Pith. sign in

REVIEW 4 major objections 8 minor 19 references

A deep-research agent that recursively verifies its provisional answers and uses unresolved constraints to target the next round of search outperforms far larger systems on deep-search benchmarks.

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 07:21 UTC pith:CVVJFDHX

load-bearing objection Strong systems paper with a genuinely useful architecture, but the headline ablation is confounded and the confidence threshold is undisclosed. the 4 major comments →

arxiv 2607.21461 v2 pith:CVVJFDHX submitted 2026-07-23 cs.AI

AREX: Towards a Recursively Self-Improving Agent for Deep Research

Shuqi Lu , Chaofan Li , Kun Luo , Zhang Zhang , Hui Wang , Hongwang Xiao , Lei Xiong , Jiahao Wang
show 16 more authors
Sen Wang Xiyan Jiang Wanli Li Yuyang Hu Hongjin Qian Bingyu Yan Jianlyu Chen Ziyi Xia Yingxia Shao Kang Liu Zhicheng Dou Di He Chaozhuo Li Qiwei Ye Zhongyuan Wang Zheng Liu
This is my paper
classification cs.AI
keywords recursive self-improvementdeep research agentsdiscovery-verification asymmetryconstraint-wise verificationautonomous context updatinglong-horizon reinforcement learningmixture-of-expertskey-step supervision
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.

AREX is a family of recursive self-improving deep-research agents built on a discovery–verification asymmetry: finding an answer that satisfies many constraints is hard, but checking a candidate constraint-wise is comparatively cheap. The paper's core claim is that an agent should convert a provisional answer into a partially verified research state, then launch a new round of targeted research on whatever remains unresolved — rather than simply searching longer. Across BrowseComp, WideSearch, DeepSearchQA, GAIA, xbench-2510, and Humanity's Last Exam, the 122B-total/10B-active AREX-Base substantially outperforms comparable-scale baselines and remains competitive with far larger models. The load-bearing mechanistic result is that combining an inner research loop with an outer self-improvement loop and an autonomous context-update tool raises BrowseComp accuracy from 59.6 to 82.5.

Core claim

The central claim is that verification can serve as the transition operator between research rounds, not merely a final filter. AREX alternates an inner loop that gathers evidence and produces a provisional answer with supporting evidence and a confidence score, and an outer loop that adopts a simple decision rule: accept if confidence exceeds a threshold, refine by preserving useful findings and targeting the unresolved constraints, or restart if the trajectory is judged too noisy. Sustaining this recursion over long horizons requires a learned context-update tool that compresses the interaction history into an improvement state preserving verified findings, unresolved constraints, rejected

What carries the argument

The central mechanism is the bi-level recursive loop: an inner research policy (search, browse, integrate, finish) and an outer self-improvement policy that audits the structured result constraint-wise and emits accept/refine/restart decisions (Equation 10). The second load-bearing component is the autonomous context-update tool, which the model invokes itself to condense its trajectory into a compact improvement state; the paper shows it is used proactively (80.3% of BrowseComp cases, usually well below the 128K-token limit) and that replacing the full history with this state alone yields 11.8 accuracy points. A third component is step-aware training: key-step detectors identify evidence di

Load-bearing premise

The load-bearing premise is that the model's self-reported confidence score (produced by the finish tool) genuinely reflects how likely the provisional answer is to be correct — yet the paper's own Figure 3 shows 33% of incorrect outputs with ACU falling in the 90–100 confidence band.

What would settle it

Run AREX on BrowseComp with the outer loop's refinement objective replaced by a randomly selected unresolved constraint from the task template (keeping the budget, tools, and training fixed); if accuracy stays near 82.5 rather than dropping toward the no-outer-loop level, the claim that targeted follow-up from unresolved constraints drives the gain would be falsified.

