Pith. sign in

REVIEW 4 major objections 5 minor 158 references

SearchOS claims that making search progress explicit and shared lifts long-horizon web search completeness, with 80.3 F1 on WideSearch and a 13.4-point gain on GISA set questions.

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

2026-08-01 23:42 UTC pith:4HOMSJW4

load-bearing objection A genuinely detailed systems paper for agentic search whose architecture deserves attention, but the headline benchmark margins are not yet credible until baseline settings and run-level variance are disclosed. the 4 major comments →

arxiv 2607.15257 v1 pith:4HOMSJW4 submitted 2026-07-16 cs.AI cs.IR

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

Yuyao Zhang , Junjie Gao , Zhengxian Wu , Jiaming Fan , Jin Zhang , Shihan Ma , Yao Yao , Weiran Qi
show 6 more authors
Chuyan Jin Guiyu Ma Xingzhong Xu Kai Yang Ji-Rong Wen Zhicheng Dou
This is my paper
classification cs.AI cs.IR
keywords multi-agent searchopen-domain information seekingrelational schema completiongrounded citationsSearch-Oriented Context Managementpipeline parallelismagent middlewareweb search agents
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 claims that long-horizon web search fails because progress, evidence, and failures live only in conversation history. SearchOS tries to fix this by making search state a system-level object: agents fill a relational schema of entities and attributes, every value is tied to a cited source, and shared memories track unresolved gaps, evidence, coverage, and failures. The result, if the experiments are taken at face value, is the best reported completeness on both WideSearch and GISA, including an 80.3 item F1 on WideSearch and a 76.5 set F1 on GISA, 13.4 points above the strongest baseline on the latter. A sympathetic reader would care because the mechanism—externalized state rather than prompt engineering—could transfer to any long-horizon information-gathering task.

Core claim

The central discovery is that open-domain information seeking can be modeled as relational schema completion with grounded citations, and that this formulation yields measurable gains when combined with externalized search state. SearchOS treats a user request as a set of linked tables; agents discover rows, fill attributes, and anchor each value to a URL and excerpt. SOCM keeps four persistent memories—pending tasks, an evidence graph, a coverage map, and failure records—outside agent contexts, and a middleware harness enforces grounding and stall detection outside prompts. On WideSearch the system reaches 80.3 item F1 and on GISA 76.5 set F1, beating all evaluated baselines on every headli

What carries the argument

The relational search schema (tables, primary keys, foreign keys) is the target, and SOCM is the engine: Frontier Task tracks schedulable gaps, Evidence Graph stores atomic value-source-anchor triples, Coverage Map marks cells filled/uncertain/missing, and Failure Memory records dead queries and inaccessible sources. Pipeline-parallel continuous dispatch refills released agent slots with unresolved coverage gaps; the Search Tool Middleware Harness intercepts model/tool calls to inject context, extract and bind evidence, and detect stalls or budget overruns. A hierarchy of 280 strategy and access skills supplies reusable search knowledge.

Load-bearing premise

The headline benchmark gains depend on the assumption that every comparison baseline was run under the same backbone model, budget caps, and scoring protocol as SearchOS; the paper reports SearchOS's configuration but not the baselines'.

What would settle it

Re-run the main comparisons with all methods under identical budget caps, backend, and evaluation pipeline, with per-method settings disclosed; if the GISA Set-F1 margin over the strongest baseline falls below a few points, the central claim of systematic completeness gains is not supported.

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

If this is right

  • If SearchOS's claims hold, completeness-sensitive benchmarks (set enumeration, table filling) improve substantially—13.4 F1 points on set questions in the paper's main evaluation.
  • Because each value is grounded in a source URL and excerpt, the final report carries verifiable citations; coverage gaps remain visible rather than being silently filled.
  • Pipeline-parallel dispatch reduced end-to-end time by about 24% and used fewer LLM calls in the paper's ablation, suggesting efficiency gains come with completeness.
  • Skills, acting as reusable search strategy and site-specific access routines, cut search calls by 39% and page calls by 43% in the reported ablation.
  • The architecture separates global orchestration from local search, so new roles or models can be added without changing the shared state and enforcement layer.

Where Pith is reading between the lines

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

  • The relational-schema view suggests a natural evaluation metric for any open-ended research task: count how many schema cells are filled and cited; this could generalize beyond the two benchmarks.
  • Dynamic schema selection—choosing single vs. multi-table during search—may matter most when tasks have one-to-many relations; a testable extension is to run the same system with schema planning disabled and measure the gap.
  • Failure Memory across runs implies a route to amortize search knowledge: if post-mortems seed future tasks, repeated failures on similar sources should decline; that claim is not directly tested here.
  • The middleware harness's role-specific projections could allow heterogeneous backbone models to share one search state; testing with different base models would show whether enforcement transfers.

