Pith. sign in

REVIEW 4 major objections 4 minor 77 references

Not Worth Another Token: Marginal Value Estimation for Efficient Deep Research Agents

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The central claim is that where you prune a deep research agent matters more than which scoring rule you use, with post-retrieval pruning delivering the largest savings and pre-synthesis pruning mainly refining the final report.

desk verdict Solid efficiency measurements, fragile quality claims—worth refereeing but the 'little quality degradation' conclusion needs multi-judge support. read the letter →

arxiv 2608.08389 v1 pith:P7QAZEZN submitted 2026-08-09 cs.AI cs.IRcs.MA

Harshitha Kolukuluru , Reshma Ashok , Kirat Arora , Evan William Ciccarelli , Nischal Ashok Kumar , Lunyiu Nie , Franck Dernoncourt , Samyadeep Basu
show 2 more authors
Ryan A. Rossi Nedim Lipka
This is my paper · ORCID
classification cs.AIcs.IRcs.MA
keywords deepresearchagentsmarginalvalueestimationstage-awarepruningcontexttokenefficiencyretrieval-augmentedgenerationlearnedMMR
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

Long-horizon research agents gather evidence iteratively, and context grows while each extra piece of evidence adds less value. This paper asks where in the pipeline low-value content should be pruned, comparing pre-retrieval, post-retrieval, and pre-synthesis interventions under one shared setup. Its central claim is that placement matters more than the scoring rule: pruning right after retrieval cuts tokens from 375.4k to 114.6k per report and explored nodes from 29.0 to 8.84 with only a small quality drop, and three-stage pruning reaches 73.3% token reduction. Late pruning barely saves cost and mainly refines the final synthesis context, and no single rule dominates across quality, efficiency, and faithfulness. The practical upshot is that system builders should pick both the pruning stage and the scoring rule according to the objective they care about.

What carries the argument

The central mechanism is a stage-aware pruning policy built on a unified marginal-value scorer $V(x|C_t,Q)$ applied with a stage-specific threshold at pre-retrieval, post-retrieval, and pre-synthesis. The scorer is instantiated as MMR (query relevance minus maximum similarity to already retained context), geometric residual novelty (distance from the candidate to the span of retained context), centroid drift (change in the semantic center), determinantal point process gain (volume added to a diverse subset), submodular coverage normalized by token cost, combined and lexical variants, an LLM judge, and a learned value model. Placement dominates because the token accounting shows result processing consumes most of the budget: pruning a branch before its retrieval and processing costs are paid saves all downstream tokens, whereas pre-synthesis pruning only shrinks the synthesis prompt.

What would settle it

Re-score all 100 generated reports with two or three independent LLM judges and compare method-level quality rankings; if the ranking of pre-synthesis Hybrid above post-retrieval MMR reverses across judges, the claims of little quality degradation and a best quality-efficiency trade-off lose support, while the token and node savings remain intact.

Watch

Extended reading notes

Core claim

The paper's discovery is that marginal-value pruning in deep research agents is primarily a placement problem. In a fixed deep-research pipeline evaluated on 100 queries, the authors compare heuristic, lexical, LLM-based, and learned scorers at three intervention points and find that post-retrieval pruning with MMR, a score that balances query relevance against redundancy, is the strongest single stage for cost control, reducing tokens from 375.4k to 114.6k and explored nodes from 29.0 to 8.84 while keeping about 97.9 percent of baseline quality. Pre-synthesis pruning can improve quality, with the Hybrid variant reaching 60.68 versus 57.83, but leaves the search tree essentially untouched. Two-stage pruning gives the best balanced trade-offs, with the CD + SC combination, centroid drift after retrieval plus submodular coverage before synthesis, improving quality by 1.64 points while cutting tokens by 63.4 percent. Three-stage MMR achieves the largest compression, 73.3 percent fewer tokens and about 73 percent fewer explored nodes, at a modest quality cost. No method dominates across quality, efficiency, relevance, and faithfulness, and the learned controller remains a proof of concept.

