Pith.

Computation and Language · arXiv:2607.26604 v1 · submitted 29 Jul 2026 · 26 references

WikiLoop: Jointly Learning to Build and Navigate Agent-Native Wikis with Downstream Feedback

Haoliang Ming, Feifei Li, Wenhui Que

The paper says

“Knowledge-base construction and querying are typically optimized in isolation: retrieval-augmented agents operate over a fixed, externally maintained index, whereas construction receives no signal from downstream use.”

more of the abstract

We present WikiLoop, a feedback-coupled framework that jointly learns to build and navigate an agent-native Wiki, a persistent linked-page knowledge base designed for machine navigation. A role-conditioned shared policy supports two interfaces: a Navigator retrieves evidence from the Wiki to answer queries, and a Builder proposes structured edits evaluated through downstream navigation. The Navigator follows a sufficiency-before-efficiency objective that applies retrieval-cost penalties only after full evidence has been collected. The Builder learns from utility differences: a frozen Navigator scores each candidate edit by its change in downstream performance, while a guard penalty discourages regressions on unrelated queries. Training combines sequential role-specific optimization with a final joint stage over role-homogeneous batches. With Qwen3.5-9B as the common backbone, WikiLoop reaches 62.6 aggregate Answer Correctness on AuthTrace, 6.3 points above LLM-Wiki, base, with the largest gains on multi-document queries. Controlled comparisons support the intended effects of both objectives, and the learned edits remain useful to a held-out Navigator. Paired comparisons indicate that the final shared policy largely retains both role-specific capabilities, improves Navigator and end-to-end Answer Correctness by 0.4 points relative to the corresponding specialist references, and consolidates both interfaces into one model. Without dataset-specific training, WikiLoop also improves over the same-backbone LLM-Wiki, base on HotpotQA and MuSiQue.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 1 Aug 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Well-designed training recipe for coupling wiki construction to downstream navigation, with honest ablations; the joint-consolidation claim is undercut by an unmatched-training-update confound, but the main result stands. This line is wrong

4 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 26 references say.

The paper says, at §4.4, Table 3; §A.1, Table S1

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The claim that Joint RL 'improves Navigator and end-to-end AC by 0.4 points relative to the corresponding specialist references' is confounded by additional training updates. Per Table S1, Joint RL runs 90 updates at a 2:1 NAV:BUILD ratio from θB, adding 60 NAV-reward updates and 30 Builder-reward updates to the 120 NAV and 90 BUILD updates used for the specialists.

more of objection 1

The final shared policy therefore receives 180 NAV updates and 120 BUILD updates, versus 120 and 90 for the specialist references. No matched-update control is reported (e.g., 60 additional NAV-RL updates from θN, or 30 additional BUILD-RL updates from θB followed by continued Navigator training). The paired-bootstrap intervals in Table 3 condition on the already-trained policies and do not control for training effort. As written, the abstract's claim that the final shared policy 'improves' over the specialists cannot be attr

The paper says, at §3.3 and §5

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The Builder reward is computed from isolated before/after utility on separate Wiki copies, and no learned edit is ever committed to the training Wiki during BUILD-RL or Joint RL. The paper explicitly flags persistent write-back and long-horizon co-adaptation as open directions.

more of objection 2

This is a load-bearing limitation because the reward being optimized is marginal utility in isolation, whereas the deployed Wiki is built by sequentially applying many edits whose effects interact. The end-to-end gains on AuthTrace indirectly suggest the proxy is useful, but the paper does not test whether isolated utility correlates with the utility of the same edit when applied to the persistent, evolving Wiki. A small-scale comparison with persistent write-back, or at least a post-hoc correlation analysis between isolated and accumulated edit utility, would substantially strengthen the central claim that downst

The paper says, at §A.1

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

Most reported configurations are trained and evaluated once, as stated in the supplement: 'Unless stated otherwise, each reported configuration is trained and evaluated once.' The confidence intervals and p-values are computed over evaluation queries or instances conditional on the trained policies; they do not reflect variance due to training seeds, data shuffling, or rollout sampling.

more of objection 3

This is especially relevant for Table 3, where the +0.4-point improvement is small and the paired-bootstrap interval excludes zero only under the single-seed conditioning. At minimum, the authors should report multiple training seeds for the primary comparisons (Joint SFT, NAV-RL, BUILD-RL, Joint RL) or explicitly qualify the effect sizes as single-seed estimates. Without this, the precise magnitude of the headline improvements is uncertain.

