Pith. sign in

REVIEW 4 major objections 5 minor 218 references

Decoupling page selection from evidence extraction, with fetched pages kept in a per-question workspace, leads BrowseComp accuracy on every backbone tested.

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 15:09 UTC pith:M4HSGHVC

load-bearing objection Persistence as a design axis is worth taking seriously, but best-of-three reporting on 200 questions leaves the 2–4.5 pp BrowseComp margins inside noise. the 4 major comments →

arxiv 2608.02097 v1 pith:M4HSGHVC submitted 2026-08-03 cs.AI cs.IR

Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents

Qi Liu , Yiqun Chen , Zidan Chen , Yan Gao , Yi Wu , Yao Hu , Jiaxin Mao , Fengbin Zhu
show 1 more author
Tat-Seng Chua
This is my paper
classification cs.AI cs.IR
keywords search agentsdeep researchdocument interfacepersistent workspaceFetch-then-ExploreBrowseCompWideSearchReAct
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.

The paper argues that how a search agent reads a page matters as much as how it finds one, and that the dominant interfaces—visit-and-read and transient browsing—both discard the page after the moment it is opened. It proposes Fetch-then-Explore, a tool profile where fetching a page records its full text in a per-question filesystem workspace, and evidence is pulled later via grep and read. Under a fixed ReAct harness and search backend, this profile leads BrowseComp accuracy at all three backbones and generally matches or exceeds baselines on WideSearch. Behavioral analysis attributes the gain to returning to pages after leaving them, which transient interfaces rarely do. If correct, this shows persistence—not heavier fetching or bigger context—is a productive axis for long-horizon search agents.

Core claim

The paper's central claim is that separating the act of selecting a page from the act of extracting evidence, and retaining selected pages in an external per-question workspace, improves end-task accuracy over interfaces that fuse extraction into fetch time or hold a single transient page. In a controlled comparison holding harness, search backend, and backbone fixed, the Fetch-then-Explore profile achieves BrowseComp accuracies of 44.5, 52.5, and 70.5 across three backbones, ahead of the second-best profile by 2.0 to 4.5 percentage points, and matches or exceeds baselines on WideSearch Row-F1. The defining behavior is revisiting: 12–20% of documents are re-queried on BrowseComp and 67–76% o

What carries the argument

The central mechanism is the per-question file-system workspace C_q, together with the tool profile {fetch, grep, read, list_fetched}. fetch stores the full normalized page body off-context and returns only a header; grep performs regex search across the whole cache or a named file; read pulls bounded line windows, with grep's line numbers plugging directly into read offsets. This decouples selection from extraction and makes extraction deferred and repeatable, while nothing is evicted. The ablation confining grep to a single file shows the workspace pays only when queries can span the whole cache.

Load-bearing premise

The reported accuracy gains rest on comparing the best of three runs per cell without variance or significance tests, on a fixed 200-question sample from a live, changing web; if those margins are within run-to-run variation or web drift, the central claim does not hold.

What would settle it

Re-run the same five profiles on a fixed snapshot of the 200 BrowseComp questions with 10 seeds per cell and compute per-profile accuracy distributions; the claim would be falsified if the Fetch-then-Explore advantage over the second-best profile shrinks to overlapping confidence intervals, or if a single-page transient baseline matches it when given the same total turn budget.

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

If this is right

  • Search-agent systems can improve accuracy by changing only the document-access toolset, without retraining the policy or changing the search backend.
  • Document access and search behave as substitutes: agents with a persistent workspace issue fewer search calls, so retention may reduce cost as well as improve accuracy.
  • Evaluation of search agents should measure re-querying of held pages, not just fetch counts, since the revisit rate is the behavior that separates the winning interface.
  • The benefit concentrates on single-fact multi-hop questions (BrowseComp) rather than broad-table tasks (WideSearch), where fetch-time summaries already achieve high coverage.
  • Extensions the paper names include layering cross-trajectory memory on the same substrate and enriching the cache with entity/structure indexes to surface candidates that lexical grep cannot.