Load-bearing premise

The quality-trade-off conclusions assume that a fixed rubric-based LLM judge yields stable relative scores, even though the paper's own appendix shows the same reports scoring more than 30 points apart under different judges.

Editorial extensions

If this is right

  • A cheap relevance-redundancy rule applied right after retrieval can cut per-report token use by about 69.5% and explored branches by about 70% while staying within about 98% of baseline quality on the fixed rubric.
  • Pruning only before synthesis is a quality-refinement tool, not a cost-saving one: it can lift rubric quality by 2.85 points but leaves node count at the full 29.0 and token cost near baseline.
  • Two-stage pruning offers the most balanced operating points, such as centroid drift plus submodular coverage improving quality by 1.64 points while cutting tokens by 63.4%.
  • Three-stage pruning is for maximal compression: MMR reaches 73.3% token reduction and about 73% fewer explored nodes, at a cost of roughly 1.9 points of quality.
  • No single pruning rule wins on quality, efficiency, and source-grounding at once, so the choice of stage and scorer should follow the target objective.

Reading between the lines

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

  • If the cost structure is similar in other tree-structured research agents, with result processing dominating the token budget, the stage-ordering result should transfer, while the exact quality numbers will shift with the judge, generator, and task mix.
  • The learned controller's limited showing may reflect its small supervision set of 360 decisions labeled by one submodular family; training it on end-to-end report quality or on multiple pruning objectives could make learned pruning competitive with tuned heuristics.
  • Because quality and citation recall diverge under pruning, deployments should add an explicit metric for retention of minority or contradictory evidence, since aggressive early pruning may discard caveats that matter in high-stakes domains.
  • The 100-report sample and single pipeline make the efficiency ordering the most portable claim; the quality ordering should be re-estimated on any new benchmark before it is used to choose a method.
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 / 4 minor

Summary. The paper studies marginal-value-based pruning for long-horizon deep research agents, comparing heuristic, lexical, LLM-based, and learned pruning strategies at pre-retrieval, post-retrieval, and pre-synthesis stages in one-, two-, and three-stage configurations. Using a GPT-Researcher pipeline on 100 DeepResearchGym queries, it reports that stage placement matters more than the scoring rule: post-retrieval pruning yields the largest end-to-end token, node, and runtime savings, while pre-synthesis pruning mainly refines final report quality. The headline efficiency results include a 73.3% token reduction for three-stage MMR and a 69.5% reduction for post-retrieval MMR, with quality scores within a few points of baseline. The paper also presents cross-benchmark efficiency evidence on DeepResearch Bench and a small-scale learned pre-retrieval controller that does not outperform hand-designed heuristics.

Significance. If the efficiency findings hold, the paper offers useful practical guidance for context management in agentic retrieval systems, and the stage-aware comparison is a valuable framing. The paper has notable strengths: efficiency metrics (tokens, nodes, runtime) are reported with standard errors; the evaluation uses a shared underlying pipeline with cached evidence, so most observed efficiency differences are attributable to pruning; and the authors are unusually candid in Section 9 about judge dependence and the limited scope of quality conclusions. The cross-benchmark results in Appendix B.5 provide directional evidence that the stage-ordering efficiency result generalizes. However, the quality-sensitive claims—'little quality degradation' and specific 'best quality-efficiency trade-offs'—are not established at the current level of evidence, and several internal inconsistencies in stage attribution and configuration control affect the central comparison.