The paper says, at §3.3 and §4.1

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The Builder reward is produced by the authors' own frozen Navigator on the authors' own AuthTrace benchmark, and the closest baseline (LLM-Wiki, base) is from the same group. The held-out Navigator (a different-seed run of the same NAV-RL pipeline) and the zero-shot transfer to HotpotQA/MuSiQue mitigate the most direct circularity.

more of objection 4

However, they do not fully establish that the learned edits are useful to an independent downstream consumer: both Navigators share the same training data, teacher, reward design, and initial Wiki. To support the generality of 'agent-native Wiki' construction, the paper should evaluate edit utility under an independently trained or off-the-shelf navigator, or at least report edit-level transfer of the learned Builder to a navigator from a different family. If such an experiment is not feasible, the limitations section should state this dependence more prominen

5 minor objections

The paper says, at §4.2, Table 1

Pith says: minor objection 1 of 5

The HippoRAG 2 row appears misformatted: '269.836.5' should likely be '69.8 36.5'. Please fix the table.

The paper says, at §4.3, Table 2

Pith says: minor objection 2 of 5

The metrics 'ER Full' and 'Cost@Full' are used in Panel A but only defined in the table caption; the main text should define them explicitly, since they are central to interpreting the Navigator ablation.

show 3 more

The paper says, at §4.4, Table 3

Pith says: minor objection 3 of 5

The column headers 'Specialist Ref.' and 'Post-BUILD NAV' are ambiguous. Clarify that for Navigator rows the specialist reference is NAV-RL, for Builder rows it is BUILD-RL, and for end-to-end rows it is the specialist pair.

The paper says, at §4.4 and Table 3

Pith says: minor objection 4 of 5

'Holmp' should be 'Holm-adjusted p' or similar formatting; this appears several times.

The paper says, at §A.2

Pith says: minor objection 5 of 5

The description of 'Builder candidates' in Table S2 (628 / 2,385 / 335) is clear, but the text below introducing L0–L2 and quality tiers would benefit from a one-sentence summary of how tier labels are used downstream, since readers may otherwise confuse filtering with training-target construction.

What would settle it

Run a matched-update control on Qwen3.5-9B/AuthTrace: (1) continue NAV-RL from θN for 60 additional updates under the same reward and hyperparameters, and evaluate Navigator AC on the fixed Wiki; (2) continue BUILD-RL from θB for 30 additional updates, compose this continued Builder with the continued Navigator, and evaluate end-to-end All AC. If the continued NAV-RL reaches ≥60.8 AC and the matched specialist pair reaches ≥62.6 All AC, then the +0.4 gains in Table 3 are explained by additional training rather than by joint consolidation. If the controls fall below these values, the joint-stage attribution is confirmed.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 1 Aug 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 56 s · 18,500 tokens read, 12,350 written
  3. Editor wrote the verdict above 28 s · 19,293 tokens read, 2,986 written
  4. Referee major revision: 4 major objections, 5 minor 52 s · 18,812 tokens read, 5,179 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

WikiLoop makes a credible case for the main thing it claims: you can train a Builder to edit an agent-native wiki using a frozen Navigator's utility deltas and get real gains. The architecture is simple, the ablations are well designed, and the headline numbers—62.6 All AC vs LLM-Wiki base's 56.3, with the biggest gains on multi-document queries—hold up under a second judge and a blinded human rating. The zero-shot transfer to HotpotQA and MuSiQue is a smart addition and helps counter the self-benchmark worry.

What's actually new: the sufficiency-before-efficiency gate on the Navigator reward, the guarded utility-difference reward for the Builder, and the sequential recipe (NAV-RL, BUILD-RL, Joint RL) that produces a single shared policy. The ablations support the intended effect of each mechanism: gating improves AC and full-evidence completion while reducing premature stopping; the guard term cuts guard regression by two-thirds with little loss on affected queries; the held-out Navigator shows the edits generalize beyond the training evaluator. That is honest, careful empirical work.