Where Pith is reading between the lines

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

  • Inference: The best-of-three reporting without variance means the exact margins may not replicate; a natural next step is to run the comparison with many seeds and report confidence intervals on a fixed snapshot of the web.
  • Inference: The workspace is per-question, but the same substrate could hold pages across questions; the paper's own extension list points there, suggesting the mechanism's value may grow with cross-question accumulation.
  • Inference: Cache-wide grep is rarely used but essential—removing it costs as much as dropping grep entirely—so fine-tuning the policy to issue whole-cache greps may yield gains beyond the reported interface-only effect.
  • Inference: The FTE-without-cache-scope ablation implies that retaining pages without cross-page query scope can hurt accuracy relative to transient browsing; persistence and scope are coupled, and designs that add persistence alone may see no benefit.

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 proposes Fetch-then-Explore (FTE), a document-access interface for search agents that decouples page selection from evidence extraction by storing fetched pages in a per-question filesystem workspace and querying them later with grep/read tools, rather than injecting page content into the context at fetch time or holding a single transient page. The authors instantiate this interface in a unified ReAct harness with a fixed search backend and compare it against four baselines (snippet-only, two visit-and-read variants, and session-based browsing) on BrowseComp and WideSearch across three LLM backbones. They report that FTE leads BrowseComp accuracy at every backbone, generally matches or exceeds baselines on WideSearch, and that behavioral analysis attributes the gains to returning to pages after leaving them—a behavior the persistent workspace enables far more than transient interfaces. The paper also includes ablations of the extraction tools and a discussion of the design space.

Significance. If the central claim holds, the paper identifies a simple and broadly applicable design axis—persistent, externally stored document access—that improves end-task accuracy in long-horizon search agents. The controlled harness, fixed search backend, and multiple backbones are strengths, as is the attempt to isolate the document interface from other design choices. The tool-call analysis and ablations provide a useful behavioral characterization. However, the empirical foundation is currently thin: the headline margins are small relative to the measurement noise under the stated protocol, and the paper's own WideSearch results are mixed. The manuscript would be strengthened substantially by releasing per-run, per-question results and by reporting variance or confidence intervals, as well as by providing the missing BrowseComp-Plus results referenced in the text.

major comments (4)
  1. [§4.1, Table 2] The central claim rests on best-of-three reporting over a fixed 200-question BrowseComp sample with no error bars or significance tests. The decisive FTE-vs-OPEN+FIND margins are +2.0 to +4.5 pp, i.e., 4 to 9 correct answers out of 200; the paired-difference standard error for 200 binary outcomes is roughly 5 pp, so no individual margin is reliable, and three same-signed margins are only weakly informative. Best-of-three selection is an additional risk: if FTE's per-run variance differs from baselines, taking the max can inflate FTE's apparent advantage even under equal means. Please report per-run means, standard deviations, and ideally paired or bootstrap confidence intervals, and release per-question run-level data.
  2. [§4.2, WideSearch results] The abstract and conclusion claim FTE 'generally matches or exceeds' baselines on WideSearch, but Table 2 shows FTE trailing VISIT-SUMMARY by 1.2 Row-F1 on Qwen3.5-35B-A3B and leading by only 0.1 on DeepSeek-V4-Pro. Given the lack of variance estimates, these differences are within noise. The behavioral claim in §4.3 that revisits drive gains is also harder to sustain on WideSearch: FTE revisits 67–76% of documents, yet the end-task margin over VISIT-SUMMARY is essentially zero. Please temper the claim or provide statistical support for the stated 'generally matches or exceeds.'
  3. [§4.3, BrowseComp-Plus analysis] The text states: 'On the closed corpus, document-level supervision confirms that the returned-to pages are the gold ones and the dissociation holds at line resolution (Appendix??).' No such appendix content or result is present. The BrowseComp-Plus evaluation is also referenced in §4.1 as '(Appendix??)' but no results are reported. This is a missing piece of evidence for the attribution argument, which is load-bearing for the paper's behavioral explanation. Please either include the full BrowseComp-Plus results or remove the claims that depend on them.
  4. [§4.3, FTE\x-cache ablation] The ablation confines grep to a single already-fetched page and reports that accuracy falls below OPEN+FIND on BrowseComp (39.5 vs 42.0 for Qwen-35B, 45.5 vs 50.5 for Qwen-122B in the text). The authors interpret this as evidence that persistence and cache-wide scope are two halves of one mechanism. However, the arm is implemented by refusing cache-wide patterns, which adds a retry cost and changes the action surface; the authors acknowledge this as a lower bound. The interpretation is plausible but not uniquely identified: the drop could partly reflect the cost of refused calls rather than the value of joint addressability. Please discuss this confound more explicitly or provide a control that imposes the same retry cost without removing cache-wide scope.
minor comments (5)
  1. [Throughout] Several appendix cross-references are unresolved: 'Appendix??' appears in §4.1 and §4.3, and '§??' appears in §4.3 and Appendix B. Please fix these references.
  2. [§2] The related-work section contains a broken citation: 'without reliably improving answers (?)'. Please supply the missing reference.
  3. [Figure 2] The figure caption and diagram are helpful, but the contrast between FTE and OPEN+FIND would be clearer if the 'released on next open' behavior were annotated with an explicit failure example.
  4. [§4.1] The choice of temperature=1.0 with best-of-three reporting is not discussed. Since the sampling temperature is high, best-of-three may overstate expected performance; please justify the choice or report results under multiple seeds.
  5. [Table 2] The WideSearch Row-F1 and Item-F1 numbers appear without confidence intervals and without clarifying whether the same best-of-three run is used for both metrics. Please clarify the relationship between the reported metrics and the selected run.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical measurement against external benchmarks, not a derived quantity; no step reduces to its inputs.

