Pith. sign in

REVIEW 3 major objections 7 minor 42 references

Relevance should steer how agents search a corpus—not only which documents they open—so hard answers arrive with far fewer wasted tool calls.

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 · grok-4.5

2026-07-31 20:35 UTC pith:43LAMBLR

load-bearing objection Solid systems paper: relevance-as-execution-prior for scoped rg is a real interface move; efficiency gains look sturdy, accuracy headlines are still thin on n=100. the 3 major comments →

arxiv 2607.24223 v1 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 agentic searchrelevancecorpus interactionripgrepdirect corpus interactionbrowse question answeringreasoning-intensive retrievalretrieval-augmented generation
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.

This paper argues that relevance is being used too narrowly in agentic search. Conventional systems use it to pick a top-k shortlist, which cannot itself localize, join, or verify the small spans complex questions need. Direct corpus interaction fixes the interface with grep-style tools, but without a relevance prior it often finds useful clues late and burns many unproductive steps. The authors claim a third role: keep fine-grained pattern matching, but treat relevance as an execution prior that decides where search begins, which files are scanned first, and which matches remain visible when output is truncated. Their method, RARG, ranks documents into a scope, forces sequential ripgrep over that order, seeds a few query-relevant paragraphs, and can rerank local matches. On hard browsing QA and reasoning-intensive retrieval, this improves the accuracy–efficiency trade-off over retrieval-only agents, unrestricted direct interaction, and systems that only use relevance to build a workspace.

Core claim

Relevance can guide corpus interaction itself, not merely select its inputs. By turning retrieval scores into coarse-to-fine execution control—document-ordered sequential ripgrep, entry-point paragraphs, and match-level reranking—agents expose globally useful clues earlier, keep informative excerpts in view under truncation, and converge to verified answers with higher accuracy and fewer tool calls than top-k retrieval or relevance-agnostic direct search.

What carries the argument

RARG (Relevance-Aware RipGrep Search Agent): an execution prior that (1) ranks documents into a scope file and forces single-threaded ripgrep traversal in that order, (2) initializes search with a handful of query-relevant paragraphs, and (3) reranks a wider pool of grep matches so locally informative excerpts can outrank noise from document-level scores alone.

Load-bearing premise

The claim stands only if embedding scores are a good enough guide for both whole-document order and short-match visibility, and if the model reliably follows the scoped multi-step search protocol.

What would settle it

On the same hard browsing benchmark and model backbone, disable ordered scoped ripgrep and match reranking: if accuracy and tool-call cost no longer beat unrestricted direct corpus interaction and retrieval-built interaction spaces—or if gains disappear when the embedder is weak or the model ignores scoped-search instructions—the central claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Agents can keep high-resolution grep-style composition while using retrieval scores as runtime guidance rather than a hard top-k content bottleneck.
  • Document ranking alone is incomplete: controlling which local matches stay visible under truncation recovers evidence buried in lower-ranked files.
  • Relevance-ordered interaction can retain a large accuracy edge as the corpus grows, provided the retriever still captures gold documents in a bounded scope.
  • Optimal relevance granularity depends on the objective: match-level focus speeds answer convergence, while broader observation can help ranking-style recall.

Where Pith is reading between the lines

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

  • Tool stacks may increasingly treat rankers as schedulers for local operators (scan order, stop rules, observation budgets), not only as document fetchers.
  • Closing the train–eval gap on explicit local rerank queries could recover the fastest convergence path the paper saw without the accuracy drop.
  • A single agent that switches between depth-first verification and breadth-first candidate gathering from the same relevance signals could unify QA and ranking workloads.

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

3 major / 7 minor

Summary. The paper proposes RARG, a search agent built on the DCI-Agent-Lite harness that injects embedding-based relevance into grep-style corpus interaction as an "execution prior": a new embed_recall tool writes a ranked document scope to a file, the agent rg-searches along that ranked order (with a rule-based pass injecting -j1 to force single-threaded, order-preserving scanning), and two extensions add query-relevant entry-point paragraphs (RARG+) and match-level reranking of rg hits using a constructed scope-query-plus-keywords query (RARG++). On the 100-query BrowseComp-Plus setting at 100K documents, RARG++ reaches 84% vs 78% for RISE/DCI with GPT-5.4-mini at fewer tool calls, 91% with GPT-5.4, and retains a 10-point margin after scaling the corpus to 1M documents; on four BRIGHT subsets RARG+ attains the best average nDCG@10 (53.36). Behavioral analyses (scope-rank hit heatmaps, scope recall, Bash composition) support the claimed mechanism, and a generative rerank-query variant is reported honestly as degrading accuracy.