major comments (4)
  1. [§6.4, §A.2, Table 4, §9, Abstract] The abstract and Section 7 claims of 'little quality degradation' and specific 'best quality-efficiency trade-offs' are not supported by the reported quality evidence. The headline quality differences are small (e.g., post-retrieval MMR 56.62 vs. baseline 57.83; three-stage MMR 55.90; CD+SC 59.47), yet Table 4 shows that the same reports receive scores ranging from 34.37 to 91.50 depending on the judge, and Appendix A.2 states that quality, relevance, and faithfulness metrics are reported as means only with no standard errors or significance tests. Against a judge-induced swing of roughly 57 points, a 1.2-point gap or a 2.9-point gain is not interpretable. The paper's own Section 9 correctly cautions that quality-sensitive conclusions should be treated as relative comparisons under a fixed judge, but the abstract and conclusions do not carry that caveat. Either multi-judge stability analysis and significance testing are needed, or the quality-sensitive claims must be removed or substantially weakened.
  2. [§3.1, Tables 1, 5, 6, 7, 8] The stage comparison contains internal inconsistencies that undermine the controlled comparison. First, the row labeled 'LLM' under one-stage Post-Retrieval Pruning in Table 1 is actually a pre-retrieval configuration: Table 7 gives it a Pre-Retrieval latency of 105.64s and no Post-Retrieval latency, and Table 8 reports Query Pruning tokens. This violates the paper's stated exclusion of Pre-Retrieval-only configurations from one-stage analyses and mislabels where pruning occurs. Second, one-stage Centroid Drift explores 10.47 nodes (Table 6) while two-stage Centroid Drift explores 9.02 nodes, even though the two-stage configuration should prune at Post-Retrieval identically and additionally at Pre-Synthesis; pre-synthesis pruning cannot change the number of retrieval nodes. This implies that post-retrieval thresholds or other settings differ between the one- and two-stage runs, contradicting Section 6.3's claim that configurations 'differ only in how and where pruning is applied.' The stage-ordering conclusions require a consistent attribution of each configuration's actual pruning stages.
  3. [§3.1] The central claim that 'early pruning yields the largest end-to-end savings' is not fully tested because the earliest intervention point, Pre-Retrieval, is never evaluated as a standalone one-stage configuration. Section 3.1 excludes Pre-Retrieval-only and partial Pre-Retrieval combinations from the one- and two-stage analyses with the assertion that 'Pre-Retrieval decisions are necessarily more predictive and error-sensitive than later-stage decisions.' This is a hypothesis, not a measurement, and it directly affects the paper's headline ordering result. The current evidence supports 'Post-Retrieval pruning saves more than Pre-Synthesis pruning,' but the more general claim about early pruning requires either adding Pre-Retrieval-only results or narrowing the conclusion to the stages actually compared.
  4. [§B.1, Hybrid formula] The Hybrid scoring formula in Appendix B.1 is internally inconsistent and may describe a different method than the one evaluated. The formula is Score(x) = α·sim(e(x), q) − β·V_GRN(x|C) + γ·V_SC(x|C, Q), with the text stating that the second term 'rewards orthogonal novelty.' However, V_GRN is defined as the residual norm after projection, which is larger for more novel candidates, so the negative sign penalizes novelty rather than rewarding it. Since Pre-Synthesis Hybrid is the paper's strongest one-stage quality configuration, this sign error or typo is load-bearing. The authors should correct the formula or clarify the intended sign convention and confirm that the implemented method matches the corrected definition.
minor comments (4)
  1. [Table 5] Table 5 refers to the LLM-based pruning method as 'LLM (§B.1)' in several rows, but the LLM strategy is defined in Section B.2, not B.1; the cross-references are inconsistent with Table 1 and the appendix.
  2. [Table 6] Table 6 is titled 'Efficiency comparison across all methods' but omits the two-stage GRN and DPP rows that appear in Table 5; either add the missing rows or adjust the title and caption.
  3. [Table 7] The stage-level runtime values are not additive (e.g., baseline Research/Scraping plus Result Processing exceeds the total runtime), which is likely due to overlapping parallel execution; the caption should state this explicitly to avoid misreading.
  4. [§B.7] The threshold sweeps in Appendix B.7 are computed on a 10-query subset, while the main results use 100 queries; the main text should state this subset size and note that the 2% stability criterion is evaluated on the same noisy quality metric flagged in Section 9.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pruning heuristics are fixed formulas with hand-set thresholds, efficiency is measured against external benchmarks, and the learned controller is explicitly exploratory and not load-bearing.

