennan, Robert and Neubig, Graham , title =

    Wang, Xingyao and Rosenberg, Simon and Michelini, Juan and Smith, Calvin and Tran, Hoang H. and Nyst, Engel and Malhotra, Rohit and Zhou, Xuhui and Chen, Valerie and Brennan, Robert and Neubig, Graham , title =. Proceedings of Machine Learning and Systems 8 (MLSys 2026) , year =

    2026

Pith tools

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