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 SearchOS, a multi-agent framework for open-domain information seeking. The core idea is to treat an information-seeking request as relational schema completion with grounded citations: the system discovers entities, fills table attributes, and anchors each value to source evidence. A Search-Oriented Context Management (SOCM) component externalizes shared state as a Frontier Task pool, an Evidence Graph, a Coverage Map, and Failure Memory. Pipeline-parallel orchestration dispatches unresolved schema gaps, and a Search Tool Middleware Harness intercepts model/tool interactions to ground evidence, detect stalls, and enforce budgets. The system also includes 280 hand-built hierarchical skills. Experiments on WideSearch and GISA report state-of-the-art results, including 80.3 item-level F1 on WideSearch and 76.5 set F1 on GISA, with the latter claimed to exceed the strongest baseline by 13.4 points. The paper includes detailed appendices with implementation interfaces and process-level case studies.

Significance. If the empirical claims hold, SearchOS is a meaningful systems contribution: it moves search progress out of the model context into explicit, shared, durable state and provides middleware-level enforcement of grounding and budget constraints. The paper is unusually detailed in its system description, ships code and interface specifications, and includes qualitative trajectory analyses that illustrate the intended control loop. The central design ideas — relational schema completion, externalized coverage/evidence/failure state, pipeline-parallel dispatch, and middleware governance — are coherent and relevant to current long-horizon agent research. However, the significance is currently limited by evaluation-control concerns. The main benchmark claim rests on a comparison whose apples-to-apples status is not established, and there is an internal inconsistency between the headline F1 and the scheduling-ablation F1. These issues are fixable, but until they are addressed the quantitative claims should be treated with caution.

major comments (4)
  1. [§4.2–§4.3, Table 2] The headline claim 'leads all metrics' depends on an apples-to-apples comparison, but the paper specifies the experimental configuration only for SearchOS. It lists baselines in §4.2 and gives SearchOS's GLM-5 backbone, Qwen3.5-35B-A3B extraction model, 50 orchestrator iterations, 8 sub-agents, 20 searches/sub-agent, 1800 s, and Max@3 in §4.3, without stating the corresponding settings for ReAct, Plan-and-Solve, Table-as-Search, A-MapReduce, or Web2BigTable. If baseline numbers were taken from prior papers with weaker backbones, different budget caps, or non-Max@3 reporting, the reported margins, especially +13.4 GISA Set F1, would be evaluation-protocol artifacts rather than framework advantages. The authors should re-run all baselines in the same harness with the same backbone, budget, and scoring, or explicitly state the provenance of each baseline number and justify comparability.
  2. [§5.2, Tables 4–5] Table 2 reports WideSearch Item F1 80.3 for SearchOS (Max@3), while Table 5 reports Continuous scheduling Item F1 86.75 (mean across 30 trajectories) and Table 4 reports round-wise gains over batch of +2.35, +15.00, and +1.85 pp. A mean over trajectories cannot exceed a per-case best-of-three maximum of the same protocol, so these numbers cannot both describe the same benchmark/evaluation. If Table 5 uses a different subset, different trajectories, or a different scoring version, that must be stated and reconciled. As written, the internal inconsistency undermines confidence in the reported effect of pipeline-parallel scheduling.
  3. [§3.4, §4.4, Table 6] The 280 pre-built skills include domain-specific and site-specific access skills, and §3.4 explicitly defers large-scale skill synthesis to follow-up work. The paper provides no leakage analysis showing that these hand-built skills do not encode WideSearch/GISA task patterns or direct answers. The skill ablation (Table 6) attributes +2.0 Item F1 and +3.4 Row F1 to skills, but if skills were constructed with knowledge of the test benchmarks, this gain is not a generalizable system property. Please report how skills were curated relative to the benchmark queries and provide a task-agnostic leakage check (e.g., skill usefulness on unseen domains or a held-out skill set).
  4. [§4.3, §4.4, Table 2] All main results are reported as Max@3 with no variance, confidence intervals, or significance tests. With stochastic LLM agents and best-of-three selection, differences as small as +0.1 Item Precision on WideSearch and large but possibly noisy gains such as +13.4 GISA Set F1 cannot be assessed. Please report per-case means and standard deviations (or paired significance tests) for both SearchOS and baselines, and state how many runs the baseline numbers average over. This is necessary to support the 'leads all metrics' claim.
minor comments (5)
  1. [Abstract and §4.4] The paper claims SearchOS 'leads all metrics,' but the GISA Item EM row in Table 2 is tied at 50.0 with Web2BigTable. Please say 'leads or ties' or define how ties are handled.
  2. [§4.3] The phrase 'scaled by ×100' is ambiguous. State explicitly that F1/precision/recall values are reported as percentages.
  3. [§4.1] GISA (Zhu et al., 2026) was co-created by two of the present co-authors (J.-R. Wen and Z. Dou). The paper should disclose this relationship and confirm that SearchOS was evaluated on the benchmark as provided, without development against test answers.
  4. [Figure 4] Figure 4(a), 'Frequent skill keywords,' lacks axis labels and a clear frequency scale; as printed it is difficult to interpret.
  5. [Tables 4–5] Table 4 reports 'median across 10 cases' while Table 5 reports 'mean across 30 trajectories per policy.' Please unify the experimental units and definitions so the reader can compare the two ablations.

