Pith. sign in

REVIEW 4 major objections 5 minor 42 references

RARG turns relevance scores into an execution order for grep-based corpus search, improving accuracy per tool call over retrieval and direct-corpus-interaction agents.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 03:27 UTC pith:43LAMBLR

load-bearing objection RARG is a real, useful idea—relevance as an execution order for grep—but the headline gains over RISE/DCI rest on a confounded baseline setup and a 100-question sample, so the numbers need re-verification before I'd treat them as solid. the 4 major comments →

arxiv 2607.24223 v2 pith:43LAMBLR submitted 2026-07-27 cs.CL

A New Role for Relevance: Guiding Corpus Interaction in Agentic Search

Jiangnan Li , Yuqing Li , Mo Yu , Jinchao Zhang , Jie Zhou This is my paper
classification cs.CL
keywords relevanceinteractioncorpussearchrargagentscluesconvergence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

LLM agents that search large text corpora usually have two options. One is to retrieve a short ranked list of documents with an embedding model; this is fast but can miss a decisive detail buried in a long document. The other is to let the agent run grep-like searches across raw files; this is flexible but treats every file as equally promising, wasting turns on noise. RARG combines both: it first ranks the corpus with an embedding model, then makes the ripgrep tool scan files in that rank order, so the most promising files are searched first. On top of this, RARG+ seeds the agent with a few query-relevant paragraphs as a starting hint, and RARG++ reranks the many grep matches so the model only sees the most informative excerpts.

On a 100-question subset of BrowseComp-Plus, RARG++ reports 84% accuracy (vs 78% for RISE and DCI) while using roughly 24 tool calls instead of 29 and 99. On the BRIGHT retrieval benchmark, RARG+ reaches 53.36 average nDCG@10, narrowly ahead of the NeMo agent at 52.89. The paper also tests scaling to a 1M-document corpus and shows RARG++ keeping a 10-point margin over a BM25-based baseline.

The main caveats are that the QA benchmark is only 100 questions, no error bars or significance tests are reported, and the headline RISE and DCI numbers in the main table are cited from a prior paper under possibly different context-management settings. The ablations and the qualitative case study are consistent with the proposed mechanism, but the strength of the efficiency comparison is limited by these issues.

Core claim

"By turning relevance into an execution prior over rg exploration, RARG advances the accuracy–efficiency frontier over retrieval-based and direct-interaction agents" (Figure 1 caption). Concretely, the paper reports RARG++ at 84% accuracy with 23.9 tool calls on the 100-query BrowseComp-Plus sample vs 78%/28.7 for RISE and 78%/99.1 for DCI (GPT-5.4-mini), and RARG+ at 53.36 avg nDCG@10 on BRIGHT vs NeMo's 52.89. These specific numbers are the load-bearing evidence for the claim.

Load-bearing premise

The accuracy–efficiency comparison rests on the 100-query RISE subset being representative of BrowseComp-Plus and on cited RISE/DCI numbers being directly comparable to RARG's, even though RARG uses a different context-compaction schedule (last-40 tool results, 230K threshold; Section 4.1) than the RISE harness that generated the baseline numbers, and no variance is reported. If the baseline numbers are not re-run under identical settings, the tool-cost and accuracy differences may be confounded.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces RARG, an agentic search method that combines DCI-style grep interaction with embedding-based relevance at three levels: document-order traversal (RARG), entry-point paragraph initialization (RARG+), and match-level reranking (RARG++). It evaluates on a 100-query BrowseComp-Plus sample and four BRIGHT subsets, reporting higher accuracy and/or fewer tool calls than RISE, DCI, and NeMo. The paper argues that relevance should be an execution prior for corpus interaction rather than merely a content filter, and it presents analyses of scope recall, rg coverage, command composition, and a qualitative case study.