full rationale

The paper makes no formal derivation whose conclusion is equivalent to an input. Its central claim—that Fetch-then-Explore leads BrowseComp accuracy at every backbone and generally matches or exceeds baselines on WideSearch—is an empirical comparison measured against two external open-web benchmarks (BrowseComp, WideSearch) with official gold answers and official evaluation protocols. The five tool profiles are experimental conditions; FTE's tools (fetch/grep/read/list_fetched) are defined independently of the reported accuracy ordering, and none of the reported numbers is computed from a fitted parameter, a self-citation, or a uniqueness theorem. The behavioral analysis (revisit rate, tool-call n-grams, fetch reuse) is measured on the same reported runs and is presented as a post-hoc explanation, not as the evidence for the accuracy claim; 'returning to a page' is operationally defined for every profile and is not the same quantity as end-task accuracy, so the explanation is not definitionally equivalent to the result. The judge model overlaps with one agent backbone, but the paper explicitly notes this applies identically to all profiles in that row, so it cannot manufacture intra-backbone deltas. A few references (e.g., WideSearch, the deep-research survey) include present co-authors, but they are used only to situate the work and are not load-bearing. Best-of-three reporting and the absence of error bars are measurement-protocol risks, not circularity. No circular step is identified; the score of 1 reflects only the presence of minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests primarily on implementation choices (tool bounds, step budget, benchmark subsample) and on domain assumptions about the harness, the judge, and the stability of the live web. No physics-style invented entities are introduced; the per-question workspace is a software artifact with external, falsifiable consequences. The most load-bearing assumptions are the reliability of LLM judging and the representativeness of a best-of-three measurement on a noisy live environment.

free parameters (3)
  • tool output bounds = GREP_MAX_MATCHES=20; read limit 200 lines; default grep context 3 lines; 50 KB output cap
    Chosen by hand and held fixed across all runs; these bounds shape how much evidence the agent sees per call and could affect accuracy if varied, though the paper does not tune them.
  • per-question step budget = 300 turns
    A resource limit chosen for all profiles; the number of search and extraction calls is bounded by this budget, and different budgets could change the ranking.
  • BrowseComp subsample = 200 of 1,266 questions
    A fixed random sample drawn under resource constraints; the accuracy numbers are estimates on this subset, and the paper does not report the sampling seed or variance.
axioms (4)
  • domain assumption ReAct loop with native tool calling is a faithful harness for search agents
    All profiles run in the same ReAct agent, so conclusions generalize only if this harness represents how deployed search agents work.
  • domain assumption LLM-as-a-judge scoring on BrowseComp and WideSearch is a reliable measure of answer quality
    Accuracy and F1 are determined by an LLM judge rather than a deterministic match; the judge overlaps with one agent backbone, and the paper assumes this does not distort intra-backbone deltas.
  • domain assumption Serper search results and crawl4ai page rendering are representative and stable enough across runs
    The open web changes between the three runs of each cell; the paper treats the live environment as a fixed benchmark, and the best-of-three protocol may absorb content drift.
  • domain assumption Backbones can learn to use grep and read from JSON schemas alone
    No per-profile prompting or tool-use demonstrations are given; the behavioral data show the weakest backbone pages via read more than grep, indicating a capability-interface interaction that the design takes for granted.
invented entities (1)
  • per-question filesystem workspace (C_q) independent evidence
    purpose: Stores fetched page bodies off-context so grep/read can extract evidence on demand and pages remain queryable after the agent moves on to other documents
    The workspace's predicted effects, higher revisit rates and higher end-task accuracy on external benchmarks, are falsifiable and measurable outside the paper; it is the only intervention that distinguishes FTE from the OPEN+FIND baseline.

