Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Deep Researcher with Test-Time Diffusion

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A deep research agent that treats report writing as a denoising process claims state-of-the-art results, beating OpenAI Deep Research in head-to-head comparisons.

desk verdict The engineering recipe is plausible and the short-form results are solid, but the long-form SOTA claim rests on a modestly calibrated LLM judge with no human check on the actual comparisons. read the letter →

arxiv 2507.16075 v1 pith:HCAXKRXZ submitted 2025-07-21 cs.CL

Rujun Han , Yanfei Chen , Zoey CuiZhu , Lesly Miculicich , Guan Sun , Yuanjun Bi , Weiming Wen , Hui Wan
show 10 more authors
Chunfeng Wen Solène Maître George Lee Vishy Tirumalashetty Emily Xue Zizhao Zhang Salem Haykal Burak Gokturk Tomas Pfister Chen-Yu Lee
This is my paper
classification cs.CL
keywords deepresearchagentstest-timescalingdiffusionprocessretrieval-augmentedgenerationself-evolutionlong-formreportLLM-as-a-judgemulti-hopreasoning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that deep research agents plateau when they rely on generic test-time scaling, and that the right structure is to treat report generation as a diffusion process, starting from a preliminary draft and iteratively refining it. Each round uses the current draft to drive new search queries, and retrieved information works as the 'denoising' signal that revises the draft. A component-wise self-evolution algorithm further improves the plan, questions, answers, and final report. If the claim holds, TTD-DR outperforms existing deep research agents on long-form report and multi-hop reasoning benchmarks, and the design principle could shift how agents use test-time compute.

What carries the argument

The central object is the denoising-with-retrieval loop shown in Algorithm 1, anchored by the evolving report draft. In each iteration the current draft is used to generate the next search question, the retrieved answer provides a concrete delta, and the draft is revised to remove noise, creating a continuous feedback loop where the report itself guides the search. The second mechanism, component-wise self-evolution, generates multiple variants of each workflow component, evaluates them with an LLM judge, revises them, and merges the survivors, thereby improving the context that feeds the denoising loop. Together these two mechanisms form the test-time diffusion process that the paper argues is responsible for the reported gains.

What would settle it

Have a panel of human raters, matching the ones whose agreement was measured at 69.0%, score the 205 LongForm Research report pairs behind the 69.1% win rate, and also score the DeepConsult pairs behind the 74.5% number; if human preference for TTD-DR over OpenAI Deep Research is not substantially above chance, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that TTD-DR, a test-time diffusion framework, achieves state-of-the-art results by modeling research report generation as an iterative denoising process guided by retrieval. A preliminary draft is treated as a noisy starting point and an evolving skeleton; each iteration feeds the current draft to the search-question generator, retrieves and synthesizes an answer, then revises the draft to remove imprecision and incompleteness. Component-wise self-evolution samples multiple variants of plans, search questions, answers, and report sections, scores them with an LLM-as-a-judge, revises them, and merges the best, so that each stage contributes high-quality context to the denoising loop. The paper reports win rates of 69.1% and 74.5% against OpenAI Deep Research on LongForm Research and DeepConsult, and higher correctness on HLE-search, HLE-full, and GAIA, supporting the claim that the framework outperforms existing deep research agents.

Load-bearing premise

The long-form win-rate claims rest on Gemini-1.5-pro as an LLM judge whose agreement with human raters (60.8%) is below the 69.0% human-human agreement, so the reported 69.1% and 74.5% win rates could be materially different if humans were the judges.