Circularity Check

0 steps flagged

No significant circularity: the paper's components are constructive and its headline results are empirical; the GISA benchmark overlap and under-specified baseline settings are validity concerns, not derivation-level circularity.

full rationale

The paper is a systems/empirical contribution rather than a derivation chain, and I find no step where a stated result reduces by definition or by self-citation to its own inputs. The relational schema-completion objective (Eqs. 1–2), SOCM state (Eqs. 3–9), coverage criterion (Eq. 8), middleware evidence-acceptance rule (Eq. 14), stall detector (Eq. 17), and pipeline dispatch rule (Eq. 10) are constructive definitions and implementation mechanisms; none is a fitted parameter later relabeled as a prediction. The 280 pre-built skills are hand-authored, and the paper explicitly defers skill synthesis from data ('In follow-up work, we will detail large-scale search agent skill synthesis from data sources, search trajectories, and user intent. Stay tuned!', §3.4), which is an acknowledged scope limitation, not a circularity. The main results on GISA do involve a benchmark co-created by two senior authors of this paper (Zhu et al., 2026), and Section 4.3 does not state baseline configurations, so the comparison may be imperfectly controlled; however, those are experimental-validity and reporting concerns, not cases where the paper's claim is equivalent to its inputs by construction. No cited uniqueness theorem or ansatz is imported from the authors' prior work to force the design. Under the hard rule requiring a specific exhibited reduction, no significant circularity is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Five hand-chosen or unspecified numeric components affect the reported results (budget caps, Max@3 selection, stall-window and sensor thresholds, evidence-quality scores τ/α/γ, and the Bind/Anchor acceptance functions); none have sensitivity analyses. Five domain assumptions underpin the claim, the most load-bearing being baseline comparability and the task-agnosticism of the 280 hand-built skills. Three software artifacts are introduced (SOCM, middleware harness, skill library); they have empirical support via ablations and trajectory traces but are never isolated from one another, so their individual contributions are unmeasured. No physical or mathematical entities are postulated.

free parameters (5)
  • Session budget caps = 50 orchestrator iterations; 8 sub-agents; 20 searches/sub-agent; 1800 s
    Set by hand in §4.3; no sensitivity analysis; results likely depend on them.
  • Max@3 run selection = best of three runs
    §4.3; inflates point estimates; used for headline numbers in Tables 2 and 6.
  • Stall window w and budget thresholds B_iter/B_search/B_time = unspecified
    Eqs. (17)-(18) define stall state and budget pressure without giving numeric values.
  • Evidence scoring (tier τ, alignment α, confidence γ) = unspecified
    Eq. (7) lexicographic cell-evidence selection; the component scores are not operationalized.
  • Evidence acceptance functions Bind and Anchor = unspecified
    Eq. (14): acceptance requires schema binding and span anchoring; underlying criteria are not defined.
axioms (5)
  • domain assumption Open-domain information seeking is faithfully modeled as relational schema completion with grounded citations
    §2 defines the task this way; evaluation metrics then measure schema completion, entangling formulation and metric.
  • domain assumption Pipeline-parallel utilization gains from DNN training transfer to agent scheduling
    §3.2 cites Huang et al. 2019 and Narayanan et al. 2019 and assumes the analogy; supporting ablation covers only 10-30 short cases.
  • domain assumption Baselines were evaluated under conditions comparable to SearchOS
    §4.2-4.3 list baselines but give no per-baseline backbone, budget, or run count; without this, Table 2 gains are uncontrolled.
  • domain assumption Best-of-three (Max@3) summarizes performance faithfully
    §4.3; no variance reported, so run-to-run stability is unknown and point estimates are optimistic.
  • domain assumption The 280 pre-built skills are task-agnostic rather than benchmark-tuned
    §3.4 describes hand-built skills refined from trajectories and defers data-driven synthesis; no leakage analysis is given.
invented entities (3)
  • SOCM state (Frontier Task, Evidence Graph, Coverage Map, Failure Memory) independent evidence
    purpose: Externalize search progress as shared, persistent system state across orchestrator, explore, search, and writer agents
    Concretely specified (§3.1), exemplified (App. A), and evaluated end-to-end, but never ablated as a whole component.
  • Search Tool Middleware Harness independent evidence
    purpose: Intercept model/tool interactions to prepare context, ground evidence, and enforce budgets and stall detection
    Trajectory-level evidence in Fig. 5 shows interventions restoring stalled progress; only joint effects are measured.
  • Hierarchical skill library (280 skills) independent evidence
    purpose: Reuse search strategies and site-specific access procedures across sessions
    Ablation shows +2.0 item F1 and −36.6% session time, but skills are hand-built and data-driven synthesis is deferred (§3.4); benchmark leakage is not ruled out.