pith-pipeline@v1.3.0-daily-deepseek · 20349 in / 11270 out tokens · 81102 ms · 2026-08-04T15:09:00.787886+00:00 · methodology

0 comments
read the original abstract

Search agents now answer questions that take dozens of searches to settle, yet how such an agent reads a page has drawn far less attention than how it finds one. Nearly all of them use one of two document interfaces, and both tie a page to the moment it is opened. \emph{Visit-and-read} injects a reading of the page into the message history at fetch time, fixing that reading before the agent knows which fact it will need. Stateful \emph{browsing} instead extracts on demand from the page in hand, but holds one page at a time and releases it as soon as the agent opens another. Either way, a page that turns out to matter many turns later has to be fetched and rendered into context all over again. We propose \textbf{Fetch-then-Explore}, which separates page selection from evidence extraction and keeps what it selects: pages are recorded in a per-question workspace on the filesystem rather than the context window or a transient session, and evidence is pulled from them on demand later. Selection becomes almost free, extraction can wait until the agent knows what to look for and be repeated as its hypothesis sharpens, and pages are not released when the agent moves on, so evidence accumulates across the trajectory. In a unified ReAct harness with fixed search, we compare Fetch-then-Explore against snippet-only, visit-and-read, and browsing baselines on two open-web benchmarks, BrowseComp and WideSearch, across three agent backbones. It leads BrowseComp accuracy at every backbone and generally matches or exceeds the baselines on WideSearch, and a behavioral analysis traces the gains to the workspace's defining move: returning to a page after leaving it, which it does far more than any transient interface, so evidence missed on a first pass can still be recovered later.

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

218 extracted references · 140 linked inside Pith

  1. [1]

    2026 , archivePrefix =

    VideoSearch-R1: Iterative Video Retrieval and Reasoning via Soft Query Refinement , author =. 2026 , archivePrefix =. 2607.00446 , url =

    Pith/arXiv arXiv 2026

  2. [2]

    2026 , archivePrefix =

    Multi-Turn Agentic Scientific Literature Search via Workflow Induction , author =. 2026 , archivePrefix =. 2607.00597 , url =

    Pith/arXiv arXiv 2026

  3. [3]

    2026 , archivePrefix =

    Can LLM-as-a-Judge Reliably Verify Rubrics in Agentic Scenarios? , author =. 2026 , archivePrefix =. 2606.29920 , url =

    Pith/arXiv arXiv 2026

  4. [4]

    2026 , archivePrefix =

    KbSD: Knowledge Boundary aware Self-Distillation for Behavioral Calibration in Agentic Search , author =. 2026 , archivePrefix =. 2606.29863 , url =

    Pith/arXiv arXiv 2026

  5. [5]

    2026 , archivePrefix =

    When Search Agents Should Ask: DiscoBench for Clarification-Aware Deep Search , author =. 2026 , archivePrefix =. 2606.27669 , url =

    Pith/arXiv arXiv 2026

  6. [6]

    2026 , archivePrefix =

    ProMSA:Progressive Multimodal Search Agents for Knowledge-Based Visual Question Answering , author =. 2026 , archivePrefix =. 2606.27974 , url =

    Pith/arXiv arXiv 2026

  7. [7]

    2026 , archivePrefix =

    Lacuna: A Research Map for Machine Learning , author =. 2026 , archivePrefix =. 2606.26246 , url =

    Pith/arXiv arXiv 2026

  8. [8]

    2026 , archivePrefix =

    MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG , author =. 2026 , archivePrefix =. 2606.26793 , url =

    Pith/arXiv arXiv 2026

  9. [9]