Significance. If the comparisons hold, the contribution is timely and useful: it identifies a clear design gap in agentic search, the method is simple and reproducible, code is released, and the analyses go beyond a single benchmark. The two-resolution relevance guidance (document-level traversal order and match-level visibility) is a clean idea with potential impact on practical search agents. However, the empirical support is currently not strong enough: the main baseline comparison is confounded by harness differences, the sample sizes are small with no uncertainty quantification, and the cost metric is not defined. These are load-bearing issues that can be fixed by re-running baselines under identical conditions and reporting uncertainty, but they need to be addressed before the central claim can be accepted.

major comments (4)
  1. [Section 4.1; Table 1] Table 1 marks RISE, RISE-BM25, Retrieval-Agent, and DCI with '△ denotes the results are cited from RISE [35]', but Section 4.1 says 'We run both DCI-Agent-Lite and the RISE-based baselines using the official RISE implementation.' This is internally inconsistent. If the baseline numbers are cited, they were generated under RISE's context-compaction schedule, not RARG's. Section 4.1 states that RARG keeps the most recent 40 tool results and raises the compaction threshold to 230K, whereas DCI compaction keeps 12 turns. Context compaction affects how much evidence the model sees before older results are cleared and how often it must re-search; it can therefore change both accuracy and tool counts. The 84 vs 78 accuracy gap and 23.9 vs 28.7 tool-call gap could in part be driven by this scheduling difference. Please re-run all Δ-marked baselines under the identical harness, or justify that th
  2. [Table 1; Table 3] All headline QA results are on 100 queries (Section 4.1). With n=100, a 78% vs 84% difference has a standard error of roughly 4 percentage points; the difference is within a 95% confidence interval around either estimate. No confidence intervals, bootstrap estimates, or significance tests are reported anywhere in Table 1. On BRIGHT (Table 3), RARG+ exceeds NeMo by 0.47 nDCG@10 (53.36 vs 52.89), and the per-domain columns show overlapping values; no variance is reported. The paper should report uncertainty (e.g., bootstrap CIs or per-query breakdowns) and should avoid claiming an accuracy-efficiency frontier on the current sample sizes.
  3. [Figure 1; Section 4.1] Figure 1's x-axis is labeled 'Cost ($/query)' with values in dollars, but the experiments section only reports average tool calls (Table 1). The claimed '−46% cost' is not derivable from the paper: RARG++ vs RISE on GPT-5.4-mini is 23.9 vs 28.7 tool calls (−16.7%), and on GPT-5.4 it is 25.43 vs 34.30 (−25.9%). Moreover, tool-call counts are not a uniform cost unit, since RARG++ invokes embedding-based reranking on rg output (M up to 500; Section 3.4) and RARG's document ranking and paragraph splitting also consume compute, while DCI's Bash calls are plain grep. Please specify a token/dollar cost model, or relabel the efficiency axis as tool calls and correct the percentages.
  4. [Section 3.2; Appendix B.1] Section 3.2 says embed_recall 'returns only the mapping' from scope file to query, and RARG+ is defined by appending top-10 query-relevant paragraphs (Section 3.3). However, the system prompt in Appendix B.1 (Fig. 5) states that embed_recall 'also returns a few reference paragraphs from top-ranked documents.' If that prompt is used for plain RARG, then the RARG vs RARG+ comparison (80 vs 81 in Table 1, and the case study's T7→T2 shift) is confounded by whether the initialization is already present; if it is only used for RARG+, the presentation is misleading. Please clarify which prompt is used for each variant and re-check the ablation.
minor comments (5)
  1. [Table 3] The row label 'Nemo Agent52.89' is missing a space; it should read 'Nemo Agent' with the value in the Avg. column.
  2. [Figure 4(a)] The error bars are described as standard deviation across RARG, RARG+, and RARG++ (three values), not across queries or repeated runs. This does not quantify sampling uncertainty and should be relabeled to avoid confusion.
  3. [Figure 3] The rows preceding the heatmap (1–500, 501–1,000, ...) and the numeric rows below are hard to parse. Clarify what the row values represent and how the rank bins are defined.
  4. [References] Reference [6] is incomplete: 'Reza Esfandiarpoor and NVIDIA. Nvidia nemo retriever's agentic retrieval pipeline.' lacks year, venue, and version; provide a full citation or URL.
  5. [Section 4.5] The phrase 'which explains the claim we stated in Sec. 4.3' should be rephrased as 'which supports the claim' or 'is consistent with the claim'.