full rationale

The paper's central claims are empirical comparisons rather than derivations, and no load-bearing step reduces to its own inputs. The pruning strategies (MMR, GRN, CD, DPP, SC, Hybrid) are classical fixed scoring formulas with hand-set hyperparameters, such as MMR lambda=0.35 and centroid-drift delta=0.03, and are not fitted to the reported quality or efficiency metrics. Appendix B.7 performs local threshold sweeps to confirm that the published operating points are not brittle, and efficiency numbers are objective measurements with standard errors computed on the external DeepResearchGym benchmark over a fixed GPT-Researcher baseline. The quality metric is a fixed rubric-based LLM judge, and the paper explicitly states in Section 9 and Appendix B.6 that absolute scores are judge-dependent and are interpreted only as relative comparisons under a fixed judge; the judge-sensitivity results in Table 4 are a measurement-validity concern, not a circularity. The learned pre-retrieval controller is trained on trajectories from the same benchmark, but the paper explicitly labels it a submodular-aligned exploratory proxy in Appendix B.3, and it is not used to derive the main stage-ordering or quality-efficiency conclusions. The only self-citation, Nie et al. 2026 for tree-structured deep research workflows, is contextual and not load-bearing: no uniqueness theorem or derivation is imported from it, and the pipeline itself is built on the external GPT-Researcher framework. No equation in the paper is equivalent by construction to the quantity it is claimed to predict.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The paper rests on domain assumptions about benchmark representativeness, embedding quality, and judge reliability, plus hand-set thresholds. The most consequential is the judge-stability assumption, since the quality trade-off conclusions depend on small score differences under a judge that the paper itself shows is highly variable. No new theoretical entities are introduced.

free parameters (10)
  • MMR lambda = 0.35
    Relevance-redundancy trade-off in post-retrieval and pre-synthesis MMR; hand-set; local sweep in Table 15 shows stability around this value.
  • DPP threshold tau = 0.30
    Determinant-ratio gain threshold for DPP pruning; hand-set; sweep shows stable region through 0.30, degraded at 0.40.
  • Centroid drift delta = 0.03
    Minimum centroid shift to retain a branch; hand-set; sweep shows 0.03 stable, 0.05 and 0.085 degrade quality.
  • GRN threshold tau = 0.85
    Residual-novelty threshold; hand-set; sweep shows non-monotonic quality with degradation at 0.80 and 0.90.
  • Submodular coverage threshold = 0.05
    Minimum marginal gain per token to retain; hand-set; stable through 0.08, degraded at 0.10.
  • LLM pruning delta = 0.3
    Threshold for LLM judge keep probability; hand-set; no sweep reported.
  • Lexical lambda and threshold = 0.6 and 0.2
    Weight on TF-IDF cosine vs bigram Jaccard, and lexical MMR threshold; hand-set.
  • Hybrid mixture weights = (alpha, beta, gamma) = (0.40, 0.30, 0.30)
    Weights for relevance, novelty, and coverage in the Hybrid score; chosen by the authors without sensitivity analysis.
  • Pre-synthesis context budget = max 10 contexts
    Greedy selection budget for MMR and DPP at pre-synthesis; hand-set.
  • Evaluation sample size = 100 queries
    Fixed subsample of the 1000-query DeepResearchGym benchmark; sampling method not specified.
assumptions (6)
  • domain assumption DeepResearchGym and its 100-query subsample are representative of deep research tasks; GPT-Researcher is a representative pipeline
    Central empirical conclusions are drawn from one benchmark and one framework (Section 6.1); the paper acknowledges generalization limits in Section 9.
  • domain assumption Fixed-judge rubric scores are comparable across methods under the same judge
    Quality comparisons rely on a single LLM judge; Appendix Table 4 shows judges vary by more than 30 points for identical reports, and the paper reports means without standard errors.
  • domain assumption Embedding similarity (BAAI/bge-small-en-v1.5) captures semantic relevance, novelty, and diversity for pruning decisions
    All heuristic scores operate in embedding space; if embeddings are poor, pruning signals degrade (Section 5.1).
  • ad hoc to paper The learned pre-retrieval controller's labels, derived only from submodular-family runs, are a valid utility signal
    Section B.3 states labels come only from GPTResearcher_sc runs, making the learned method a submodular-aligned proxy rather than a general learned policy.
  • ad hoc to paper Excluding Pre-Retrieval-only configurations from one- and two-stage analyses does not bias the stage comparison
    Section 3.1 excludes these configurations with the assumption that Pre-Retrieval decisions are 'necessarily more predictive and error-sensitive', which prevents isolating pre-retrieval's standalone effect.
  • standard math Standard math facts: cosine similarity, orthogonal projection, determinant ratios for PSD matrices, submodularity of facility-location objective
    Used in GRN, DPP, and SC scoring (Sections 5.1 and B.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Not Worth Another Token: Marginal Value Estimation for Efficient Deep Research Agents." pith.science (2026). https://pith.science/paper/P7QAZEZN