    2026 , archivePrefix =

    Heuresis: Search Strategies for Autonomous AI Research Agents Across Quality, Diversity and Novelty , author =. 2026 , archivePrefix =. 2606.25198 , url =

    Pith/arXiv arXiv 2026

  10. [10]

    2026 , archivePrefix =

    AGORA: An Archive-Grounded Benchmark for Agentic Workplace Document Reasoning , author =. 2026 , archivePrefix =. 2606.24526 , url =

    arXiv 2026

  11. [11]

    2026 , archivePrefix =

    Breaking the Evaluation Paradox: Evaluating High-Entropy Search with Computationally Irreducible Constraints , author =. 2026 , archivePrefix =. 2606.22783 , url =

    Pith/arXiv arXiv 2026

  12. [12]

    2026 , archivePrefix =

    PaperClaw: Harnessing Agents for Autonomous Research and Human-in-the-Loop Refinement , author =. 2026 , archivePrefix =. 2606.22610 , url =

    Pith/arXiv arXiv 2026

  13. [13]

    2026 , archivePrefix =

    Novelty-Aware Agentic Retrieval: Comparing Research Contributions Through Structured Multi-Step Reasoning , author =. 2026 , archivePrefix =. 2606.22151 , url =

    Pith/arXiv arXiv 2026

  14. [14]

    2026 , archivePrefix =

    Agentic Symbolic Search: Characterizing PDEs Beyond Hand-crafted Expressions, Meshes, and Neural Networks , author =. 2026 , archivePrefix =. 2606.20467 , url =

    Pith/arXiv arXiv 2026

  15. [15]

    2026 , archivePrefix =

    ScholarQuest: A Taxonomy-Guided Benchmark for Agentic Academic Paper Search in Open Literature Environments , author =. 2026 , archivePrefix =. 2606.20235 , url =

    Pith/arXiv arXiv 2026

  16. [16]

    2026 , archivePrefix =

    MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval , author =. 2026 , archivePrefix =. 2606.18508 , url =

    Pith/arXiv arXiv 2026

  17. [17]

    2026 , archivePrefix =

    Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search , author =. 2026 , archivePrefix =. 2606.17209 , url =

    Pith/arXiv arXiv 2026

  18. [18]

    2026 , archivePrefix =

    DRFLOW: A Deep Research Benchmark for Personalized Workflow Prediction , author =. 2026 , archivePrefix =. 2606.18191 , url =

    Pith/arXiv arXiv 2026

  19. [19]

    2026 , archivePrefix =

    Beyond Monolingual Deep Research: Evaluating Agents and Retrievers with Cross-Lingual BrowseComp-Plus , author =. 2026 , archivePrefix =. 2606.15345 , url =

    Pith/arXiv arXiv 2026

  20. [20]

    2026 , archivePrefix =

    S1-DeepResearch: Beyond Search, Toward Real-World Long-Horizon Research Agents , author =. 2026 , archivePrefix =. 2606.15367 , url =

    arXiv 2026

  21. [21]

    2026 , archivePrefix =

    DEEPRUBRIC: Evidence-Tree Rubric Supervision for Efficient Reinforcement Learning of Deep Research Agents , author =. 2026 , archivePrefix =. 2606.17029 , url =

    arXiv 2026

  22. [22]

    2026 , archivePrefix =

    SANA: What Matters for QA Agents over Massive Data Lakes? , author =. 2026 , archivePrefix =. 2606.13904 , url =

    arXiv 2026

  23. [23]

    2026 , archivePrefix =

    Self-Evolving Deep Research via Joint Generation and Evaluation , author =. 2026 , archivePrefix =. 2606.04507 , url =

    Pith/arXiv arXiv 2026

  24. [24]

    2026 , archivePrefix =

    ARBOR: Online Process Rewards via a Reusable Rubric Buffer for Search Agents , author =. 2026 , archivePrefix =. 2606.03239 , url =

    Pith/arXiv arXiv 2026

  25. [25]

    2026 , archivePrefix =

    Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses , author =. 2026 , archivePrefix =. 2606.02373 , url =

    Pith/arXiv arXiv 2026

  26. [26]

    2026 , archivePrefix =

    TVIR: Building Deep Research Agents Towards Text-Visual Interleaved Report Generation , author =. 2026 , archivePrefix =. 2606.02320 , url =