Circularity Check

0 steps flagged

No circularity: RARG's central claims are empirical evaluations on external benchmarks; no result is forced by construction or by a self-citation chain.

full rationale

The paper does not claim a formal derivation from first principles; it proposes an agent design and tests it on external benchmarks (BrowseComp-Plus via RISE's 100-query sample and BRIGHT) against baselines from other groups. The embedding models (Qwen3-Embedding-4B, llama-nv-embed-reasoning-3b) are off-the-shelf, and the scope cap, compaction thresholds, match caps, and reranking pool sizes are fixed design choices rather than parameters fitted to the reported accuracy/nDCG values. The coarse-to-fine variants (RARG/RARG+/RARG++) are compared with each other and with external baselines; the target metrics are not used as their own training signals. Self-citations [3,5,14,15,27,29,34] appear only in related work or as pointers to background ideas (e.g., query-relevant paragraphs for entry-point initialization); none is invoked as a uniqueness theorem or as the evidence for the main empirical claim. The paper's own Limitations section acknowledges dependence on embedding quality, instruction following, and corpus-induced interference — these are assumptions, not hidden circularities. The main weakness is methodological: Table 1 marks RISE/DCI numbers with '△ denotes the results are cited from RISE [35]' while Section 4.1 states that RISE-based baselines were run using the official RISE implementation, and no variance is reported on the 100-query sample. That is a comparability/robustness concern about confounded baselines, not a reduction of the claimed result to its inputs. Therefore no circular step can be exhibited.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

This is an empirical systems paper; no new theoretical entities are introduced. The main additional commitments are the hand-set hyperparameters listed above and the benchmark/protocol assumptions the evaluation rests on. No circular derivation is present.

free parameters (6)
  • scope cap (top-10,000 document paths) = 10000
    embed_recall writes at most 10,000 paths into a scope file; evidence beyond rank 10,000 is unreachable. Chosen for cost, not derived.
  • reranking pool size M and top-m shown = M=500, m=30/60 for BC+/BRIGHT
    Match-level reranking considers up to M candidates and shows top m; hand-set to fit the LLM observation budget.
  • rg match cap and match truncation = 30/60 matches, 1000/500 chars per match
    Controls the amount of raw grep output that reaches the model; a design choice that affects both accuracy and cost.
  • context compaction schedule = keep last 40 tool results, compaction threshold 230K
    Differs from DCI's default; chosen to 'better exploit the cache'. This is a potential confound in the tool-count comparison.
  • entry-point initialization (top paragraphs) = top-10 paragraphs, each 400-1000 chars
    RARG+ appends these to embed_recall output as a starting hint; hand-set in the implementation.
  • embedding model selection per benchmark and level = Q3E for BC+, NV for BRIGHT document ranking, Q3E for BRIGHT reranking
    The method relies on a strong embedding model; the choice is a design decision and depends on model strengths.
axioms (5)
  • domain assumption Embedding similarity is a valid relevance prior for ordering rg traversal
    Core mechanism in Section 3.2: if embedding scores do not correlate with evidence utility, ordering files by them will not expose clues earlier.
  • domain assumption The 100-query RISE subset is representative of BrowseComp-Plus
    All headline BC+ accuracy numbers come from a 100-query sample (Section 4.1); no confidence intervals are reported.
  • domain assumption LLM agent reliably follows the scope-based rg protocol
    The method instructs the model to use 'cat scope.txt | xargs rg' and to avoid unscopped searches; the authors acknowledge instruction-following sensitivity in Limitations.
  • domain assumption rg with -j1 preserves the input order of paths
    The -j1 flag is added to force sequential scanning so document relevance determines the order of matches; this relies on ripgrep's behavior.
  • domain assumption LLM-as-judge (GPT-5.1) produces accurate binary answers for accuracy scoring
    Accuracy on BC+ is computed with an LLM judge (Section 4.1); if the judge is biased, the reported accuracies are skewed.

pith-pipeline@v1.3.0-alltime-deepseek · 17954 in / 13765 out tokens · 195336 ms · 2026-08-04T03:27:43.943508+00:00 · methodology