@misc{pith2026260808389,
  author       = {Pith},
  title        = {Pith review of: Not Worth Another Token: Marginal Value Estimation for Efficient Deep Research Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P7QAZEZN}},
  note         = {Machine review of arXiv:2608.08389}
}
read the original abstract

Long-horizon research agents solve open-ended tasks through iterative retrieval, aggregation, and synthesis, but context grows rapidly while the marginal value of additional evidence often declines. This leads to unnecessary token cost, higher latency, and noisier inputs for final report generation. We study marginal value estimation for context management in deep research agents and present the first systematic stage-aware comparison of pruning strategies across the pipeline. We evaluate lightweight heuristic criteria and a learned value model at pre-retrieval, post-retrieval, and pre-synthesis stages. Our results show that pruning effectiveness depends more on where pruning is applied than on the specific scoring rule: early pruning yields the largest end-to-end savings, while later pruning mainly refines the final synthesis context. Lightweight heuristics reduce token usage by up to 73% with little quality degradation, learned pruning remains competitive on selected trade-offs, and no single method dominates across quality, efficiency, and faithfulness. These findings provide practical guidance for designing efficient long-horizon agentic systems.

Figures

Figures reproduced from arXiv: 2608.08389 by the authors.

Figure 1
Figure 1. Overview of the deep research pipeline. Given a user query, the system proceeds through four stages: [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Prompts used in our pipeline [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Overall quality versus token usage in thousands (averaged over the 100 reports) for all pruning strategies [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 45 canonical work pages

  1. [1]

    Andrei Z. Broder. 1997. https://doi.org/10.1109/SEQUEN.1997.666900 On the resemblance and containment of documents . In Compression and Complexity of SEQUENCES 1997, Positano, Amalfitan Coast, Salerno, Italy, June 11-13, 1997, Proceedings , pages 21--29. IEEE

    arXiv 1997

  2. [2]

    Carbonell and Jade Goldstein

    Jaime G. Carbonell and Jade Goldstein. 2017. https://doi.org/10.1145/3130348.3130369 The use of mmr, diversity-based reranking for reordering documents and producing summaries . SIGIR Forum , 51(2):209--210

    arXiv 2017

  3. [5]

    Huiqiang Jiang, Qianhui Wu, Chin - Yew Lin, Yuqing Yang, and Lili Qiu. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.825 Llmlingua: Compressing prompts for accelerated inference of large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023 , pages 13358-...

    doi:10.18653/v1/2023.emnlp-main.825 2023

  4. [6]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. https://arxiv.org/abs/2310.06839 Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression . Preprint, arXiv:2310.06839

    arXiv 2024

  5. [7]

    Alex Kulesza and Ben Taskar. 2012. https://doi.org/10.1561/2200000044 Determinantal point processes for machine learning . Found. Trends Mach. Learn., 5(2-3):123--286

    doi:10.1561/2200000044 2012

  6. [11]