The soft spot is exactly where the stress-test note points. The paper's claim that the final joint policy 'improves Navigator and end-to-end AC by 0.4 points relative to the corresponding specialist references' is confounded by training effort. Joint RL starts from θB and runs 90 updates at a fixed 2 NAV : 1 BUILD ratio, so the final policy receives 180 Navigator-reward updates and 120 Builder-reward updates, versus 120 and 90 for the specialists. There is no continued-training control. The +0.4 could simply be more gradient steps on the same rewards. This is a genuine omission, and it undercuts the abstract's 'improves while consolidating' phrase. It does not, however, undercut the main result: the Builder objective, which does not depend on the joint stage, produces the large gains.

Other soft spots: each configuration is trained once, so there's no seed variance; code and data are not released; the Builder reward is computed by the authors' own Navigator on their own AuthTrace benchmark. These are addressable conditions, not logical flaws. The paper itself flags the biggest conceptual caveat—edits are scored in isolation and never committed to the persistent Wiki during training (§5)—so a persistent write-back could change what the Builder optimizes.

Who's this for: anyone working on retrieval-augmented agents, memory writing, or KB construction. It deserves a serious referee; the review should ask for a matched-update control, training-seed variance, and artifacts. But the central mechanism is supported, and the paper is worth engaging with.

The strongest objection

The strongest quantitative evidence for the joint stage is Table 3, which reports that the final shared policy improves Navigator AC and end-to-end All AC by +0.4 over the corresponding specialist references. The comparison is not matched for training effort. Per Table S1 and §A.1, NAV-RL uses 120 updates, BUILD-RL uses 90 updates, and Joint RL uses 90 updates at a fixed 2 NAV : 1 BUILD ratio, i.e., 60 additional NAV updates and 30 additional BUILD updates, initialized from θB. Thus the final policy receives 180 NAV-reward updates and 120 Builder-reward updates, versus 120 and 90 for the specialists. The +0.4 gains could simply reflect additional gradient steps on the same rewards rather than a benefit of joint role-conditioned consolidation. No continued-training control is reported (e.g., 60 more NAV-RL updates from θN, or 30 more BUILD-RL updates from θB composed with a continued Navigator). The abstract's claim that the final shared policy 'improves' over specialists while consolidating is therefore not established by the current experimental design.

The reader's identified concern about isolated before/after edit utility is real and explicitly acknowledged in §5, but the paper's end-to-end results partially mitigate it: even if isolated ΔU is an imperfect proxy, the final built Wiki plus final Navigator achieves the headline gains and zero-shot transfer. The more decisive threat is the Table 3 attribution. The joint stage is not a clean test of consolidation because it adds training budget. From Table S1: NAV-RL 120 updates, BUILD-RL 90, Joint RL 90 with a 2:1 NAV:BUILD ratio, so Joint is NAV-RL + BUILD-RL + 60 extra NAV + 30 extra BUILD updates. The +0.4 improvements over specialists are within the range that extra gradient steps could explain, especially at the lower Joint RL learning rate. A matched-update control is standard for such a claim and is missing. This does not refute the broader conclusion that feedback-coupled training improves over the supervised initialization and LLM-Wiki, base — the 62.6 vs 56.3 gap and the role-objective ablations remain — but it removes the specific 'joint improves over specialists' component of the strongest claim. The CONDITIONAL verdict stands, with the added condition that Table 3's attribution be tested with matched training effort. Since the label does not change, verdict_should_be is UNCHANGED. I mark agreement as partial because the reader's weakest assumption is related but not identical to the training-update confound.

The referee's report

If the claims hold, the paper makes a useful contribution by demonstrating that downstream navigation utility can serve as a training signal for structured knowledge-base construction, rather than optimizing construction in isolation. The work has notable strengths: controlled ablations separately test the Navigator and Builder objectives; paired bootstrap and permutation tests are reported; the held-out Navigator provides a reasonable check against overfitting to the reward evaluator; blinded human evaluation corroborates the automatic-judge rankings; and zero-shot transfer to two benchmarks increases confidence in generalization. The paper is also unusually explicit about its limitations, including the isolated-edit training protocol and transductive knowledge-base setting. These strengths make the empirical pattern internally coherent. However, two aspects need attention before the central claims are fully supported: the joint-consolidation result is confounded by unequal training updates, and the evaluation rests on single training runs for most configurations, so the reported effect sizes, especially the small +0.4-point gains, are not yet robust.