0 comments
read the original abstract

Relevance is a query-dependent estimate of whether a document or excerpt contains useful evidence. Existing retrieval agents use relevance to select top-$k$ content, but document relevance alone cannot localize, compose, or verify the evidence required by complex questions. Direct Corpus Interaction (DCI) enables such fine-grained operations through grep-style exploration, but its relevance-agnostic search can expose useful clues late and delay convergence. Recent advances use relevance to narrow the corpus into a working space for interaction. Once interaction begins, however, relevance still does not directly guide which documents grep searches first or distinguish informative excerpts from a broad set of matches to let LLMs see them first. We introduce the Relevance-Aware RipGrep Search Agent (RARG), which turns relevance into an execution prior for corpus interaction. RARG provides coarse-to-fine relevance guidance: it orders documents for sequential 'ripgrep' traversal to expose globally relevant clues earlier, initializes promising entry points with query-relevant paragraphs, and reranks grep matches to surface informative excerpts that document-level ranking may otherwise obscure. Across challenging browse question answering and reasoning-intensive retrieval, RARG improves the accuracy--efficiency frontier over retrieval-based and direct-interaction agents. These results demonstrate that relevance-aware interaction enables faster and more reliable search convergence.

Figures

Figures reproduced from arXiv: 2607.24223 by Jiangnan Li, Jie Zhou, Jinchao Zhang, Mo Yu, Yuqing Li.

Figure 1
Figure 1. Figure 1: Accuracy/nDCG@10 versus interaction cost (average tool calls) on BrowseComp-Plus [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of RARG. embed_recall ranks the corpus into a scope file, and rg scans it in ranked order so relevant documents surface first (RARG). RARG+ adds query-relevant paragraphs as an entry point; RARG++ reranks rg matches to keep informative excerpts from lower-ranked documents visible. Document-level relevance sets where rg searches first; match-level relevance sets which matches reach the model. agent… view at source ↗
Figure 3
Figure 3. Figure 3: Relevant-document hits over scope ranks on BC+. Embedding hits are top-concentrated but [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scope quality and Bash usage under RARG on BrowseComp-Plus. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The system prompt used by RARG for the BrowseComp-Plus evaluation. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: The system prompt used by RARG for the BrowseComp-Plus evaluation. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The system prompt used by RARG for the BRIGHT retrieval evaluation. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: The system prompt used by RARG for the BRIGHT retrieval evaluation. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Compressed, turn-aligned trajectories for BC+ query 229 on the 100K corpus. The dot [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 7
Figure 7. Figure 7: Compressed, turn-aligned trajectories for BC+ query 229 on the 100K corpus. The dot [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

Review history (2 revisions) →

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

42 extracted references · 20 linked inside Pith

  1. [1]

    Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie, Kai Zou, Andrew Liu, Joshua Green, Kshama Patel, Ruoxi Meng, Mingyi Su, et al. 2025. Browsecomp-plus: A more fair and transparent evaluation benchmark of deep-research agent.arXiv preprint arXiv:2508.06600

    Pith/arXiv arXiv 2025

  2. [2]

    Tsz Ting Chung, Leyang Cui, Lemao Liu, Xinting Huang, Shuming Shi, and Dit-Yan Yeung

  3. [3]

    Tsz Ting Chung, Lemao Liu, Mo Yu, and Dit-Yan Yeung. 2025. Divlogiceval: A framework for benchmarking logical reasoning evaluation in large language models.Findings of the Association for Computational Linguistics: EMNLP, pages 901–915

    2025

  4. [4]

    Tsz Ting Chung, Lemao Liu, Mo Yu, and Dit-Yan Yeung. 2026. Many-shot cot-icl: Making in-context learning truly learn.arXiv preprint arXiv:2605.13511. 11

    Pith/arXiv arXiv 2026

  5. [5]

    Guoxuan Ding, Yuqing Li, Ziyan Zhou, Zheng Lin, Daren Zha, and Jiangnan Li. 2026. Exdr: Explanation-driven dynamic retrieval enhancement for multimodal fake news detection.arXiv preprint arXiv:2601.15820

    arXiv 2026

  6. [6]