Significance. If the results hold, the paper makes a genuinely useful conceptual contribution to agentic search: separating relevance as an execution prior (traversal order, match visibility under truncation) from relevance as a content-selection filter. This is a clean articulation of a real gap between RISE-style interaction-space construction and DCI-style unguided grep, and the mechanism claims are backed by more than headline numbers — the hit-rank heatmap (Fig. 3), scope-recall/RG-coverage measurements (Fig. 4a), and Bash-composition decomposition (Fig. 4b) directly test the proposed causal story rather than merely decorating it. The 100K→1M scale-up with long FineWeb-Edu distractors is a meaningful stress test that the authors report honestly (all methods degrade, contradicting RISE's non-decreasing trend), and the negative generative-variant result (Table 1, 75% at 17.8 tools) is candidly disclosed. Code release is promised. The efficiency half of the frontier claim (23.9 vs 99.1 DCI tool calls; flat tool counts under 10× scaling) is robustly supported; the accuracy half currently rests on small absolute differences on a 100-query sample (see major comments).

major comments (3)
  1. [§4.2, Table 1] No uncertainty estimates are reported anywhere, yet the accuracy leg of the central claim rests on small absolute differences over n=100 binary outcomes. RARG++ 84% vs RISE/DCI 78% is a 6-query gap; the coarse-to-fine progression 80→81→84 is 1 and 3 queries. Independent-proportion standard errors at n=100 are ~4–5 points per arm, so these gaps are within unpaired sampling noise unless the paired discordance is strongly one-sided. Since all methods are run on the same 100 queries, a paired analysis is available at no extra experimental cost: please report per-query discordant-pair counts and a McNemar test (or paired bootstrap CIs) for the headline comparisons. If the paired split is roughly even, the abstract's claim of 'higher accuracy with substantially fewer tool calls' should be tempered to 'DCI-level accuracy at substantially lower cost' — still a contribution, but a different one.
  2. [Table 1, GPT-5.4 row] The single largest headline gain (91% vs 82%, 'exceeding RISE by 9 points', §1 and §4.2) compares RARG++ at medium thinking effort against a RISE number cited with an explicit '?' for unknown thinking effort. Inference-time configuration materially affects agent accuracy on this task family (the paper itself varies effort across backbones). The flagship comparison is therefore uncontrolled. Please either re-run RISE with GPT-5.4 at medium effort using the official implementation (as the authors already do for the mini/nano rows), or restrict the GPT-5.4 claim to the efficiency dimension and state the confound prominently in §4.2 rather than only in a table footnote.
  3. [§4.3, Table 2] Two issues. (1) The parenthesized deltas in Table 2 (e.g., Tools '24.7(+0.8)') are never defined in the caption or text; from context they appear to be changes relative to the 100K-corpus numbers in Table 1, but this must be stated explicitly, and the mixed signs (Search −0.1, Bash +0.9) deserve one sentence of interpretation. (2) The 1M comparison uses only RISE-BM25, not RISE itself, even though §4.2 argues the navigational structuring (not the retriever) is what distinguishes RISE — and the scaling result is exactly where structure might matter. Either run RISE proper at 1M or justify why RISE-BM25 is the right comparator for the scaling claim; the current 'keeps a 10-point margin over RISE' phrasing in §1 overstates what Table 2 shows.
minor comments (7)
  1. [Fig. 4a and §4.5] Nano's lower scope recall (91.8% vs 95–97%) is attributed to 'weaker embed_recall queries and frequent, less disciplined embed_recall triggering.' Scope recall as defined is a per-scope quantity averaged within sample; more frequent re-scoping should if anything raise per-sample gold coverage. Please clarify whether recall is computed per individual scope or per union of scopes within an episode, and reconcile the explanation with the definition.
  2. [§4.1, Implementation Details] Several load-bearing hyperparameters (scope cap 10K, match caps 30/60, char truncation 1000/500, M=500/m=30, top-10 entry paragraphs, 400–1000-char paragraph splitting) are stated without ablation or sensitivity analysis. A small sweep of at least m (or the char cap) on a subset would strengthen the claim that the design is robust rather than tuned; at minimum, say how these were chosen.
  3. [Table 3] RARG++ underperforms RARG+ and RARG on BRIGHT (50.55 vs 53.36/51.75). The breadth-vs-depth explanation in §4.4 is plausible but post hoc; consider testing it directly (e.g., correlate per-query reranking benefit with gold-set size) or soften the framing from 'the optimal granularity depends on the task' to an explicitly hypothesized mechanism.
  4. [Table 1 caption] 'Parts of Search/Bash are underlined to show different behaviors of agents' — the underlining convention is not visible/meaningful as described; replace with explicit markers or a sentence in §4.2 identifying which cells are intended.
  5. [§3.2] The -j1 injection is rule-based regex matching on Bash commands; state briefly what happens when the model emits rg with unusual flag ordering, pipes, or compound commands (does the injection ever fail silently, reverting to unordered output?). Fig. 4b's scoped-rg shares suggest high but not total adherence — quantifying injection success rate would close this gap.
  6. [§4.1] The DCI prompt for BRIGHT was 'modified to explicitly require the agent to rank 10 document names' because the default performed substantially worse. This is a baseline-affecting change; please document the exact prompt change in an appendix and report the unmodified DCI number if available, to rule out that the BRIGHT DCI baseline is handicapped.
  7. [General] Typos/formatting: 'paragrpahs' (Fig. 2), '\tmp\scope_{N}.txt' uses backslashes inconsistently with the forward-slash paths shown elsewhere, Table 2 caption says 'medium reasoning' while Table 1 says 'medium effort' — unify terminology.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with external accuracy/nDCG/tool-cost metrics; relevance is an execution prior, not redefined as the evaluation target.