Editorial extensions

If this is right

  • If TTD-DR's results hold, deep research agents can outperform proprietary systems on long-form and multi-hop benchmarks using only a search tool, without requiring proprietary browsing or multimodal integrations.
  • The Pareto frontier analysis shows that adding denoising-with-retrieval revision steps yields more performance per unit latency than generic test-time scaling, so this design is a comparatively efficient way to scale test-time compute.
  • Early denoising steps already capture a large share of the final report's information, implying that a draft-centric loop reduces information loss and lets the agent converge faster than independent section-by-section search.
  • The ablation numbers attribute a substantial gain to self-evolution alone, lifting LongForm Research win rate from 39.4% to 60.9% before report-level denoising is added, which suggests component-level optimization is a major contributor by itself.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the auto-judge's alignment with human raters (60.8%) is below human-human agreement (69.0%), the long-form win rates are best read as an upper bound until a human evaluation confirms them.
  • Editorial inference: the draft acts as a shared mutable state that could serve as a dense reward signal for reinforcement-learning agent tuning, which the paper leaves to future work.
  • Editorial inference: the same denoising loop could generalize to multi-tool research, treating browsing, code execution, or multimodal retrieval as distinct denoising operators, which would be a natural test of whether the mechanism transfers beyond search-only agents.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Test-Time Diffusion Deep Researcher (TTD-DR), a deep-research agent built on Gemini-2.5-pro and Google's Agent Development Kit, which combines a three-stage backbone (plan generation, iterative search-and-synthesis, final report) with two test-time mechanisms: component-wise self-evolution and a report-level 'denoising with retrieval' loop that iteratively revises an evolving draft based on retrieved information. The agent is evaluated against OpenAI Deep Research, Perplexity, Grok, GPT-Researcher, and Open Deep Search on a private 205-query LongForm Research set, DeepConsult, HLE-search, HLE-full, and GAIA. The reported results are 69.1% and 74.5% win rates on the two long-form benchmarks, 33.9/34.3/69.1 correctness on the three short-form benchmarks, and an ablation study separating the backbone, self-evolution, and denoising-with-retrieval, together with Pareto-frontier analyses of latency versus quality.

Significance. If the long-form win-rate results held up under human evaluation, TTD-DR would be a strong and clearly ablated demonstration that a draft-centric iterative search-and-revision loop improves deep research agents over strong proprietary baselines. The short-form HLE and GAIA results are objective and consistently positive, and the paper deserves credit for publishing its LLM-judge calibration (Appendix A.3) and for making the latency/quality trade-off explicit. The main limitations are that the abstract's state-of-the-art claim for long-form generation rests on a single modestly aligned auto-rater, several benchmark margins lack confidence intervals, and the per-benchmark hyperparameters and private evaluation sets make the generality claim difficult to verify. As a conceptual contribution, the paper is primarily an engineering integration of iterative drafting, RAG, and multi-sample merging under a diffusion metaphor, rather than a new algorithmic framework.

major comments (4)
  1. [Appendix A.3, Table 3 and Section 4.1] The headline win rates on LongForm Research (69.1%) and DeepConsult (74.5%) are produced by Gemini-1.5-pro as LLM-as-a-judge, whose agreement with human raters is 60.8% versus 69.0% for human-human agreement, and no human evaluation is reported on the actual test pairs. Because the abstract's state-of-the-art claim is for long-form report generation, this is load-bearing: a judge with this alignment can carry systematic stylistic or length biases that aggregate accuracy does not reveal. Please add human side-by-side judgments on a representative sample of the LongForm Research and DeepConsult pairs, or explicitly restrict the claim to 'wins under the calibrated auto-rater.'
  2. [Section 4.1, Table 1] The short-form correctness margins on GAIA (69.1 vs 67.4, a 1.7-point gain) and HLE-search (33.9 vs 29.1) are reported without confidence intervals or significance tests; on a 200-query subset and on GAIA's validation set these margins are within plausible sampling noise. Please report bootstrap confidence intervals and, where appropriate, per-benchmark significance tests for all correctness comparisons, including the ablation rows in Table 2.
  3. [Appendix A.6, Table 4 versus Section 2.2 and Figure 5] In the reported best settings, the search-question and answer self-evolution steps are zero on every benchmark (s_q=0, s_a=0), and the report self-evolution steps are zero on HLE and GAIA (s_r=0). The '+Self-Evolution' ablation in Table 2 therefore conflates multiple-initial-state sampling with merging and the evolutionary revision loop described in Section 2.2; for HLE/GAIA it also does not evolve the final report. Please define precisely which variant of the self-evolution algorithm was run, and reinterpret the ablation gains accordingly.
  4. [Appendix A.4 and Table 4] HLE-search is a 200-query subset created by a Gemini-1.5-pro categorization, and the main experimental settings differ by benchmark (e.g., n_r=1 vs 5, s_r=1 vs 0) with no validation split described. The paper should state whether the per-benchmark hyperparameters in Table 4 were selected using the same test sets, and if so, discuss the multiple-comparison risk; otherwise the 'state-of-the-art on a wide array of benchmarks' claim is hard to assess.
minor comments (5)
  1. [Algorithm 1, Line 7] The exit_loop predicate is referenced but never defined; please specify the stopping condition and how it interacts with the maximum revision count N=20 stated in Section 3.4.
  2. [Section 3.3 and reproducibility] No code or release of the private LongForm Research queries is provided; please release code, configuration files, or at least anonymized example outputs and judge prompts to make the evaluation reproducible.
  3. [Figures 12 and 13] Figures 12 and 13 appear to have nearly identical captions for the same HLE-search Pareto frontier; please confirm they are distinct figures or merge them.
  4. [Section 3.5] The first sentence says 'We compare our RA systems'; this should be 'DR systems'.
  5. [Section 3.5] For proprietary baselines whose outputs were scraped, please state retrieval dates and product versions, since these systems evolve rapidly and the comparison is otherwise time-stamped only by the arXiv submission.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claimed results are external benchmark measurements, and the self-citations are contextual rather than load-bearing.