pith-pipeline@v1.3.0-alltime-deepseek · 18907 in / 20123 out tokens · 168251 ms · 2026-08-01T23:42:50.881826+00:00 · methodology

0 comments
read the original abstract

Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search attempts fail to yield useful evidence, current single- and multi-agent systems can become trapped in repetitive loops, wasting search budgets and ultimately compromising the quality and completeness of the final output. We introduce SearchOS, a system-level multi-agent framework that turns fragile, implicit search progress into explicit, persistent, and shared state. First, we formulate open-domain information seeking as relational schema completion with grounded citations, where agents discover entities, populate attributes across linked tables, and anchor each value to source evidence. Then we design Search-Oriented Context Management (SOCM), which externalizes the evolving state into Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory. Built on SOCM, SearchOS applies a pipeline-parallel scheduling mechanism that overlaps the execution of sub-agents and continuously refills freed slots with tasks targeting unresolved coverage gaps to improve utilization and throughput. To schedule and control the execution of search agents, SearchOS introduces a Search Tool Middleware Harness that intercepts model and tool interactions to record grounded evidence and react to stalls or budget exhaustion, and provides a reusable hierarchical skill system comprising strategy and access skills to augment the agents' search process and avoid repeating failed search patterns across runs. On WideSearch and GISA, SearchOS leads all metrics among the evaluated single- and multi-agent baselines, paving the way toward robust information-seeking collaboration.

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

158 extracted references · 41 linked inside Pith

  1. [1]

    Abhyankar, Z

    R. Abhyankar, Z. He, V. Srivatsa, H. Zhang, and Y. Zhang. I nfer C ept: Efficient intercept support for augmented large language model inference. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 81--95. PMLR, 2024. URL https://proceedings.mlr.press/v235/abhyankar24a.html

    2024

  2. [2]

    Effective context engineering for AI agents

    Anthropic . Effective context engineering for AI agents. Anthropic Engineering, Sept. 2025. URL https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents

    2025

  3. [3]

    P. A. Bernstein. Synthesizing third normal form relations from functional dependencies. ACM Transactions on Database Systems (TODS), 1 0 (4): 0 277--298, 1976

    1976

  4. [4]

    M. Chen, L. Sun, T. Li, H. Sun, Y. Zhou, C. Zhu, H. Wang, J. Z. Pan, W. Zhang, H. Chen, F. Yang, Z. Zhou, and W. Chen. Research: Learning to reason with search for llms via reinforcement learning, 2025 a . URL https://arxiv.org/abs/2503.19470

    Pith/arXiv arXiv 2025

  5. [6]

    W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C.-M. Chan, H. Yu, Y. Lu, Y.-H. Hung, C. Qian, Y. Qin, X. Cong, R. Xie, Z. Liu, M. Sun, and J. Zhou. AgentVerse : Facilitating multi-agent collaboration and exploring emergent behaviors. In International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=EHg5GDnyq1

    2024

  6. [7]

    Z. Chen, K. Liu, Q. Wang, J. Liu, W. Zhang, K. Chen, and F. Zhao. MindSearch : Mimicking human minds elicits deep ai searcher. In International Conference on Learning Representations, 2025 b . URL https://openreview.net/forum?id=xgtXkyqw1f

    2025

  7. [8]

    E. F. Codd. A relational model of data for large shared data banks. Communications of the ACM, 13 0 (6): 0 377--387, 1970