    Hui Lin and Jeff Bilmes. 2011. https://aclanthology.org/P11-1052/ A class of submodular functions for document summarization . In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 510--520, Portland, Oregon, USA. Association for Computational Linguistics

    work page 2011

  7. [12]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. https://doi.org/10.1162/TACL\_A\_00638 Lost in the middle: How language models use long contexts . Trans. Assoc. Comput. Linguistics, 12:157--173

    doi:10.1162/tacl 2024

  8. [14]

    Rossi, and Swarat Chaudhuri

    Lunyiu Nie, Nedim Lipka, Ryan A. Rossi, and Swarat Chaudhuri. 2026. https://arxiv.org/abs/2510.05145 Efficient tree-structured deep research with adaptive resource allocation . Preprint, arXiv:2510.05145

    arXiv 2026

Show all 77 references
  1. [15]

    Nikos Pagonas, Yeounoh Chung, Kostis Kaffes, and Arvind Krishnamurthy. 2025. https://arxiv.org/abs/2510.14126 Cortex: Workflow-aware resource pooling and scheduling for agentic serving . Preprint, arXiv:2510.14126

    2025

  2. [16]

    Radev, Hongyan Jing, Magorzata Sty, and Daniel Tam

    Dragomir R. Radev, Hongyan Jing, Magorzata Sty, and Daniel Tam. 2004. https://doi.org/10.1016/J.IPM.2003.10.006 Centroid-based summarization of multiple documents . Inf. Process. Manag., 40(6):919--938

    2004 doi

  3. [17]

    Juan Enrique Ramos. 2003. https://api.semanticscholar.org/CorpusID:14638345 Using tf-idf to determine word relevance in document queries

    2003

  4. [19]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. http://papers.nips.cc/paper\_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html Reflexion: language agents with verbal reinforcement learning . In Advances in ...

    2023

  5. [20]

    Tropp and Anna C

    Joel A. Tropp and Anna C. Gilbert. 2007. https://doi.org/10.1109/TIT.2007.909108 Signal recovery from random measurements via orthogonal matching pursuit . IEEE Trans. Inf. Theory , 53(12):4655--4666

    2007

  6. [22]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023 a . http://papers.nips.cc/paper\_files/paper/2023/hash/271db9922b8d1f4dd7aaef84ed5ac703-Abstract-Conference.html Tree of thoughts: Deliberate problem solving with large lang...

    2023

  7. [23]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023 b . https://openreview.net/forum?id=WE\_vluYUL-X React: Synergizing reasoning and acting in language models . In The Eleventh International Conference on Learning Representation...

    2023

  8. [24]

    ICLR 2026 Workshop on Lifelong Agents: Learning, Aligning, Evolving , year=

    Efficient Tree-Structured Deep Research with Adaptive Resource Allocation , author=. ICLR 2026 Workshop on Lifelong Agents: Learning, Aligning, Evolving , year=

    2026

  9. [25]

    Process Reward Models for

    Choudhury, Sanjiban , journal=. Process Reward Models for

  10. [26]

    The Use of

    Carbonell, Jaime and Goldstein, Jade , booktitle=. The Use of

  11. [27]

    Proceedings

    On the Resemblance and Containment of Documents , author=. Proceedings. Compression and Complexity of. 1997 , organization=

    1997

  12. [28]

    Ramos, Juan and others , booktitle=. Using. 2003 , organization=

    2003

  13. [29]

    Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , pages=

    A Class of Submodular Functions for Document Summarization , author=. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , pages=

  14. [30]

    Foundations and Trends

    Determinantal Point Processes for Machine Learning , author=. Foundations and Trends. 2012 , publisher=

    2012

  15. [31]

    Information Processing & Management , volume=

    Centroid-Based Summarization of Multiple Documents , author=. Information Processing & Management , volume=. 2004 , publisher=

    2004

  16. [32]

    IEEE Transactions on Information Theory , volume=

    Signal Recovery from Random Measurements via Orthogonal Matching Pursuit , author=. IEEE Transactions on Information Theory , volume=. 2007 , publisher=

    2007

  17. [33]

    2023 , eprint=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. 2023 , eprint=

    2023

  18. [34]