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

If this is right

  • If the recursive verification loop is the real driver, then deep-research agents should treat verification as an active control signal that decides when to continue, what to look for next, and when to stop — not just as a scoring step at the end.
  • Learned, self-invoked context management organized around the research state (verified findings, unresolved constraints, rejected paths) can outperform fixed truncation or external summarization for long-horizon tasks.
  • Concentrating supervision and RL credit on a small number of decision-critical steps (evidence discovery, redirection, context updates) is more effective than uniform token-level training when budgets are matched.
  • A 10B-active-parameter agent using this architecture can be competitive with models using substantially more activated parameters on several deep-research benchmarks.
  • The same double-loop structure could generalize to any task where a candidate solution can be decomposed into independently checkable constraints, including code and mathematical verification.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the confidence score's reliability is the fulcrum for the entire accept/refine/restart rule; the authors' own Figure 3 shows 33% of incorrect outputs with the context-update tool landing in the 90–100 confidence bin, so a robust deployment would need calibrated confidence or a richer stopping signal than a scalar threshold.
  • A natural extension the authors do not explore is using the update_context tool as a domain-agnostic pretrained module, decoupling state maintenance from the research policy; a testable prediction is that such a module would improve transfer to new long-horizon tasks.
  • Because the outer loop's refinement objective is derived from unresolved constraints, one can test whether the gain comes from the targeting itself or merely from having a second research round: replacing the targeted objective with a random unsolved constraint should, under the paper's claim, degrade accuracy.
  • The paper's framing suggests a broader recipe: any verification-aware agent that preserves 'what is established' and tracks 'what remains to check' can turn a limited inference budget into compounding progress, which may matter more as models become cheaper per token.

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 / 8 minor

Summary. AREX introduces a bi-level recursive research agent for deep research. The inner loop gathers evidence through search/browsing tools and periodically invokes a learned update_context tool that compresses the interaction history into a structured research state; it terminates with a provisional answer, supporting evidence, and a self-reported confidence score. The outer loop then either accepts the answer (confidence ≥ τ), refines it by preserving verified evidence and issuing a targeted next-round objective, or restarts from the original query. The paper trains dense 4B and 122B-A10B MoE variants using synthetic constraint-based tasks, filtered teacher trajectories, multi-stage agentic mid-training with key-step supervision, and step-aware reinforcement learning. It reports strong results on BrowseComp, WideSearch, DeepSearchQA, GAIA, xbench-2510, and HLE, and attributes the main gains to autonomous context updating (ACU) and the outer self-improvement loop in Table 3.

Significance. If the mechanism is validated, this is a meaningful contribution: it offers a concrete instantiation of recursive verification-driven refinement and shows that a 10B-active MoE can be competitive with much larger research agents on external benchmarks. The paper has real strengths: it evaluates on external benchmarks rather than only on in-house tasks, the ACU ablation in Table 3 is internally framed as a controlled comparison, Appendix B is honestly caveated, and the models are released. However, the central outer-loop ablation is confounded by compute and training-distribution differences, and the confidence-based control loop is underspecified because τ and τ_conf are never reported. Those issues must be addressed before the central mechanistic claim is established.