full rationale

The paper's central claim is an empirical comparison, not a derivation. TTD-DR is a pipeline that plans, searches, answers, and iteratively revises a draft; the final report is then scored on held-out queries from LongForm Research, DeepConsult, HLE, and GAIA. No fitted parameter is renamed as a prediction: the win rates and correctness scores are measured outcomes of the pipeline, and the self-evolution and denoising loops are algorithmic mechanisms rather than definitions of the evaluation target. The paper calibrates its long-form judge against 200 human-rated pairs (Appendix A.3, Table 3) and uses the calibrated judge only as the evaluator, not as a term in the method's construction. The citations to the authors' prior work (e.g., Han et al., 2024) are contextual references for evaluation challenges and domain coverage, and the self-evolution citations (Lee et al., 2025; Novikov et al., 2025) are inspiration sources, not load-bearing uniqueness claims. The legitimate concern is evaluator alignment (Gemini-1.5-pro achieves 60.8% agreement with humans vs. 69.0% human-human), but that is a validity and robustness issue, not circularity under the definitions used here. Short-form HLE/GAIA results use official ground-truth prompts and are externally anchored. Hence no step reduces by construction to its own input.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework is a system with hand-chosen hyperparameters rather than a fitted model. The most consequential choices are the per-benchmark initial-state counts and the use of a single model family for generation and judging. There are no newly invented physical or mathematical entities.

free parameters (7)
  • maximum denoising/search steps N = 20
    Set to 20 for all benchmarks (Section 3.4) as the maximum revision steps in Algorithm 1.
  • number of initial search query states n_q = 5
    Table 4 sets n_q=5 for all datasets; this controls the parallel sampling width for search questions.
  • number of initial answer states n_a = 3
    Table 4 sets n_a=3 for all datasets; this controls the parallel sampling width for answers.
  • number of initial report states n_r = 1 for LongForm/DeepConsult, 5 for HLE/GAIA
    Table 4 varies n_r by benchmark, indicating per-dataset tuning.
  • search question self-evolving steps s_q = 0 for all benchmarks
    Table 4 sets s_q=0, contradicting the prose that self-evolution applies to every stage's agents.
  • answer self-evolving steps s_a = 0 for all benchmarks
    Table 4 sets s_a=0, so the answer stage only does parallel sampling and merging, not iterative self-evolution.
  • HLE-search subset size = 200 sampled queries
    Section 3.3 randomly samples 200 of the queries labeled as search-requiring by Gemini-1.5-pro.
assumptions (5)
  • domain assumption Gemini-1.5-pro pairwise ratings are a reliable proxy for human judgments of report helpfulness and comprehensiveness.
    Invoked in Section 3.2 to produce all long-form win rates; Table 3 shows only 60.8% agreement with human raters.
  • domain assumption The 205 LongForm Research queries are representative real-world deep research requests.
    This private dataset is not released and no inter-annotator statistics are given for it.
  • domain assumption Gemini-1.5-pro's categorization of HLE queries into 'search' vs 'reasoning' produces a valid HLE-search subset.
    The subset is used as a benchmark in Tables 1-2, but the exact sampled queries are not released and the labeling model is from the same family as the system.
  • domain assumption Google-search grounding returns sufficient relevant documents for every search question.
    The entire retrieval component depends on this, but no retrieval quality analysis is provided.
  • ad hoc to paper The exit_loop condition in Algorithm 1 terminates at a sensible stopping point.
    Line 7 of Algorithm 1 calls exit_loop, but the criterion is never defined; the loop may simply run all 20 steps.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Researcher with Test-Time Diffusion." pith.science (2026). https://pith.science/paper/HCAXKRXZ