Pith's reading of the paper (a machine's summary, not the authors')

This paper tries to prove that constructing a knowledge base and navigating it should not be optimized separately: the same policy can learn both roles, with the builder rewarded by how much its edits improve downstream answers. A navigator retrieves evidence from a linked-page wiki, and a builder proposes structured edits; a frozen navigator scores each candidate edit by the before/after change on affected queries, with a guard penalty protecting unrelated queries. The shared policy first specializes each role, then consolidates them. The result is a 6.3-point aggregate accuracy gain over a static-wiki baseline on the AuthTrace benchmark, concentrated on multi-document queries, with transfer to HotpotQA and MuSiQue. A sympathetic reader would care because it offers a path where persistent knowledge artifacts are shaped by the agents that actually use them.

What carries the argument. The carrier of the argument is the pair of role-specific reward signals inside a single role-conditioned policy. The Navigator reward applies retrieval-cost penalties only after evidence recall is complete (sufficiency-before-efficiency), which prevents premature stopping. The Builder reward is the guarded before/after utility difference: each candidate patch is applied to an isolated copy of the Wiki, scored by the frozen Navigator on affected and guard query sets, and the guard term penalizes regressions on unrelated queries while edit-cost and structural penalties shape the patch. A final joint reinforcement-learning stage interleaves role-homogeneous batches to consolidate both interface

Core claim. The paper claims that a knowledge base's construction can be guided by the same feedback that measures its downstream usefulness: a Builder proposes structured edits, a frozen Navigator scores the edit's marginal effect on query answering, and both roles share one policy. On the AuthTrace benchmark this feedback-coupled training raises aggregate answer correctness to 62.6, 6.3 points above a static-wiki baseline, with the largest gains (over 12 points) on questions requiring two or more documents. The learned edits also transfer to a held-out Navigator and, without dataset-specific training, to HotpotQA and MuSiQue. The authors frame this as evidence that downstream navigation utility is an

The premise it rests on. The builder's credit is computed from edits applied to an isolated copy and scored by a frozen navigator; if the utility of an edit in isolation differs from its utility after it becomes part of the persistent wiki that later navigators see, the builder optimizes the wrong objective.

If the paper is right (Pith's inferences, not the authors')

If the claim holds, structured knowledge-base construction can be trained end-to-end from downstream query utility rather than from reconstruction or hand-designed objectives.

The largest gains on multi-document queries imply that feedback-coupled construction is most valuable when navigation must compose evidence across pages.

A single shared policy can retain both builder and navigator capabilities, avoiding the cost and complexity of two separately deployed models.

Edits trained with one navigator remain useful to a held-out navigator, suggesting the learned construction signal generalizes beyond the evaluator.

Zero-shot gains on HotpotQA and MuSiQue indicate the approach transfers to new corpora without dataset-specific training.

If isolated, discard-after-scoring evaluation underestimates coordinated edits, a persistent write-back variant might either amplify the gains (edits compounding) or expose regression chains that the guard term does not catch.

The same before/after downstream-utility loop could apply to other persistent agent artifacts — code repositories, memory stores, or tool configurations — wherever construction and consumption are currently decoupled.

One could test the sufficiency-before-efficiency principle as a generic reward gate for any multi-step retrieval policy, independent of wikis.

The joint consolidation stage may be acting as a regularizer against role interference; ablating only the joint stage would isolate how much of the capability retention comes from the shared-policy update.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Knowledge-base construction and querying are typically optimized in isolation: retrieval-augmented agents operate over a fixed, externally maintained index, whereas construction receives no signal from downstream use. We present WikiLoop, a feedback-coupled framework that jointly learns to build and navigate an agent-native Wiki, a persistent linked-page knowledge base designed for machine navigation. A role-conditioned shared policy supports two interfaces: a Navigator retrieves evidence from the Wiki to answer queries, and a Builder proposes structured edits evaluated through downstream navigation. The Navigator follows a sufficiency-before-efficiency objective that applies retrieval-cost penalties only after full evidence has been collected. The Builder learns from utility differences: a frozen Navigator scores each candidate edit by its change in downstream performance, while a guard penalty discourages regressions on unrelated queries. Training combines sequential role-specific optimization with a final joint stage over role-homogeneous batches. With Qwen3.5-9B as the common backbone, WikiLoop reaches 62.6 aggregate Answer Correctness on AuthTrace, 6.3 points above LLM-Wiki, base, with the largest gains on multi-document queries. Controlled comparisons support the intended effects of both objectives, and the learned edits remain useful to a held-out Navigator. Paired comparisons indicate that the final shared policy largely retains both role-specific capabilities, improves Navigator and end-to-end Answer Correctness by 0.4 points relative to the corresponding specialist references, and consolidates both interfaces into one model. Without dataset-specific training, WikiLoop also improves over the same-backbone LLM-Wiki, base on HotpotQA and MuSiQue.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 0.73