major comments (4)
  1. [§5.3, Table 3; §3.2, §4.1] The 11.1-point gain attributed to the outer loop (71.4→82.5 with ACU) is confounded. The w/ outer-loop condition may use up to 5 outer-loop operations (Sec. 5.1), while w/o outer loop accepts the first finish output; no matched turn/token budget is reported for the two rows. Moreover, the model was trained on trajectories that include outer-loop refinement (Sec. 3.2, 4.1), so the w/o row is a train/inference mismatch and may understate what a single-round model could achieve. The 22.9-point headline in Table 3 therefore cannot be assigned specifically to the recursive double-loop architecture. Please add a matched-budget single-round control and/or train a variant without outer-loop trajectories.
  2. [§2.3, Eq. (10); Eq. (17); §5.3, Fig. 3] The accept/refine/restart rule is governed by a threshold τ that is never reported, and the teacher-trajectory filter threshold τ_conf in Eq. (17) is also undisclosed. This makes the control loop non-reproducible and leaves open the possibility that gains were selected over a favorable threshold. The calibration evidence is weak: Fig. 3 shows that 33.0% of incorrect outputs with ACU have confidence in the 90–100 bin, so high confidence does not reliably indicate correctness. Please report τ and τ_conf, and provide a threshold sweep or calibration analysis.
  3. [§5.2, Table 1; Abstract] The abstract's claim that AREX 'substantially outperforms comparable-scale baselines' is not supported by the table: AREX-Base trails Kimi-K2.6 on BrowseComp (82.5 vs 83.2), trails MiroThinker-H1 on GAIA and xbench-2510, and many margins are within a few points. No error bars, significance tests, or multiple runs are reported, so 'substantially' is unjustified. In addition, the HLE block mixes full-HLE and text-only scores (per the footnote), making direct ranking within that block unreliable. Please report uncertainty estimates and separate the HLE protocols clearly.
  4. [§5.1, §5.2] The parameter-efficiency claim ('competitive with models using substantially more activated parameters') is not matched for inference-time compute. Deep-research performance scales with test-time budget; AREX is given a 300-turn/5-round ceiling, but no token or turn budgets are reported for the comparison systems. Without matched compute or at least reported budgets, the capability-to-parameter conclusion is not established. Please report average/median turns and tokens for AREX and the baselines, or add compute-matched comparisons.
minor comments (8)
  1. [§2.3, Eq. (9)] G_θ, the trajectory assessor, is never defined. Clarify whether it is the same policy, a separate model, and how it is trained.
  2. [§5.1] The text says '5 outer self-improvement loop operations' but elsewhere implies '5 recursive rounds.' State the exact maximum number of finish calls per episode.
  3. [§5.3, Fig. 3] Report the number of outputs in each confidence bin and add a reliability diagram/calibration curve; histograms alone overstate the separation between correct and incorrect outputs.
  4. [Table 2] The row labeled 'Upper bound 128,000' is ambiguous; label it as the active context-window size.
  5. [§3.2] The teacher models used to collect trajectories are not named. State which models generated the trajectories and how many trajectories were retained.
  6. [Appendix B] Define the acceptance criteria for 'accepted generated trajectories' in the self-distillation experiment.
  7. [§4.1] The rule-based key-step detectors are described only qualitatively; report the actual detector thresholds or at least their precision/recall on a validation set.
  8. [References] The citation 'Team et al. (2026)' is used ambiguously; disambiguate which team/paper is meant.

Circularity Check

0 steps flagged

No significant circularity: the central claims rest on external benchmarks and controlled ablations; the identified confounds are experimental-design concerns, not derivation-by-construction.

full rationale

The paper's load-bearing results are accuracies on external benchmarks (BrowseComp, WideSearch, DeepSearchQA, HLE) and within-system ablations. The ACU ablation compares the same trained model with and without context updating under a matched interaction budget, so it is a genuine controlled comparison rather than a construction-equivalent prediction. The outer-loop comparison in Table 3 is not fully matched—the w/ outer-loop condition can use additional research rounds and the model was trained on trajectories that include outer-loop transitions—but this is a methodological confound (compute and train/inference distribution mismatch), not a case where the claimed result is defined in terms of its inputs or where a fitted parameter is renamed as a prediction. The confidence-based decision rule (Eq. 10) uses a self-reported score, and the training-data filter (Eq. 17) is a separate threshold on teacher trajectories; the paper provides empirical calibration evidence in Fig. 3, so the control signal is not circular by construction even if the threshold τ is unreported. Key-step annotations are grounded in rule-based detectors and verifiable task structure, not in the model's own outputs, and the step-level loss analysis is a post-hoc diagnostic, not a training target. No load-bearing self-citation or imported uniqueness theorem appears. Therefore no circular step can be exhibited under the required standard.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on a substantial set of undisclosed hyperparameters (τ, λ_key, β, τ_conf, detector thresholds) and on domain assumptions about synthetic-task transfer and teacher-trajectory validity. The invented entities are internal components with no independent evidence outside the paper's own benchmark evaluations.