@misc{pith2026250716075,
  author       = {Pith},
  title        = {Pith review of: Deep Researcher with Test-Time Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCAXKRXZ}},
  note         = {Machine review of arXiv:2507.16075}
}
read the original abstract

Deep research agents, powered by Large Language Models (LLMs), are rapidly advancing; yet, their performance often plateaus when generating complex, long-form research reports using generic test-time scaling algorithms. Drawing inspiration from the iterative nature of human research, which involves cycles of searching, reasoning, and revision, we propose the Test-Time Diffusion Deep Researcher (TTD-DR). This novel framework conceptualizes research report generation as a diffusion process. TTD-DR initiates this process with a preliminary draft, an updatable skeleton that serves as an evolving foundation to guide the research direction. The draft is then iteratively refined through a "denoising" process, which is dynamically informed by a retrieval mechanism that incorporates external information at each step. The core process is further enhanced by a self-evolutionary algorithm applied to each component of the agentic workflow, ensuring the generation of high-quality context for the diffusion process. This draft-centric design makes the report writing process more timely and coherent while reducing information loss during the iterative search process. We demonstrate that our TTD-DR achieves state-of-the-art results on a wide array of benchmarks that require intensive search and multi-hop reasoning, significantly outperforming existing deep research agents.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Preconditioned Test-Time Adaptation for Out-of-Distribution Debiasing in Narrative Generation

    cs.CL 2026-03 unverdicted novelty 6.0 of 10

    CAP-TTA triggers context-aware preconditioned LoRA updates on high bias-risk OOD prompts to reduce toxicity in LLM narrative generation while preserving fluency and avoiding catastrophic forgetting.

  2. Open Data Synthesis For Deep Research

    cs.CL 2025-08 conditional novelty 6.0 of 10

    InfoSeek automatically synthesizes hierarchical constraint-satisfaction QA data from webpages; a 3B agent trained on it reaches 16.5% accuracy on BrowseComp-Plus, outperforming Qwen3-32B and Gemini 2.5 Flash.

  3. Deep Research Agents: A Systematic Examination And Roadmap

    cs.AI 2025-06 conditional novelty 4.0 of 10

    A survey that organizes LLM-powered deep research agents into static versus dynamic workflows and single versus multi agent architectures, and reviews their benchmarks and open challenges.

Reference graph

Works this paper leans on

55 extracted references · 21 canonical work pages · cited by 3 Pith papers

  1. [1]

    Alzubi, C

    S. Alzubi, C. Brooks, P. Chiniya, E. Contente, C. von Gerlach, L. Irwin, Y. Jiang, A. Kaz, W. Nguyen, S. Oh, H. Tyagi, and P. Viswanath. Open deep search: Democratizing search with open-source reasoning agents. 03 2025. URL https://arxiv.org/abs/2503.20201

    arXiv 2025

  2. [2]

    J. Baek, S. K. Jauhar, S. Cucerzan, and S. J. Hwang. Researchagent: Iterative research idea generation over scientific literature with large language models. April 2024

    work page 2024

  3. [3]

    Catalano

    A. Catalano. Patterns of graduate students' information seeking behavior: a meta‐synthesis of the literature", journal of documentation. Patterns of graduate students' information seeking behavior: a meta‐synthesis of the literature, 69 0 (2): 0 243--274, 2013. URL https://doi.org/10.1108/00220411311300066

    doi:10.1108/00220411311300066 2013

  4. [4]