    Pith/arXiv arXiv 2026

  27. [27]

    2026 , archivePrefix =

    Deep Research as Rubric for Reinforcement Learning , author =. 2026 , archivePrefix =. 2606.01091 , url =

    Pith/arXiv arXiv 2026

  28. [28]

    2026 , archivePrefix =

    Masking Stale Observations Helps Search Agents -- Until It Doesn&\#39;t: A Regime Map and Its Mechanism , author =. 2026 , archivePrefix =. 2606.00408 , url =

    Pith/arXiv arXiv 2026

  29. [29]

    2026 , archivePrefix =

    LongTraceRL: Learning Long-Context Reasoning from Search Agent Trajectories with Rubric Rewards , author =. 2026 , archivePrefix =. 2605.31584 , url =

    Pith/arXiv arXiv 2026

  30. [30]

    2026 , archivePrefix =

    AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle , author =. 2026 , archivePrefix =. 2605.31468 , url =

    Pith/arXiv arXiv 2026

  31. [31]

    2026 , archivePrefix =

    COMPASS: Cognitive MCTS-Guided Process Alignment for Safe Search Agents , author =. 2026 , archivePrefix =. 2605.30838 , url =

    Pith/arXiv arXiv 2026

  32. [32]

    2026 , archivePrefix =

    MosaicLeaks:Privacy Risks in Querying-in-the-Open for Deep Research Agents , author =. 2026 , archivePrefix =. 2605.30727 , url =

    Pith/arXiv arXiv 2026

  33. [33]

    2026 , archivePrefix =

    Planner-Centric Reinforcement Learning for Deep Research with Structure-Aware Reward , author =. 2026 , archivePrefix =. 2605.30824 , url =

    Pith/arXiv arXiv 2026

  34. [34]

    2026 , archivePrefix =

    LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis , author =. 2026 , archivePrefix =. 2605.30434 , url =

    Pith/arXiv arXiv 2026

  35. [35]

    2026 , archivePrefix =

    Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness for Interleaved Report Generation , author =. 2026 , archivePrefix =. 2605.29861 , url =

    Pith/arXiv arXiv 2026

  36. [36]

    2026 , archivePrefix =

    SAAS: Self-Aware Reinforcement Learning for Over-Search Mitigation in Agentic Search , author =. 2026 , archivePrefix =. 2605.29796 , url =

    Pith/arXiv arXiv 2026

  37. [37]

    2026 , archivePrefix =

    Beyond Trajectory Rewards: Step-level Credit Assignment for Agentic Search via Graph Modeling , author =. 2026 , archivePrefix =. 2605.29697 , url =

    Pith/arXiv arXiv 2026

  38. [39]

    2026 , archivePrefix =

    LiveBrowseComp: Are Search Agents Searching, or Just Verifying What They Already Know? , author =. 2026 , archivePrefix =. 2605.28721 , url =

    Pith/arXiv arXiv 2026

  39. [40]

    2026 , archivePrefix =

    Plan Before Search: Search Agents Need Plan , author =. 2026 , archivePrefix =. 2605.28354 , url =

    Pith/arXiv arXiv 2026

  40. [41]

    2026 , archivePrefix =

    Retrieval, Reward, and Training Protocols: What Matters in Training Search Agents? , author =. 2026 , archivePrefix =. 2605.27881 , url =

    Pith/arXiv arXiv 2026

  41. [42]

    2026 , archivePrefix =

    AI Research Agents Narrow Scientific Exploration , author =. 2026 , archivePrefix =. 2605.27905 , url =

    Pith/arXiv arXiv 2026

  42. [43]

    2026 , archivePrefix =

    ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence , author =. 2026 , archivePrefix =. 2605.26340 , url =

    Pith/arXiv arXiv 2026

  43. [44]

    2026 , archivePrefix =

    VeriTrace: Evolving Mental Models for Deep Research Agents , author =. 2026 , archivePrefix =. 2605.26081 , url =

    Pith/arXiv arXiv 2026

  44. [45]

    2026 , archivePrefix =

    Can LLMs Time Travel? Enhancing Temporal Consistency in Legal Agentic Search through Reinforcement Learning , author =. 2026 , archivePrefix =. 2605.25920 , url =