free parameters (5)
  • confidence threshold τ
    Eq. (10) accept/refine/restart depends on τ; no value reported; likely tuned to optimize benchmark accuracy.
  • key-step reward shaping strength λ_key
    Sec. 4.2 controls the auxiliary bonus for annotated key steps; value not reported.
  • KL penalty coefficient β
    Sec. 4.2 KL penalty against reference policy; value not reported.
  • teacher trajectory confidence filter τ_conf
    Eq. (17) removes trajectories with low teacher confidence; threshold not reported.
  • key-step detector thresholds
    Sec. 4.1 rule-based detectors for evidence discovery, path rejection, and context updates rely on undisclosed thresholds.
axioms (5)
  • domain assumption Constraint-wise verification is tractable for deep research tasks
    Sec. 1 discovery–verification asymmetry; assumes a valid answer's constraints can be checked independently. Not proven, but common in retrieval tasks.
  • domain assumption Synthetic tasks with automatically verified answers transfer to real deep-research benchmarks
    Sec. 3.1 training uses synthetic tasks generated from templates; transfer is not demonstrated outside the reported benchmarks.
  • domain assumption Teacher trajectories from unnamed strong teacher models provide valid supervision
    Sec. 3.2 teacher models are not named and trajectory validity filters are heuristic.
  • ad hoc to paper Self-reported confidence s^(k) is informative for correctness
    Eq. (8)/(10); Fig. 3 partially contradicts this (33% of errors at 90–100 confidence).
  • domain assumption Evaluation tool environment matches training tool environment
    Sec. 5.1 both use search/visit/update_context/finish, but APIs are unspecified.
invented entities (2)
  • autonomous context-update tool (update_context) no independent evidence
    purpose: Compress the interaction trajectory into a compact improvement state preserving verified findings, unresolved constraints, rejected candidates, and next-step plan.
    Only evaluated within this paper's benchmarks; no external falsifiable handle.
  • improvement state z_t^(k) no independent evidence
    purpose: Carry verified evidence and unresolved constraints across recursive research rounds.
    Internal representation; no external predictive handle beyond the paper's own experiments.

pith-pipeline@v1.3.0-alltime-deepseek · 17984 in / 12523 out tokens · 114654 ms · 2026-08-01T07:21:03.254274+00:00 · methodology

0 comments
read the original abstract

Deep research requires agents to find answers that jointly satisfy multiple constraints. Discovering such answers is costly, whereas verifying a candidate can often be decomposed into tractable constraint-wise checks. This discovery--verification asymmetry suggests that a research agent should do more than simply search longer: it should recursively improve its current answer by verifying intermediate results and using the partially verified state to guide subsequent refinement. We introduce AREX, a family of Recursively Self-Improving (RSI) deep research agents. AREX alternates between an inner research loop that gathers evidence and constructs a provisional answer, and an outer self-improvement loop that audits the answer constraint-wise, identifies unresolved claims, and launches targeted follow-up research. To sustain RSI over long horizons, AREX learns an autonomous context-update tool that compresses growing interaction history into a compact improvement state preserving verified evidence and unresolved constraints, without relying on an external model. We train AREX on verified synthetic tasks and high-quality trajectories through agentic mid-training and long-horizon reinforcement learning. To mitigate sparse final rewards during long horizon learning, we emphasize key steps where decisive evidence is acquired or erroneous research directions are corrected. We instantiate a dense 4B model and a 122B-A10B Mixture-of-Experts model. Across BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam (HLE), and other reasoning and tool-use benchmarks, AREX substantially outperforms comparable-scale baselines and remains competitive with models using substantially more activated parameters.

Figures

Figures reproduced from arXiv: 2607.21461 by Bingyu Yan, Chaofan Li, Chaozhuo Li, Di He, Hongjin Qian, Hongwang Xiao, Hui Wang, Jiahao Wang, Jianlyu Chen, Kang Liu, Kun Luo, Lei Xiong, Qiwei Ye, Sen Wang, Shuqi Lu, Wanli Li, Xiyan Jiang, Yingxia Shao, Yuyang Hu, Zhang Zhang, Zheng Liu, Zhicheng Dou, Zhongyuan Wang, Ziyi Xia.