    Transactions of the Association for Computational Linguistics , year=

    Lost in the Middle: How Language Models Use Long Contexts , author=. Transactions of the Association for Computational Linguistics , year=

  19. [35]

    , journal=

    Choi, Nayoung and Zhang, Jonathan and Choi, Jinho D. , journal=

  20. [36]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and others , journal=

  21. [37]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

    2023

  22. [38]

    Maria Emilia Mazzolenis and Ruirui Zhang , year=. Agent

  23. [39]

    2510.00568 , archivePrefix=

    Shiyu Li and Yang Tang and Yifan Wang and Peiming Li and Xi Chen , year=. 2510.00568 , archivePrefix=

    arXiv

  24. [40]

    2026 , note=

    Nalar: An Agent Serving Framework , author=. 2026 , note=

    2026

  25. [41]

    2025 , note=

    Cortex: Workflow-Aware Resource Pooling and Scheduling for Agentic Serving , author=. 2025 , note=

    2025

  26. [42]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  27. [43]

    Publications Manual , year = "1983", publisher =

    1983

  28. [44]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    1981

  29. [45]

    2025 , primaryClass=

    DeepResearchGym: A Free, Transparent, and Reproducible Evaluation Sandbox for Deep Research , author=. 2025 , primaryClass=

    2025

  30. [46]

    Scalable Training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable Training of

  31. [47]

    Unlocking Context Constraints of

    Li, Yucheng , journal=. Unlocking Context Constraints of

  32. [48]

    2024 , archivePrefix=

    Researchy Questions: A Dataset of Multi-Perspective, Decompositional Questions for LLM Web Agents , author=. 2024 , archivePrefix=

    2024

  33. [49]

    Jiang, Huiqiang and Wu, Qianhui and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , journal=

  34. [50]

    Jiang, Huiqiang and Wu, Qianhui and Luo, Xufang and Li, Dongsheng and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili , journal=

  35. [51]

    Xu, Fangyuan and Shi, Weijia and Choi, Eunsol , journal=

  36. [52]

    Dan Gusfield , title =. 1997

    1997

  37. [53]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , note =

    2015

  38. [54]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  39. [55]

    2026 , eprint=

    Efficient Tree-Structured Deep Research with Adaptive Resource Allocation , author=. 2026 , eprint=

    2026

  40. [56]

    CoRR , volume =

    Sanjiban Choudhury , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.10325 , eprinttype =

    2025 doi arXiv

  41. [57]

    Carbonell and Jade Goldstein , title =

    Jaime G. Carbonell and Jade Goldstein , title =. 2017 , url =

    2017

  42. [58]

    Broder , editor =

    Andrei Z. Broder , editor =. On the resemblance and containment of documents , booktitle =. 1997 , url =

    1997

  43. [59]

    2003 , url=

    Using TF-IDF to Determine Word Relevance in Document Queries , author=. 2003 , url=

    2003

  44. [60]

    A Class of Submodular Functions for Document Summarization

    Lin, Hui and Bilmes, Jeff. A Class of Submodular Functions for Document Summarization. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies. 2011

    2011

  45. [61]

    Alex Kulesza and Ben Taskar , title =. Found. Trends Mach. Learn. , volume =. 2012 , url =

    2012

  46. [62]

    Radev and Hongyan Jing and Magorzata Sty and Daniel Tam , title =

    Dragomir R. Radev and Hongyan Jing and Magorzata Sty and Daniel Tam , title =. Inf. Process. Manag. , volume =. 2004 , url =

    2004

  47. [63]

    Tropp and Anna C

    Joel A. Tropp and Anna C. Gilbert , title =. 2007 , url =

    2007

  48. [64]

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , booktitle =

    Shunyu Yao and Dian Yu and Jeffrey Zhao and Izhak Shafran and Tom Griffiths and Yuan Cao and Karthik Narasimhan , editor =. Tree of Thoughts: Deliberate Problem Solving with Large Language Models , booktitle =. 2023 , url =

    2023

  49. [65]