    Pith/arXiv arXiv 2026

  45. [46]

    2026 , archivePrefix =

    An Interactive Paradigm for Deep Research , author =. 2026 , archivePrefix =. 2605.24266 , url =

    Pith/arXiv arXiv 2026

  46. [47]

    2026 , archivePrefix =

    QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks , author =. 2026 , archivePrefix =. 2605.24218 , url =

    Pith/arXiv arXiv 2026

  47. [48]

    2026 , archivePrefix =

    Design and Report Benchmarks for Knowledge Work , author =. 2026 , archivePrefix =. 2605.23262 , url =

    Pith/arXiv arXiv 2026

  48. [49]

    2026 , archivePrefix =

    SGR-Bench: Benchmarking Search Agents on State-Gated Retrieval , author =. 2026 , archivePrefix =. 2605.22219 , url =

    Pith/arXiv arXiv 2026

  49. [50]

    2026 , archivePrefix =

    SciAtlas: A Large-Scale Knowledge Graph for Automated Scientific Research , author =. 2026 , archivePrefix =. 2605.22878 , url =

    Pith/arXiv arXiv 2026

  50. [51]

    2026 , archivePrefix =

    DeepWeb-Bench: A Deep Research Benchmark Demanding Massive Cross-Source Evidence and Long-Horizon Derivation , author =. 2026 , archivePrefix =. 2605.21482 , url =

    Pith/arXiv arXiv 2026

  51. [52]

    2026 , archivePrefix =

    AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration , author =. 2026 , archivePrefix =. 2605.20025 , url =

    Pith/arXiv arXiv 2026

  52. [53]

    2026 , archivePrefix =

    ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning , author =. 2026 , archivePrefix =. 2605.20176 , url =

    Pith/arXiv arXiv 2026

  53. [54]

    2026 , archivePrefix =

    Time to REFLECT: Can We Trust LLM Judges for Evidence-based Research Agents? , author =. 2026 , archivePrefix =. 2605.19196 , url =

    Pith/arXiv arXiv 2026

  54. [55]

    2026 , archivePrefix =

    Argus: Evidence Assembly for Scalable Deep Research Agents , author =. 2026 , archivePrefix =. 2605.16217 , url =

    Pith/arXiv arXiv 2026

  55. [56]

    2026 , archivePrefix =

    Is Grep All You Need? How Agent Harnesses Reshape Agentic Search , author =. 2026 , archivePrefix =. 2605.15184 , url =

    Pith/arXiv arXiv 2026

  56. [57]

    2026 , archivePrefix =

    EcoGEO: Trajectory-Aware Evidence Ecosystems for Web-Enabled LLM Search Agents , author =. 2026 , archivePrefix =. 2605.12887 , url =

    Pith/arXiv arXiv 2026

  57. [58]

    2026 , archivePrefix =

    AgentDisCo: Towards Disentanglement and Collaboration in Open-ended Deep Research Agents , author =. 2026 , archivePrefix =. 2605.11732 , url =

    Pith/arXiv arXiv 2026

  58. [59]

    2026 , archivePrefix =

    CuSearch: Curriculum Rollout Sampling via Search Depth for Agentic RAG , author =. 2026 , archivePrefix =. 2605.11611 , url =

    Pith/arXiv arXiv 2026

  59. [60]

    2026 , archivePrefix =

    WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation , author =. 2026 , archivePrefix =. 2605.10912 , url =

    Pith/arXiv arXiv 2026

  60. [61]

    2026 , archivePrefix =

    AutoLLMResearch: Training Research Agents for Automating LLM Experiment Configuration - Learning from Cheap, Optimizing Expensive , author =. 2026 , archivePrefix =. 2605.11518 , url =

    Pith/arXiv arXiv 2026

  61. [62]

    2026 , archivePrefix =

    Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient? , author =. 2026 , archivePrefix =. 2605.10848 , url =

    Pith/arXiv arXiv 2026

  62. [63]

    2026 , archivePrefix =

    RubricEM: Meta-RL with Rubric-guided Policy Decomposition beyond Verifiable Rewards , author =. 2026 , archivePrefix =. 2605.10899 , url =

    Pith/arXiv arXiv 2026

  63. [64]

    2026 , archivePrefix =

    Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents , author =. 2026 , archivePrefix =. 2605.10832 , url =