Figure 1
Figure 1. Figure 1: Benchmark performance of AREX 1 arXiv:2607.21461v2 [cs.AI] 24 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AREX’s recursive self-improvement framework. The inner research loop maintains the research state and externalizes a provisional answer with supporting evidence and confidence, while the outer self-improvement loop accepts, refines, or restarts based on confidence and trajectory assessment. and efficient deep research agents. Our main contributions are summarized as follows: • We formulate multi-constraint… view at source ↗
Figure 3
Figure 3. Figure 3: Confidence distributions for correct and incorrect outputs. Each bar is [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average step loss after full-trajectory mid-training. [PITH_FULL_IMAGE:figures/full_fig_p014_4.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

19 extracted references · 1 canonical work pages

  1. [1]

    doi: 10.1038/s41586-025-09962-4

    CenterforAISafety,ScaleAI,andHLEContributorsConsortium.Abenchmarkofexpert-levelacademicquestionstoassessAIcapabili- ties.Nature,649(8099):1139–1146,2026. doi: 10.1038/s41586-025-09962-4. https://doi.org/10.1038/s41586-025-09962-4. Preprint title: Humanity’s Last Exam; arXiv:2501.14249. DeepSeek-AI. DeepSeek-V3 technical report.arXiv preprint arXiv:2412.19437,

    Pith/arXiv arXiv 2026

  2. [5]

    Albert Q

    doi: 10.48550/arXiv.2506.18096.https://arxiv.org/abs/2506.18096. Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierr...

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.18096

  3. [6]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan O

    doi: 10.48550/arXiv.2401.04088.https://arxiv.org/abs/2401.04088. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan O. Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-R1: Training LLMs to reason and leverage search engines with reinforcement learning. InThe Second Conference on Language Modeling, 2025.https://openreview.net/forum?id=Rwhi91ideu...

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.04088 2025

  4. [7]

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe

    doi: 10.48550/arXiv.2507.02592.https://arxiv.org/abs/2507.02592. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InInternational Conference on Learning Representations,

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.02592

  5. [8]

    Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al

    https://openreview.net/forum?id=v8L0pN6EOi. Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al. Deepseek-v3. 2: Pushing the frontier of open large language models.arXiv preprint arXiv:2512.02556,

    Pith/arXiv arXiv

  6. [9]

    doi: 10.18653/v1/2026.acl-long.966

    Association for Computational Linguistics. doi: 10.18653/v1/2026.acl-long.966. https://aclanthology.org/2026.acl-long.966/. Earlier preprint: arXiv:2510.06727. GrégoireMialon,ClémentineFourrier,CraigSwift,ThomasWolf,YannLeCun,andThomasScialom. GAIA:Abenchmarkforgeneral AIassistants. InInternationalConferenceonLearningRepresentations,2024. https://openrevi...

    arXiv 2026

  7. [10]

    2604.14518.https://arxiv.org/abs/2604.14518

    doi: 10.48550/arXiv. 2604.14518.https://arxiv.org/abs/2604.14518. 15 Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess,andJohnSchulman. WebGPT:Browser-ass...

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.14518

  8. [12]

    Official technical webpage; not a scholarly paper

    https://openai.com/index/ introducing-deep-research/. Official technical webpage; not a scholarly paper. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560,

    Pith/arXiv arXiv

  9. [14]

    Tongyi deepresearch technical report.arXiv preprint arXiv:2510.24701,

    Tongyi DeepResearch Team et al. Tongyi deepresearch technical report.arXiv preprint arXiv:2510.24701,

    Pith/arXiv arXiv

  10. [15]

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui

    doi: 10.48550/ arXiv.2510.24701.https://arxiv.org/abs/2510.24701. Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce LLMs step-by-step without human annotations. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pap...