    1970

  8. [9]

    Fourney, G

    A. Fourney, G. Bansal, H. Mozannar, C. Tan, E. Salinas, E. Zhu, F. Niedtner, G. Proebsting, G. Bassman, J. Gerrits, J. Alber, P. Chang, R. Loynd, R. West, V. Dibia, A. Awadallah, E. Kamar, R. Hosn, and S. Amershi. Magentic-One : A generalist multi-agent system for solving complex tasks. arXiv preprint arXiv:2411.04468, 2024. URL https://arxiv.org/abs/2411.04468

    Pith/arXiv arXiv 2024

  9. [10]

    D. Gao, Z. Li, X. Pan, W. Kuang, Z. Ma, B. Qian, F. Wei, W. Zhang, Y. Xie, D. Chen, L. Yao, H. Peng, Z. Zhang, L. Zhu, C. Cheng, H. Shi, Y. Li, B. Ding, and J. Zhou. AgentScope : A flexible yet robust multi-agent platform. arXiv preprint arXiv:2402.14034, 2024. URL https://arxiv.org/abs/2402.14034

    Pith/arXiv arXiv 2024

  10. [11]

    J. Gao, W. Fu, M. Xie, S. Xu, C. He, Z. Mei, B. Zhu, and Y. Wu. Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl, 2025. URL https://arxiv.org/abs/2508.07976

    arXiv 2025

  11. [12]

    Y. Gu, Y. Shu, H. Yu, X. Liu, Y. Dong, J. Tang, J. Srinivasa, H. Latapie, and Y. Su. Middleware for LLM s: Tools are instrumental for language agents in complex environments. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7646--7663, 2024. doi:10.18653/v1/2024.emnlp-main.436. URL https://aclanthology.org/2...

    doi:10.18653/v1/2024.emnlp-main.436 2024

  12. [13]

    S. Hong, M. Zhuge, J. Chen, X. Zheng, Y. Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin, L. Zhou, C. Ran, L. Xiao, C. Wu, and J. Schmidhuber. MetaGPT : Meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=VtmBAGCN7o

    2024

  13. [14]

    Huang, Y

    Y. Huang, Y. Cheng, A. Bapna, O. Firat, M. X. Chen, D. Chen, H. Lee, J. Ngiam, Q. V. Le, Y. Wu, and Z. Chen. GPipe : Efficient training of giant neural networks using pipeline parallelism. In Advances in Neural Information Processing Systems, volume 32, pages 103--112, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/093f65e080a295f8076b1c5722a46a...

    2019

  14. [17]

    B. Jin, H. Zeng, Z. Yue, D. Wang, H. Zamani, and J. Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025 a . URL https://arxiv.org/abs/2503.09516

    Pith/arXiv arXiv 2025

  15. [19]

    S. Kim, S. Moon, R. Tabrizi, N. Lee, M. W. Mahoney, K. Keutzer, and A. Gholami. An LLM compiler for parallel function calling. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 24370--24391. PMLR, 2024. URL https://proceedings.mlr.press/v235/kim24y.html

    2024

  16. [21]

    J. Li, X. Xiao, Y. Zhang, C. Liu, L. Zhao, X. Liao, Y. Ji, J. Wang, J. Gu, Y. Ge, W. Xu, X. Fang, X. Xu, T. Zhao, Y. Kim, T. Wang, J. Hamm, S. Krishnaswamy, J. Huan, and C. K. Reddy. Agent harness engineering: A survey. OpenReview preprint, 2026 a . URL https://openreview.net/forum?id=eONq7FdiHa

    2026

  17. [23]

    X. Li, J. Jin, G. Dong, H. Qian, Y. Wu, J.-R. Wen, Y. Zhu, and Z. Dou. Webthinker: Empowering large reasoning models with deep research capability. Advances in Neural Information Processing Systems, 38: 0 120091--120131, 2026 b

    2026

  18. [24]

    Lopopolo

    R. Lopopolo. Harness engineering: Leveraging Codex in an agent-first world. OpenAI Engineering, Feb. 2026. URL https://openai.com/index/harness-engineering/

    2026

  19. [25]

    K. Luo, H. Qian, Z. Liu, Z. Xia, S. Xiao, S. Bao, J. Zhao, and K. Liu. Infoflow: Reinforcing search agent via reward density optimization, 2025. URL https://arxiv.org/abs/2510.26575

    arXiv 2025

  20. [26]

    K. Mei, X. Zhu, W. Xu, M. Jin, W. Hua, Z. Li, S. Xu, R. Ye, Y. Ge, and Y. Zhang. AIOS : LLM agent operating system. In Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=L4HHkCDz2x