    Q. Chen, M. Yang, L. Qin, J. Liu, Z. Yan, J. Guan, D. Peng, Y. Ji, H. Li, M. Hu, Y. Zhang, Y. Liang, Y. Zhou, J. Wang, Z. Chen, and W. Che. Ai4research: A survey of artificial intelligence for scientific research. 07 2025. URL https://arxiv.org/pdf/2507.01903

    arXiv 2025

  5. [5]

    M. S. Chitwood. Do you know the steps of the writing process?, 2022. URL https:https://melanieschitwood.com/do-you-know-the-steps-of-the-writing-process/

    work page 2022

  6. [6]

    Coelho, J

    J. Coelho, J. Ning, J. He, K. Mao, A. Paladugu, P. Setlur, J. Jin, J. Callan, J. Magalhães, B. Martins, and C. Xiong. Deepresearchgym: A free, transparent, and reproducible evaluation sandbox for deep research. 05 2025. doi:10.48550/arXiv.2505.19253

    doi:10.48550/arxiv.2505.19253 2025

  7. [7]

    Deerflow, 2025

    DeerFlow. Deerflow, 2025. URL https://github.com/bytedance/deer-flow

    work page 2025

  8. [8]

    Flower and J

    L. Flower and J. R. Hayes. A cognitive process theory of writing. College Composition and Communication, 32 0 (4): 0 365--387, 1981. ISSN 0010096X. URL http://www.jstor.org/stable/356600

    work page 1981

Show all 55 references
  1. [9]

    Gemini diffusion, 2025

    Gemini. Gemini diffusion, 2025. URL https://deepmind.google/models/gemini-diffusion/

    2025

  2. [10]

    Gottweis, W.-H

    J. Gottweis, W.-H. Weng, A. Daryin, T. Tu, A. Palepu, P. Sirkovic, A. Myaskovsky, F. Weissenberger, K. Rong, R. Tanno, K. Saab, D. Popovici, J. Blum, F. Zhang, K. Chou, A. Hassidim, B. Gokturk, A. Vahdat, P. Kohli, and V. Natarajan. Towards an ai co-scientist. 02 2025. doi:10....

    2025 doi arXiv

  3. [11]

    Grok, 2025

    Grok. Grok, 2025. URL https://grok.com/

    2025

  4. [12]

    J. Guan, W. Wu, Z. Wen, P. Xu, H. Wang, and M. Huang. Amor: A recipe for building adaptable modular knowledge agents through process feedback. 2024. URL https://arxiv.org/abs/2402.01469

    2024 arXiv

  5. [13]

    R. Han, Y. Zhang, P. Qi, Y. Xu, J. Wang, L. Liu, W. Y. Wang, B. Min, and V. Castelli. RAG - QA arena: Evaluating domain robustness for long-form retrieval augmented question answering. In Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, editors, Proceedings of the 2024 Conference on ...

    2024 doi

  6. [14]

    X. Hu, H. Fu, J. Wang, Y. Wang, Z. Li, R. Xu, Y. Lu, Y. Jin, L. Pan, and Z. Lan. Nova: An iterative planning and search approach to enhance novelty and diversity of llm generated ideas. 2024. URL https://arxiv.org/abs/2410.14255

    2024 arXiv

  7. [15]

    Ichihara, Y

    Y. Ichihara, Y. Jinnai, T. Morimura, K. Abe, K. Ariu, M. Sakamoto, and E. Uchibe. Evaluation of best-of-n sampling strategies for language model alignment. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=H4S4ETc8c9

    2025

  8. [16]

    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. arXiv preprint arXiv:2503.09516, 2025

    2025 arXiv

  9. [17]

    Kimi-researcher end-to-end rl training for emerging agentic capabilities, 2025

    Kimi-Researcher. Kimi-researcher end-to-end rl training for emerging agentic capabilities, 2025. URL https://moonshotai.github.io/Kimi-Researcher/

    2025

  10. [18]

    K.-H. Lee, I. Fischer, Y.-H. Wu, S. B. Dave Marwood, D. Schuurmans, and X. Chen. Evolving deeper llm thinking. 2025. URL https://arxiv.org/abs/2501.09891

    2025 arXiv

  11. [19]