    Pith/arXiv arXiv doi:10.48550/arxiv.2510.24701

  11. [16]

    doi: 10.18653/v1/2024.acl-long.510.https://aclanthology.org/2024.acl-long.510/

    Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.510.https://aclanthology.org/2024.acl-long.510/. Zihan Wang, Kangrui Wang, Qineng Wang, Pingyue Zhang, Linjie Li, Zhengyuan Yang, Kefan Yu, Minh Nhat Nguyen, Licheng Liu, Eli Gottlieb, Monica Lam, Yiping Lu, Kyunghyun Cho, Jiajun Wu, Li Fei-Fei, Lijuan Wang, Yejin Choi, and Manling ...

    Pith/arXiv arXiv 2024

  12. [17]

    Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese

    doi: 10.48550/arXiv.2504.20073.https://arxiv.org/abs/2504.20073. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. BrowseComp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516,

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.20073

  13. [18]

    Ryan Wong, Jiawei Wang, Junjie Zhao, Li Chen, Yan Gao, Long Zhang, Xuan Zhou, Zuo Wang, Kai Xiang, Ge Zhang, Wenhao Huang, Yang Wang, and Ke Wang

    doi: 10.48550/arXiv.2504.12516.https://arxiv.org/abs/2504.12516. Ryan Wong, Jiawei Wang, Junjie Zhao, Li Chen, Yan Gao, Long Zhang, Xuan Zhou, Zuo Wang, Kai Xiang, Ge Zhang, Wenhao Huang, Yang Wang, and Ke Wang. WideSearch: Benchmarking agentic broad info-seeking. InThe Fourteenth International Conference on Learning Representations, 2026.https://openrevi...

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.12516 2026

  14. [19]

    doi: 10.18653/v1/2026.findings-acl.956

    Association for Computational Linguistics. doi: 10.18653/v1/2026.findings-acl.956. https://aclanthology.org/2026.findings-acl.956/. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Bryan Kian Hsiang Low, and Paul Pu Liang. MEM1: Learning to synergize memory and reasoning for efficient long-horizon agents. InThe Fourteenth Internat...

    work page doi:10.18653/v1/2026.findings-acl.956 2026

  15. [2021]

    doi: 10.48550/arXiv.2112.09332.https://arxiv.org/abs/2112.09332. OpenAI. Introducing deep research. OpenAI technical post, February

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.09332

  16. [2023]

    https: //arxiv.org/abs/2310.08560

    doi: 10.48550/arXiv.2310.08560. https: //arxiv.org/abs/2310.08560. Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026.https://qwen.ai/blog?id=qwen3.5. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themse...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.08560 2026

  17. [2024]

    William Fedus, Barret Zoph, and Noam Shazeer

    doi: 10.48550/arXiv.2412.19437.https: //arxiv.org/abs/2412.19437. William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022.https://jmlr.org/papers/v23/21-0998.html. Jiaxuan Gao, Wei Fu, Minyang Xie, Shusheng Xu, Chuyi ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.19437 2022

  18. [2025]

    doi: 10.18653/v1/2025.acl-long.1575.https://aclanthology.org/2025.acl-long.1575/

    Association for Computational Linguistics. doi: 10.18653/v1/2025.acl-long.1575.https://aclanthology.org/2025.acl-long.1575/. Yuxuan Huang, Yihang Chen, Haozheng Zhang, Kang Li, Huichi Zhou, Meng Fang, Linyi Yang, Xiaoguang Li, Lifeng Shang, Songcen Xu, Jianye Hao, Kun Shao, and Jun Wang. Deep research agents: A systematic examination and roadmap.arXiv pre...

    Pith/arXiv arXiv 2025

  19. [2026]

    https://arxiv.org/abs/2601.20975

    doi: 10.48550/arXiv.2601.20975. https://arxiv.org/abs/2601.20975. Mengkang Hu, Tianxing Chen, Qiguang Chen, Yao Mu, Wenqi Shao, and Ping Luo. HiAgent: Hierarchical working memory management for solving long-horizon agent tasks with large language model. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Lo...

    doi:10.48550/arxiv.2601.20975