“Paired comparisons indicate that the final shared policy largely retains both role-specific capabilities, improves Navigator and end-to-end Answer Correctness by 0.4 points relative to the corresponding specialist references, and consolidates both interfaces into one model.”

How the main claim is stated: hedged possibility 0.77

What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.22 · human participants 0.04 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Knowledge-base construction and querying are typically optimized in isolation: retrieval-augmented agents operate over a fixed, externally maintained index, whereas construction receives no signal from downstream use. background 0.66 0.06 0.03 asserted 1.00
We present WikiLoop, a feedback-coupled framework that jointly learns to build and navigate an agent-native Wiki, a persistent linked-page knowledge base designed for machine navigation. contribution 1.00 0.90 0.15 asserted 1.00
A role-conditioned shared policy supports two interfaces: a Navigator retrieves evidence from the Wiki to answer queries, and a Builder proposes structured edits evaluated through downstream navigation. method 0.94 0.82 0.06 asserted 1.00
The Navigator follows a sufficiency-before-efficiency objective that applies retrieval-cost penalties only after full evidence has been collected. method 0.99 0.83 0.03 asserted 1.00
The Builder learns from utility differences: a frozen Navigator scores each candidate edit by its change in downstream performance, while a guard penalty discourages regressions on unrelated queries. method 1.00 0.94 0.04 asserted 1.00
Training combines sequential role-specific optimization with a final joint stage over role-homogeneous batches. method 1.00 0.86 0.03 asserted 0.99
With Qwen3.5-9B as the common backbone, WikiLoop reaches 62.6 aggregate Answer Correctness on AuthTrace, 6.3 points above LLM-Wiki, base, with the largest gains on multi-document queries. result 1.00 0.97 0.32 asserted 1.00
Controlled comparisons support the intended effects of both objectives, and the learned edits remain useful to a held-out Navigator. result 0.83 0.95 0.15 asserted 0.99
Paired comparisons indicate that the final shared policy largely retains both role-specific capabilities, improves Navigator and end-to-end Answer Correctness by 0.4 points relative to the corresponding specialist references, and consolidates both interfaces into one model. result 0.98 0.96 0.30 hedged possibility 0.77
Without dataset-specific training, WikiLoop also improves over the same-backbone LLM-Wiki, base on HotpotQA and MuSiQue. result 1.00 0.94 0.12 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 7,703 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (1)
Figure 1 of the paper
Figure 1. Overview of WikiLoop. The Navigator retrieves evidence from a structured Wiki under a sufficiency-before-efficiency

Figures are the authors', reproduced from the paper.