    D. Li, B. Jiang, L. Huang, A. Beigi, C. Zhao, Z. Tan, A. Bhattacharjee, Y. Jiang, C. Chen, T. Wu, K. Shu, L. Cheng, and H. Liu. From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv: 2411.16594, 2024

    2024

  12. [20]

    X. Li, G. Dong, J. Jin, Y. Zhang, Y. Zhou, Y. Zhu, P. Zhang, and Z. Dou. Search-o1: Agentic search-enhanced large reasoning models. CoRR, abs/2501.05366, 2025 a . doi:10.48550/ARXIV.2501.05366. URL https://doi.org/10.48550/arXiv.2501.05366

    2025 doi arXiv

  13. [21]

    X. Li, J. Jin, G. Dong, H. Qian, Y. Zhu, Y. Wu, J.-R. Wen, and Z. Dou. Webthinker: Empowering large reasoning models with deep research capability. 2025 b . URL https://arxiv.org/abs/2504.21776

    2025 arXiv

  14. [22]

    Liang, Z

    T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, Z. Tu, and S. Shi. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023

    2023 arXiv

  15. [23]

    A. Lim, S. Jain, and V. Seng. Deepconsult: A deep research benchmark for consulting / business queries, 2025. URL https://github.com/Su-Sea/ydc-deep-research-evals

    2025

  16. [24]

    Y. Liu, H. Zhou, Z. Guo, E. Shareghi, I. Vulic, A. Korhonen, and N. Collier. Aligning with human judgement: The role of pairwise preference in large language model evaluators. arXiv preprint arXiv:2403.16950, 2024

    2024 arXiv

  17. [25]

    C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha. The AI S cientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292, 2024

    2024 arXiv

  18. [26]

    Madaan, N

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark. Self-refine: Iterative refinement with self-feedback. In Thirty-seventh Conference on ...

    2023

  19. [27]

    Mialon, C

    G. Mialon, C. Fourrier, C. Swift, T. Wolf, Y. LeCun, and T. Scialom. Gaia: a benchmark for general ai assistants. 11 2023. URL https://arxiv.org/abs/2311.12983

    2023 arXiv

  20. [28]

    S. Nie, F. Z. 1, Z. You, X. Zhang, J. Ou, J. Hu, J. Zhou, Y. Lin, J.-R. Wen, and C. Li. Large language diffusion models. 2025. URL https://arxiv.org/abs/2502.09992

    2025 arXiv

  21. [29]

    Novikov, N

    A. Novikov, N. Vu, M. Eisenberger, E. Dupont, P.-S. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, M. P. K. Abbas Mehrabian, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog. Alphaevolve: A coding agent for scientific and algor...

    2025

  22. [30]

    Introducing deep research, 2025

    OpenAI. Introducing deep research, 2025. URL https://openai.com/index/introducing-deep-research/

    2025

  23. [31]

    Introducing perplexity deep research, 2025

    Perplexity. Introducing perplexity deep research, 2025. URL https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research

    2025

  24. [32]

    L. Phan, A. Gatti, Z. Han, N. Li, J. Hu, H. Zhang, C. B. C. Zhang, M. Shaaban, J. Ling, S. Shi, M. Choi, A. Agrawal, A. Chopra, A. Khoja, R. Kim, R. Ren, J. Hausenloy, O. Zhang, M. Mazeika, T. Nguyen, D. Anderson, I. A. Shah, M. Doroshenko, A. C. Stokes, M. Mahmood, J. Lee, O....

    2025 arXiv

  25. [33]

    J. Qiu, X. Qi, T. Zhang, X. Juan, J. Guo, Y. Lu, Y. Wang, Z. Yao, Q. Ren, X. Jiang, X. Zhou, D. Liu, L. Yang, Y. Wu, K. Huang, S. Liu, H. Wang, and M. Wang. Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution. arXiv...