    Liu and Kevin Lin and John Hewitt and Ashwin Paranjape and Michele Bevilacqua and Fabio Petroni and Percy Liang , title =

    Nelson F. Liu and Kevin Lin and John Hewitt and Ashwin Paranjape and Michele Bevilacqua and Fabio Petroni and Percy Liang , title =. Trans. Assoc. Comput. Linguistics , volume =. 2024 , url =

    2024

  50. [66]

    Choi , title =

    Nayoung Choi and Jonathan Zhang and Jinho D. Choi , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2601.07994 , eprinttype =

    2026 doi arXiv

  51. [67]

    Narasimhan and Yuan Cao , title =

    Shunyu Yao and Jeffrey Zhao and Dian Yu and Nan Du and Izhak Shafran and Karthik R. Narasimhan and Yuan Cao , title =. The Eleventh International Conference on Learning Representations,. 2023 , url =

    2023

  52. [68]

    Reflexion: language agents with verbal reinforcement learning , booktitle =

    Noah Shinn and Federico Cassano and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao , editor =. Reflexion: language agents with verbal reinforcement learning , booktitle =. 2023 , url =

    2023

  53. [69]

    CoRR , volume =

    Maria Emilia Mazzolenis and Ruirui Zhang , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2507.19543 , eprinttype =

    2025 doi arXiv

  54. [70]

    CoRR , volume =

    Shiyu Li and Yang Tang and Yifan Wang and Peiming Li and Xi Chen , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2510.00568 , eprinttype =

    2025 doi arXiv

  55. [71]

    CoRR , volume =

    Marco Laju and Donghyun Son and Saurabh Agarwal and Nitin Kedia and Myungjin Lee and Jayanth Srinivasa and Aditya Akella , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2601.05109 , eprinttype =

    2026 doi

  56. [72]

    2025 , eprint=

    Cortex: Workflow-Aware Resource Pooling and Scheduling for Agentic Serving , author=. 2025 , eprint=

    2025

  57. [73]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1973 , url =

    1973

  58. [74]

    CoRR , volume =

    Chenrui Tie and Shengxiang Sun and Jinxuan Zhu and Yiwei Liu and Jingxiang Guo and Yue Hu and Haonan Chen and Junting Chen and Ruihai Wu and Lin Shao , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.10090 , eprinttype =

    2025 doi

  59. [75]

    2026 , url =

    Taehwan Kim and Seunghwa Hyun and Jangseok Yu and Jongsun Park , title =. 2026 , url =

    2026

  60. [76]

    DeepResearchGym:

    Jo. DeepResearchGym:. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2505.19253 , eprinttype =

    2025 doi

  61. [77]

    Scalable training of L\(

    Galen Andrew and Jianfeng Gao , editor =. Scalable training of L\(. Machine Learning, Proceedings of the Twenty-Fourth International Conference. 2007 , url =

    2007

  62. [78]

    CoRR , volume =

    Yucheng Li , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2304.12102 , eprinttype =

    2023 doi arXiv

  63. [79]

    Researchy Questions:

    Corby Rosset and Ho. Researchy Questions:. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2402.17896 , eprinttype =

    2024 doi arXiv

  64. [80]

    LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models , booktitle =

    Huiqiang Jiang and Qianhui Wu and Chin. LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models , booktitle =. 2023 , url =

    2023

  65. [81]

    2024 , eprint=

    LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression , author=. 2024 , eprint=

    2024

  66. [82]

    CoRR , volume =

    Fangyuan Xu and Weijia Shi and Eunsol Choi , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2310.04408 , eprinttype =

    2023 doi arXiv

  67. [83]

    1997 , url =

    Dan Gusfield , title =. 1997 , url =. doi:10.1017/CBO9780511574931 , isbn =

    1997 doi

  68. [84]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. CoRR , volume =. 2015 , url =

    2015

  69. [85]

    Rie Kubota Ando and Tong Zhang , title =. J. Mach. Learn. Res. , volume =. 2005 , url =

    2005

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.