    2025

  21. [29]

    Narayanan, A

    D. Narayanan, A. Harlap, A. Phanishayee, V. Seshadri, N. R. Devanur, G. R. Ganger, P. B. Gibbons, and M. Zaharia. PipeDream : Generalized pipeline parallelism for DNN training. In Proceedings of the 27th ACM Symposium on Operating Systems Principles, pages 1--15, 2019. doi:10.1145/3341301.3359646. URL https://doi.org/10.1145/3341301.3359646

    arXiv 2019

  22. [31]

    Packer, S

    C. Packer, S. Wooders, K. Lin, V. Fang, S. G. Patil, I. Stoica, and J. E. Gonzalez. MemGPT : Towards LLM s as operating systems. arXiv preprint arXiv:2310.08560, 2023. URL https://arxiv.org/abs/2310.08560

    Pith/arXiv arXiv 2023

  23. [33]

    C. Qian, W. Liu, H. Liu, N. Chen, Y. Dang, J. Li, C. Yang, W. Chen, Y. Su, X. Cong, J. Xu, D. Li, Z. Liu, and M. Sun. ChatDev : Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174--15186, 2024. doi:10.18653/v1/2024.acl-long.810. URL h...

    doi:10.18653/v1/2024.acl-long.810 2024

  24. [35]

    Y. Qin, Z. Cai, D. Jin, L. Yan, S. Liang, K. Zhu, Y. Lin, X. Han, N. Ding, H. Wang, et al. Webcpm: Interactive web search for chinese long-form question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8968--8988, 2023

    2023

  25. [36]

    Shinn, F

    N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html

    2023

  26. [38]

    Trivedi, N

    H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal. Interleaving 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

    2023

  27. [39]

    H. Wang, C. M. Poskitt, and J. Sun. AgentSpec : Customizable runtime enforcement for safe and reliable LLM agents. arXiv preprint arXiv:2503.18666, 2025 a . URL https://arxiv.org/abs/2503.18666

    Pith/arXiv arXiv 2025

  28. [40]

    L. Wang, W. Xu, Y. Lan, Z. Hu, Y. Lan, R. K.-W. Lee, and E.-P. Lim. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), 2023

    2023

  29. [41]

    Z. Wang, X. Zheng, K. An, C. Ouyang, J. Cai, Y. Wang, and Y. Wu. Stepsearch: Igniting llms search ability via step-wise proximal policy optimization, 2025 b . URL https://arxiv.org/abs/2505.15107

    Pith/arXiv arXiv 2025

  30. [45]

    R. Ye, Z. Zhang, K. Li, H. Yin, Z. Tao, Y. Zhao, L. Su, L. Zhang, Z. Qiao, X. Wang, P. Xie, F. Huang, S. Chen, J. Zhou, and Y. Jiang. Agentfold: Long-horizon web agents with proactive context management, 2025. URL https://arxiv.org/abs/2510.24699

    arXiv 2025

  31. [46]

    Zhang, Z

    Y. Zhang, Z. Dou, X. Li, J. Jin, Y. Wu, Z. Li, Y. Qi, and J. Wen. Neuro-symbolic query compiler. In W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, editors, Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025 , pages 12138--12155. Association for Computational Linguistics, 2025. URL https://ac...

    2025

  32. [47]

    Zhang, H

    Y. Zhang, H. Lu, J. Jin, H. Qian, S. Li, Z. Yang, Y. Zhu, J.-R. Wen, and Z. Dou. Web sitemap knowledge can enhance autonomous browsing. In Findings of the Association for Computational Linguistics: ACL 2026, pages 29307--29321, 2026

    2026

  33. [48]

    Zheng, D

    Y. Zheng, D. Fu, X. Hu, X. Cai, L. Ye, P. Lu, and P. Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 414--431, 2025

    2025

  34. [49]

    Zhong, L

    W. Zhong, L. Guo, Q. Gao, H. Ye, and Y. Wang. MemoryBank : Enhancing large language models with long-term memory. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19724--19731, 2024. doi:10.1609/aaai.v38i17.29946

    doi:10.1609/aaai.v38i17.29946 2024

  35. [51]

    National Science Review , volume=

    Neuro-symbolic approaches in artificial intelligence , author=. National Science Review , volume=. 2022 , publisher=

    2022

  36. [52]

    Neuron , volume=

    Direct fit to nature: an evolutionary perspective on biological and artificial neural networks , author=. Neuron , volume=. 2020 , publisher=