    Nvidia nemo retriever’s agentic retrieval pipeline

    Reza Esfandiarpoor and NVIDIA. Nvidia nemo retriever’s agentic retrieval pipeline

  7. [7]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997

    Pith/arXiv arXiv 2023

  8. [8]

    Tz-Huan Hsu, Jheng-Hong Yang, and Jimmy Lin. 2026. Rethinking agentic search with pi-serini: Is lexical retrieval sufficient?arXiv preprint arXiv:2605.10848

    Pith/arXiv arXiv 2026

  9. [9]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516

    Pith/arXiv arXiv 2025

  10. [10]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. InProceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 6769–6781

    2020

  11. [11]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval- augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474

    2020

  12. [12]

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. 2025. Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366

    Pith/arXiv arXiv 2025

  13. [13]

    Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji-Rong Wen, Yutao Zhu, and Zhicheng Dou. 2025. Webthinker: Empowering large reasoning models with deep research capability.arXiv preprint arXiv:2504.21776

    Pith/arXiv arXiv 2025

  14. [14]

    Yuqing Li, Jiangnan Li, Zheng Lin, Ziyan Zhou, Junjie Wu, Weiping Wang, Jie Zhou, and Mo Yu. 2025. Mindscape-aware retrieval augmented generation for improved long context understanding.arXiv preprint arXiv:2512.17220

    Pith/arXiv arXiv 2025

  15. [15]

    Yuqing Li, Jiangnan Li, Mo Yu, Zheng Lin, Weiping Wang, and Jie Zhou. 2026. Mia- signature: Approximating global activation for long-context understanding.arXiv preprint arXiv:2605.06416

    Pith/arXiv arXiv 2026

  16. [16]

    Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, Yuyang Bai, Shangbin Feng, Hangxiao Zhu, Ming Zhong, Yuyu Zhang, Jianwen Xie, Yejin Choi, James Zou, Jiawei Han, Wenhu Chen, Jimmy Lin, Dongfu Jiang, and Yu Zhang. 2026. Beyond semantic sim- ilarity: Rethinking retrieval for agentic search via direct corpus interaction.arXiv preprint arXiv:2605.05242

    Pith/arXiv arXiv 2026

  17. [17]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts.Transac- tions of the association for computational linguistics, 12:157–173

    2024

  18. [18]

    Yi Lu, Zhuofeng Li, Ping Nie, Haoxiang Zhang, Yuyu Zhang, Kai Zou, Wenhu Chen, Jimmy Lin, Dongfu Jiang, and Yu Zhang. 2026. Dr-dci: Scaling direct corpus interaction via dynamic workspace expansion.arXiv preprint arXiv:2606.14885

    arXiv 2026

  19. [19]

    OpenAI. 2026. Gpt-5.4 nano model. https://developers.openai.com/api/docs/ models/gpt-5.4-nano

    2026

  20. [20]

    OpenAI. 2026. Introducing GPT-5.4. https://openai.com/index/ introducing-gpt-5-4/. 12

    2026

  21. [21]

    2009.The probabilistic relevance framework: BM25 and beyond, volume 4

    Stephen Robertson and Hugo Zaragoza. 2009.The probabilistic relevance framework: BM25 and beyond, volume 4. Now Publishers Inc

    2009

  22. [22]

    Alireza Salemi, Chang Zeng, Atharva Nijasure, Jui-Hui Chung, Razieh Rahimi, Fernando Diaz, and Hamed Zamani. 2026. Grepseek: Training search agents for direct corpus interaction.arXiv preprint arXiv:2605.29307

    Pith/arXiv arXiv 2026

  23. [23]

    Aditi Singh, Abul Ehtesham, Saket Kumar, and Tala Talaei Khoei. 2025. Agentic retrieval- augmented generation: A survey on agentic rag.arXiv preprint arXiv:2501.09136

    Pith/arXiv arXiv 2025

  24. [24]

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. 2025. R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592

    Pith/arXiv arXiv 2025

  25. [25]