    2025 arXiv

  26. [34]

    O. D. Research. Open deep research, 2025. URL https://github.com/langchain-ai/open_deep_research

    2025

  27. [35]

    Researcher

    G. Researcher. Gpt researcher, 2025. URL https://github.com/assafelovic/gpt-researcher

    2025

  28. [36]

    Roucher, A

    A. Roucher, A. V. del Moral, merve, T. Wolf, and C. Fourrier. Open-source deepresearch – freeing our search agents, 2025. URL https://huggingface.co/blog/open-deep-research

    2025

  29. [37]

    Schmidgall, Y

    S. Schmidgall, Y. Su, Z. Wang, X. Sun, J. Wu, X. Yu, J. Liu, Z. Liu, and E. Barsoum. Agent laboratory: Using llm agents as research assistants. 2025. URL https://arxiv.org/abs/2501.04227

    2025 arXiv

  30. [38]

    H. Shen, J. Zhang, B. Xiong, R. Hu, S. Chen, Z. Wan, X. Wang, Y. Zhang, Z. Gong, G. Bao, et al. Efficient diffusion models: A survey. Transactions on Machine Learning Research (TMLR), 2025

    2025

  31. [39]

    W. Shi, H. Tan, C. Kuang, X. Li, X. Ren, C. Zhang, H. Chen, Y. Wang, L. Shang, F. Yu, and Y. Wang. Pangu deepdiver: Adaptive search intensity scaling via open-web reinforcement learning. arXiv preprint arXiv:2505.24332, 2025

    2025

  32. [40]

    C. Si, D. Yang, and T. Hashimoto. Can llms generate novel research ideas? a large-scale human study with 100+ nlp researchers. 2024. URL https://arxiv.org/abs/2409.04109

    2024 arXiv

  33. [41]

    Stelmakh, Y

    I. Stelmakh, Y. Luan, B. Dhingra, and M.-W. Chang. ASQA : Factoid questions meet long-form answers. In Y. Goldberg, Z. Kozareva, and Y. Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8273--8288, Abu Dhabi, United A...

    2022 doi

  34. [42]

    J. Tang, L. Xia, Z. Li, and C. Huang. Ai-researcher: Autonomous scientific innovation. 2025. URL https://arxiv.org/abs/2505.18705

    2025 arXiv

  35. [43]

    Trivedi, N

    H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal. Musique: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10: 0 539--554, 2022. doi:10.1162/tacl_a_00475. URL https://aclanthology.org/2022.tacl-1.31/

    2022 doi

  36. [44]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, b. ichter, F. Xia, E. Chi, Q. V. Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Proc...

    2022

  37. [45]

    Yamada, R

    Y. Yamada, R. T. Lange, C. Lu, S. Hu, C. Lu, J. Foerster, J. Clune, and D. Ha. The ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search. 2025. URL https://arxiv.org/abs/2504.08066

    2025 arXiv

  38. [46]

    L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, and M.-H. Yang. Diffusion models: A comprehensive survey of methods and applications. 2022. URL https://arxiv.org/abs/2209.00796

    2022

  39. [47]

    Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning. H otpot QA : A dataset for diverse, explainable multi-hop question answering. In E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii, editors, Proceedings of the 2018 Conference on Empirical Me...

    2018 doi

  40. [48]

    J. Yoon, H. Cho, Y. Bengio, and S. Ahn. Fast monte carlo tree diffusion: 100x speedup via parallel sparse planning. 06 2025. URL https://arxiv.org/abs/2506.09498

    2025

  41. [49]

    Zhang, X

    K. Zhang, X. Yang, W. Y. Wang, and L. Li. Redi: efficient learning-free diffusion inference via trajectory retrieval. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org, 2023

    2023

  42. [50]

    Zheng, S

    Y. Zheng, S. Sun, L. Qiu, D. Ru, C. Jiayang, X. Li, J. Lin, B. Wang, Y. Luo, R. Pan, Y. Xu, Q. Min, Z. Zhang, Y. Wang, W. Li, and P. Liu. O pen R esearcher: Unleashing AI for accelerated scientific research. In D. I. Hernandez Farias, T. Hope, and M. Li, editors, Proceedings o...

    2024 doi

  43. [51]

    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. 2025. URL https://arxiv.org/abs/2504.03160

    2025 arXiv

  44. [52]

    Świechowski, K

    M. Świechowski, K. Godlewski, B. Sawicki, and J. Mańdziuk. Monte carlo tree search: a review of recent modifications and applications. Artificial Intelligence Review, 56, 07 2022. doi:10.1007/s10462-022-10228-y

    2022 doi

  45. [53]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  46. [54]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  47. [55]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.