    2020

  37. [53]

    arXiv preprint arXiv:2404.14851 , year=

    From matching to generation: A survey on generative information retrieval , author=. arXiv preprint arXiv:2404.14851 , year=

    Pith/arXiv arXiv

  38. [54]

    arXiv preprint arXiv:2402.19473 , year=

    Retrieval-augmented generation for ai-generated content: A survey , author=. arXiv preprint arXiv:2402.19473 , year=

    Pith/arXiv arXiv

  39. [55]

    arXiv preprint arXiv:2409.10102 , year=

    Trustworthiness in retrieval-augmented generation systems: A survey , author=. arXiv preprint arXiv:2409.10102 , year=

    Pith/arXiv arXiv

  40. [56]

    Advances in Neural Information Processing Systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in Neural Information Processing Systems , volume=

  41. [57]

    Proceedings of the 28th International Conference on Computational Linguistics , pages=

    Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps , author=. Proceedings of the 28th International Conference on Computational Linguistics , pages=

  42. [58]

    Transactions of the Association for Computational Linguistics , volume=

    MuSiQue: Multihop Questions via Single-hop Question Composition , author=. Transactions of the Association for Computational Linguistics , volume=. 2022 , publisher=

    2022

  43. [59]

    Transactions of the Association for Computational Linguistics , volume=

    Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies , author=. Transactions of the Association for Computational Linguistics , volume=. 2021 , publisher=

    2021

  44. [60]

    arXiv preprint arXiv:1809.09600 , year=

    HotpotQA: A dataset for diverse, explainable multi-hop question answering , author=. arXiv preprint arXiv:1809.09600 , year=

    Pith/arXiv arXiv

  45. [61]

    arXiv preprint arXiv:2210.03350 , year=

    Measuring and narrowing the compositionality gap in language models , author=. arXiv preprint arXiv:2210.03350 , year=

    Pith/arXiv arXiv

  46. [62]

    arXiv preprint arXiv:2405.13576 , year=

    FlashRAG: A Modular Toolkit for Efficient Retrieval-Augmented Generation Research , author=. arXiv preprint arXiv:2405.13576 , year=

    Pith/arXiv arXiv

  47. [63]

    arXiv preprint arXiv:2310.11511 , year=

    Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. arXiv preprint arXiv:2310.11511 , year=

    Pith/arXiv arXiv

  48. [64]

    arXiv preprint arXiv:2305.15294 , year=

    Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy , author=. arXiv preprint arXiv:2305.15294 , year=

    Pith/arXiv arXiv

  49. [65]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  50. [66]

    Advances in Neural Information Processing Systems , volume=

    Grammar prompting for domain-specific language generation with large language models , author=. Advances in Neural Information Processing Systems , volume=

  51. [67]

    arXiv preprint arXiv:2404.00610 , year=

    Rq-rag: Learning to refine queries for retrieval augmented generation , author=. arXiv preprint arXiv:2404.00610 , year=

    Pith/arXiv arXiv

  52. [68]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Visual programming: Compositional visual reasoning without training , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  53. [69]

    arXiv preprint arXiv:2310.15164 , year=

    LINC: A neurosymbolic approach for logical reasoning by combining language models with first-order logic provers , author=. arXiv preprint arXiv:2310.15164 , year=

    Pith/arXiv arXiv

  54. [70]

    arXiv preprint arXiv:2402.00854 , year=

    SymbolicAI: A framework for logic-based approaches combining generative models and solvers , author=. arXiv preprint arXiv:2402.00854 , year=

    Pith/arXiv arXiv

  55. [71]

    Advances in Neural Information Processing Systems , volume=

    Semantic probabilistic layers for neuro-symbolic learning , author=. Advances in Neural Information Processing Systems , volume=

  56. [72]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  57. [73]

    Advances in Neural Information Processing Systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in Neural Information Processing Systems , volume=. 2023 , url=

    2023

  58. [74]

    Advances in Neural Information Processing Systems , volume=

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face , author=. Advances in Neural Information Processing Systems , volume=

  59. [75]

    IRE Transactions on information theory , volume=

    Three models for the description of language , author=. IRE Transactions on information theory , volume=. 1956 , publisher=

    1956

  60. [76]

    Proceedings of the first annual ACM symposium on Theory of computing , pages=

    Translations on a context free grammar , author=. Proceedings of the first annual ACM symposium on Theory of computing , pages=

  61. [77]

    2007 , publisher=

    Compilers principles, techniques & tools , author=. 2007 , publisher=

    2007

  62. [78]

    arXiv preprint arXiv:1906.02916 , year=

    Multi-hop reading comprehension through question decomposition and rescoring , author=. arXiv preprint arXiv:1906.02916 , year=