full rationale

RARG’s load-bearing claims are empirical comparisons (accuracy, nDCG@10, turns/tool calls) on BrowseComp-Plus and BRIGHT against DCI, RISE, Retrieval-Agent, and NeMo. Document-level ranking, entry-point paragraphs, and match reranking use off-the-shelf embedders (Qwen3-Embedding-4B, llama-nv-embed-reasoning-3b) as priors over rg order and visibility; answers and rankings are judged externally (LLM-as-judge on BC+, nDCG@10 on BRIGHT), not by re-scoring the same relevance objective as the success metric. There is no derivation chain in which a fitted parameter is renamed a prediction, no uniqueness theorem imported from overlapping authors to forbid alternatives, and no self-definitional identity (X defined via Y then “derived” as Y). Self-citations and ecosystem positioning (DCI, RISE, related agentic-search work) are background and baseline context, not the sole justification of the central accuracy–efficiency result. Per the circularity criteria, an honest non-finding applies: score 0, no circular steps.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Load-bearing commitments are standard IR/LLM-agent assumptions plus engineering choices (scope cap, sequential -j1, constructed rerank query, compaction). No new physical entities. Free parameters are hyperparameters chosen for eval, not fitted to prove a theory identity.

free parameters (5)
  • scope_size_cap = 10000
    embed_recall writes at most 10,000 ranked paths; claimed high evidence recall but chosen by authors.
  • match_rerank_pool_M_and_top_m = M=500; m=30 (BC+), m=60 (BRIGHT)
    Rerank up to M matches, keep top m; set per benchmark without formal selection study.
  • rg_match_and_char_caps = 30 matches/1000 chars BC+; 60/500 BRIGHT
    Caps on returned matches and per-match characters control observation budget and directly affect what the LLM sees.
  • entry_point_top_paragraphs = top-10 paragraphs
    Top-10 paragraphs from top-X scope docs, 400–1000 char splits, for RARG+ initialization.
  • compaction_and_turn_limits = 40 results / 230K / 100 turns
    Keep last 40 tool results, 230K threshold, max 100 turns—affects cost and available context.
axioms (5)
  • domain assumption Embedding similarity is a useful query-dependent prior for where useful evidence is likely to lie in a corpus.
    Stated in Introduction and Method; entire execution-prior design depends on it.
  • ad hoc to paper Forcing single-threaded rg (-j1) over a ranked path list preserves relevance order and improves search convergence versus parallel unordered emission.
    Section 3.2 rule-based injection of -j1; core mechanism of RARG.
  • domain assumption LLM-as-judge (GPT-5.1) accuracy on BrowseComp-Plus is a valid primary QA metric.
    Section 4.1 follows DCI judge prompt; standard but imperfect for agent eval.
  • domain assumption Agent tool-call counts and turns are meaningful efficiency proxies for interaction cost.
    Used throughout Fig. 1 and Tables 1–3; dollar costs also shown in Fig. 1.
  • ad hoc to paper Constructed rerank queries (scope query + rule-extracted rg keywords) adequately capture global goal and local intent.
    Section 3.4; generative alternative underperforms, so this construction is load-bearing for RARG++.
invented entities (2)
  • RARG / embed_recall scope file as execution prior independent evidence
    purpose: Materialize document-level relevance as ordered paths that rg must traverse sequentially.
    New agent tool/protocol relative to plain DCI; not a physical entity but a postulated interface object central to the claim.
  • Match-level relevance insertion (RARG++) independent evidence
    purpose: Rerank a wide rg match pool so locally informative excerpts compete under truncation.
    Introduced to fix document-level rank burying good spans; validated only within this paper’s evals.

pith-pipeline@v1.2.0-grok45-kimik3 · 22945 in / 3434 out tokens · 69119 ms · 2026-07-31T20:35:22.826077+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 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 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 ↗

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

    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. 11

    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/

    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. 12

    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 Prompts for RARG A.1 System Prompt of BC+ Figure 5 shows the system prompt used by RARG on BrowseComp-Plus. It states the corpus-listing constraint, describes theembed_recalltool and the ...

    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. A.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