    Pith/arXiv arXiv 2026

  64. [65]

    2026 , archivePrefix =

    Personalized Deep Research: A User-Centric Framework, Dataset, and Hybrid Evaluation for Knowledge Discovery , author =. 2026 , archivePrefix =. 2605.10530 , url =

    Pith/arXiv arXiv 2026

  65. [66]

    2026 , archivePrefix =

    Hypothesis-Driven Deep Research with Large Language Models: A Structured Methodology for Automated Knowledge Discovery , author =. 2026 , archivePrefix =. 2605.10224 , url =

    Pith/arXiv arXiv 2026

  66. [67]

    2026 , archivePrefix =

    LoHoSearch: Benchmarking Long-Horizon Search Agents Beyond the Human Difficulty Ceiling , author =. 2026 , archivePrefix =. 2606.12837 , url =

    Pith/arXiv arXiv 2026

  67. [68]

    2026 , archivePrefix =

    EvoBrowseComp: Benchmarking Search Agents on Evolving Knowledge , author =. 2026 , archivePrefix =. 2606.13120 , url =

    Pith/arXiv arXiv 2026

  68. [69]

    2026 , archivePrefix =

    DailyReport: An Open-ended Benchmark for Evaluating Search Agents on Daily Search Tasks , author =. 2026 , archivePrefix =. 2606.12871 , url =

    Pith/arXiv arXiv 2026

  69. [70]

    2026 , archivePrefix =

    Toward Generalist Autonomous Research via Hypothesis-Tree Refinement , author =. 2026 , archivePrefix =. 2606.11926 , url =

    Pith/arXiv arXiv 2026

  70. [71]

    2026 , archivePrefix =

    Notes2Skills: From Lab Notebooks to Certainty-Aware Scientific Agent Skills , author =. 2026 , archivePrefix =. 2606.11897 , url =

    Pith/arXiv arXiv 2026

  71. [72]

    2026 , archivePrefix =

    Can AI Agents Synthesize Scientific Conclusions? , author =. 2026 , archivePrefix =. 2606.11337 , url =

    Pith/arXiv arXiv 2026

  72. [73]

    2026 , archivePrefix =

    FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents , author =. 2026 , archivePrefix =. 2606.12087 , url =

    Pith/arXiv arXiv 2026

  73. [74]

    2026 , archivePrefix =

    Search Discipline for Long-Horizon Research Agents , author =. 2026 , archivePrefix =. 2606.11522 , url =

    Pith/arXiv arXiv 2026

  74. [75]

    2026 , archivePrefix =

    Moonshine: An Autonomous Mathematical Research Agent Centered on Conjecture Generation , author =. 2026 , archivePrefix =. 2606.10806 , url =

    Pith/arXiv arXiv 2026

  75. [76]

    2026 , archivePrefix =

    Effective Reinforcement Learning for Agentic Search by Recycling Zero-Variance Queries During Training , author =. 2026 , archivePrefix =. 2606.10709 , url =

    Pith/arXiv arXiv 2026

  76. [77]

    2026 , archivePrefix =

    What Fits (Into Few Tokens) Doesn&\#39;t Overfit: Compression and Generalization in ML Research Agents , author =. 2026 , archivePrefix =. 2606.11045 , url =

    Pith/arXiv arXiv 2026

  77. [78]

    2026 , archivePrefix =

    PBSD: Privileged Bayesian Self-Distillation for Long-Horizon Credit Assignment , author =. 2026 , archivePrefix =. 2606.09348 , url =

    Pith/arXiv arXiv 2026

  78. [79]

    2026 , archivePrefix =

    SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research , author =. 2026 , archivePrefix =. 2606.09730 , url =

    Pith/arXiv arXiv 2026

  79. [80]

    2026 , archivePrefix =

    Multi-Turn Evaluation of Deep Research Agents Under Process-Level Feedback , author =. 2026 , archivePrefix =. 2606.09748 , url =

    Pith/arXiv arXiv 2026

  80. [81]

    2026 , archivePrefix =

    SlimSearcher: Training Efficiency-Aware Web Agents via Adaptive Reward Gating , author =. 2026 , archivePrefix =. 2606.07074 , url =

    Pith/arXiv arXiv 2026

Showing first 80 references.