    Pith/arXiv arXiv 1906

  63. [79]

    Transactions of the Association for Computational Linguistics , volume=

    Break it down: A question understanding benchmark , author=. Transactions of the Association for Computational Linguistics , volume=. 2020 , publisher=

    2020

  64. [80]

    arXiv preprint arXiv:2009.00751 , year=

    Text modular networks: Learning to decompose tasks in the language of existing models , author=. arXiv preprint arXiv:2009.00751 , year=

    Pith/arXiv arXiv 2009

  65. [81]

    arXiv preprint arXiv:2305.14283 , year=

    Query rewriting for retrieval-augmented large language models , author=. arXiv preprint arXiv:2305.14283 , year=

    Pith/arXiv arXiv

  66. [82]

    arXiv preprint arXiv:2303.07678 , year=

    Query2doc: Query expansion with large language models , author=. arXiv preprint arXiv:2303.07678 , year=

    Pith/arXiv arXiv

  67. [83]

    2025 , url=

    Chen, Zehui and Liu, Kuikun and Wang, Qiuchen and Liu, Jiangning and Zhang, Wenwei and Chen, Kai and Zhao, Feng , booktitle=. 2025 , url=

    2025

  68. [84]

    arXiv preprint arXiv:2402.12052 , year=

    Small Models, Big Insights: Leveraging Slim Proxy Models To Decide When and What to Retrieve for LLMs , author=. arXiv preprint arXiv:2402.12052 , year=

    Pith/arXiv arXiv

  69. [85]

    arXiv preprint arXiv:2310.06839 , year=

    Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression , author=. arXiv preprint arXiv:2310.06839 , year=

    Pith/arXiv arXiv

  70. [86]

    arXiv preprint arXiv:2310.06201 , year=

    Compressing context to enhance inference efficiency of large language models , author=. arXiv preprint arXiv:2310.06201 , year=

    Pith/arXiv arXiv

  71. [87]

    arXiv preprint arXiv:2305.06983 , year=

    Active retrieval augmented generation , author=. arXiv preprint arXiv:2305.06983 , year=

    Pith/arXiv arXiv

  72. [88]

    arXiv preprint arXiv:2403.13372 , year=

    Llamafactory: Unified efficient fine-tuning of 100+ language models , author=. arXiv preprint arXiv:2403.13372 , year=

    Pith/arXiv arXiv

  73. [89]

    arXiv preprint arXiv:2106.09685 , year=

    Lora: Low-rank adaptation of large language models , author=. arXiv preprint arXiv:2106.09685 , year=

    Pith/arXiv arXiv

  74. [90]

    arXiv preprint arXiv:2410.20753 , year=

    PlanxRAG: Planning-guided Retrieval Augmented Generation , author=. arXiv preprint arXiv:2410.20753 , year=

    Pith/arXiv arXiv

  75. [91]

    arXiv preprint arXiv:2004.04906 , year=

    Dense passage retrieval for open-domain question answering , author=. arXiv preprint arXiv:2004.04906 , year=

    Pith/arXiv arXiv 2004

  76. [92]

    2023 , eprint=

    C-Pack: Packaged Resources To Advance General Chinese Embedding , author=. 2023 , eprint=

    2023

  77. [93]

    arXiv preprint arXiv:2411.02937 , year=

    Benchmarking Multimodal Retrieval Augmented Generation with Dynamic VQA Dataset and Self-adaptive Planning Agent , author=. arXiv preprint arXiv:2411.02937 , year=

    Pith/arXiv arXiv

  78. [94]

    BSharedRAG: Backbone Shared Retrieval-Augmented Generation for the E-commerce Domain , booktitle =

    Kaisi Guan and Qian Cao and Yuchong Sun and Xiting Wang and Ruihua Song , editor =. BSharedRAG: Backbone Shared Retrieval-Augmented Generation for the E-commerce Domain , booktitle =. 2024 , url =

    2024

  79. [95]

    CoRR , volume =

    Jason Wei and Zhiqing Sun and Spencer Papay and Scott McKinney and Jeffrey Han and Isa Fulford and Hyung Won Chung and Alex Tachard Passos and William Fedus and Amelia Glaese , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2504.12516 , eprinttype =. 2504.12516 , timestamp =

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.12516 2025

  80. [96]

    2023 , eprint=

    GAIA: a benchmark for General AI Assistants , author=. 2023 , eprint=

    2023

Showing first 80 references.