    Hongjin Su, Howard Yen, Mengzhou Xia, Weijia Shi, Niklas Muennighoff, Han-yu Wang, Liu Haisu, Quan Shi, Zachary Siegel, Michael Tang, et al. 2025. Bright: A realistic and challenging benchmark for reasoning-intensive retrieval. InInternational Conference on Learning Representations, volume 2025, pages 48941–48991

    2025

  26. [26]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. Interleav- ing retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pages 10014–10037

    2023

  27. [27]

    Junjie Wu, Jiangnan Li, Yuqing Li, Lemao Liu, Liyan Xu, Jiwei Li, Dit-Yan Yeung, Jie Zhou, and Mo Yu. 2026. Situated embedding models for context-aware dense retrieval. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (V olume 2: Short Papers), pages 37–49

    2026

  28. [28]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629

    Pith/arXiv arXiv 2022

  29. [29]

    Mo Yu, Tsz Ting Chung, Chulun Zhou, Tong Li, Rui Lu, Jiangnan Li, Liyan Xu, Haoshu Lu, Ning Zhang, Jing Li, et al. 2025. Prelude: A benchmark designed to require global comprehension and reasoning over long contexts.arXiv preprint arXiv:2508.09848

    Pith/arXiv arXiv 2025

  30. [30]

    Yuqi Zeng, Qixiang Deng, Yulei Wan, Ruiquan Jiang, Xiaoqing Zheng, and Xuanjing Huang

  31. [31]

    Wenyuan Zhang, Xinghua Zhang, Haiyang Yu, Shuaiyi Nie, Bingli Wu, Juwei Yue, Tingwen Liu, and Yongbin Li. 2026. Expseek: Self-triggered experience seeking for web agents.arXiv preprint arXiv:2601.08605

    Pith/arXiv arXiv 2026

  32. [32]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176

    Pith/arXiv arXiv 2025

  33. [33]

    Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. 2025. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments.arXiv preprint arXiv:2504.03160

    Pith/arXiv arXiv 2025

  34. [34]

    Chulun Zhou, Chunkang Zhang, Guoxin Yu, Fandong Meng, Jie Zhou, Wai Lam, and Mo Yu

  35. [35]

    query",

    Shengyao Zhuang, Yuansheng Ni, Hengxin Fun, Jimmy Lin, and Xueguang Ma. 2026. Towards retrieving interaction spaces for agentic search.arXiv preprint arXiv:2606.06880. A Agent Implementation Details We detail how the compared agents are configured for a fair comparison on both benchmarks. 13 BRIGHT prompting.On BRIGHT, we found that the default DCI IR pro...

    Pith/arXiv arXiv 2026

  36. [39]

    relative/path/to/doc1.txt

  37. [40]

    relative/path/to/doc2.txt

  38. [41]

    pattern" to search only within recalled documents. May be called multiple times with different queries to create multiple scopes. Parameters. {

    relative/path/to/doc10.txt Figure 6: The system prompt used by RARG for the BRIGHT retrieval evaluation. B.3 Tool Specifications We list the tool schemas exposed to the agent.embed_recallperforms document-level relevance recall and writes a scope file;readreturns line-bounded file contents; andbashexecutes shell commands, mainly rg. The standardbashis use...

    2000

  39. [2020]

    Russell David Lyons . . . Ph.D., August 1983, Mathematics

    Person A coauthored a 1990–2005 paper with Persons B and C; Person B won the Rollo Davidson Prize in that period, and Person C published a 1990s paper whose title ends in “Line. ” What is Person A’s full name? We show the three RARG trajectories on a common turn axis in Figure 7. For readability, consecutive calls pursuing the same search objective are ma...

    1990

  40. [2024]

    InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 11057–11070

    Selection-p: Self-supervised task-agnostic prompt compression for faithfulness and transferability. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 11057–11070

    2024

  41. [2025]

    InF orty-third International Conference on Machine Learning

    Hgmem: Hypergraph-based working memory to improve multi-step rag for long-context complex relational modeling. InF orty-third International Conference on Machine Learning

  42. [2026]

    Rethinking agentic rag: Toward llm-driven logical retrieval beyond embeddings.arXiv preprint arXiv:2605.27123

    Pith/arXiv arXiv