References (26)
  1. [1] Ming, Haoliang and Li, Feifei and Wu, Xiaoqing and Que, Wenhui , journal =. Retrieval as Reasoning: Self-Evolving Agent-Native Retrieval via. 2026 , doi = on Pith
  2. [2] Wu, Xiaoqing and Li, Feifei and Ming, Haoliang and Que, Wenhui , journal =. 2026 , doi = on Pith
  3. [3] Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K\". Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume =. 2020 , url = on Pith
  4. [4] Dense Passage Retrieval for Open-Domain Question Answering , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , pages =. 2020 , publisher =. doi:10.18653/v1/2020.emnlp-main.550 , url = on Pithdoi:10.18653/v1/2020.emnlp-main.550
  5. [5] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , url = on Pith
  6. [6] Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2023 , publisher =. doi:10.18653/v1/2023.a... on Pithdoi:10.18653/v1/2023.acl-long.557
  7. [7] Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. 2024 , url = on Pith
  8. [8] Jeong, Soyeong and Baek, Jinheon and Cho, Sukmin and Hwang, Sung Ju and Park, Jong , booktitle =. Adaptive-. 2024 , publisher =. doi:10.18653/v1/2024.naacl-long.389 , url = on Pithdoi:10.18653/v1/2024.naacl-long.389
  9. [9] Fang, Jinyuan and Meng, Zaiqiao and Macdonald, Craig , booktitle =. 2026 , publisher =. doi:10.18653/v1/2026.acl-long.1793 , url = on Pithdoi:10.18653/v1/2026.acl-long.1793
  10. [10] Liu, Wenhan and Ma, Xinyu and Zhu, Yutao and Li, Yuchen and Shi, Daiting and Yin, Dawei and Dou, Zhicheng , booktitle =. Agentic-. 2026 , publisher =. doi:10.18653/v1/2026.findings-acl.785 , url = on Pithdoi:10.18653/v1/2026.findings-acl.785
  11. [11] Sarthi, Parth and Abdullah, Salman and Tuli, Aditi and Khanna, Shubh and Goldie, Anna and Manning, Christopher D. , booktitle =. 2024 , url = on Pith
  12. [12] Edge, Darren and Trinh, Ha and Cheng, Newman and Bradley, Joshua and Chao, Alex and Mody, Apurva and Truitt, Steven and Metropolitansky, Dasha and Ness, Robert Osazuwa and Larson, Jonathan , journal =. From Local to Global: A Graph. 2024 , doi = on Pith
  13. [13] Guti\'. Advances in Neural Information Processing Systems , volume =. 2024 , doi = on Pith
  14. [14] Guti\'. From. Proceedings of the 42nd International Conference on Machine Learning , series =. 2025 , publisher = on Pith
  15. [15] Guo, Zirui and Xia, Lianghao and Yu, Yanhua and Ao, Tu and Huang, Chao , booktitle =. 2025 , publisher =. doi:10.18653/v1/2025.findings-emnlp.568 , url = on Pithdoi:10.18653/v1/2025.findings-emnlp.568
  16. [16] Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , journal =. 2023 , doi = on Pith
  17. [17] Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , booktitle =. 2025 , publisher = on Pith
  18. [18] Yan, Sikuan and Yang, Xiufeng and Huang, Zuchao and Nie, Ercong and Ding, Zifeng and Li, Zonggen and Ma, Xiaowen and Bi, Jinhe and Kersting, Kristian and Pan, Jeff Z. and Sch\". arXiv preprint arXiv:2508.19828 , year =. doi:10.48550/arXiv.2508.19828 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2508.19828
  19. [19] Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents , author =. arXiv preprint arXiv:2601.01885 , year =. doi:10.48550/arXiv.2601.01885 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2601.01885
  20. [20] Yan, Jiangze and Shen, Yi and Zhang, Wenjing and Huang, Jieyun and Liu, Zhaoxiang and Wang, Ning and Wang, Kai and Lian, Shiguo , journal =. 2026 , doi = on Pith
show 6 more
  1. [21] Zhu, Yutao and Jin, Jiajie and Qian, Hongjin and Liu, Zheng and Dou, Zhicheng and Wen, Ji-Rong , booktitle =. Single. 2025 , publisher =. doi:10.18653/v1/2025.emnlp-main.243 , url = on Pithdoi:10.18653/v1/2025.emnlp-main.243
  2. [22] Multi-Agent Tool-Integrated Policy Optimization , author =. arXiv preprint arXiv:2510.04678 , year =. doi:10.48550/arXiv.2510.04678 , url = on Pithdoi:10.48550/arxiv.2510.04678
  3. [23] Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal =. 2024 , doi = on Pith
  4. [24] Group Sequence Policy Optimization , author =. arXiv preprint arXiv:2507.18071 , year =. doi:10.48550/arXiv.2507.18071 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2507.18071
  5. [25] Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D. , booktitle =. 2018 , publisher =. doi:10.18653/v1/D18-1259 , url = on Pithdoi:10.18653/v1/d18-1259
  6. [26] Trivedi, Harsh and Balasubramanian, Niranjan and Khot, Tushar and Sabharwal, Ashish , journal =. 2022 , doi = on Pith
Cited by 1 Pith paper
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 8 sentences without a confident label

Uses the method or tool of 3 papers
Uses data or values from 1 paper
Compares with 1 paper
Cites as background 4 papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment