Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification

T0 review · 3 major / 6 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Asking a large language model to emit an explicit memory-state label does not reliably improve its memory-policy decisions, because the benchmark-defined state is hard to predict and maps deterministically to the answer.

desk verdict The core negative finding is likely right, but the abstract's headline numbers conflate taxonomy exposure with the state-output field, so the primary contrast needs re-framing before acceptance. read the letter →

arxiv 2608.17247 v1 pith:JJ4SQYAJ submitted 2026-08-18 cs.AI

Yihang Chen , Pin Qian , Su Wang , Chong Peng , Huan Xu , Shuaiting Li , Yiqi Sun This is my paper · ORCID
classification cs.AI
keywords memory-policyclassificationexplicitstateelicitationpromptablationcounterfactualevaluationLLMagentsshortcutlearninglabelconditioningcontrolledaudit
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 sets out to establish that making a model output an explicit memory-state label (Active, Stale-or-Irrelevant, Conflicting-or-Superseded, Underspecified) is not a reliable way to improve memory-policy routing — the decision of whether retrieved user memory should be Used, Ignored, Updated, or Asked about. Its initial 480-example benchmark suggested large gains from a state-structured prompt, but the paper argues those gains were artifacts of a lexically separable dataset and a bundle of simultaneous prompt changes. On a controlled counterfactual set with matched four-way families, an isolated state-output field yields +0.6 percentage points for Llama-3.3-70B (p=0.81) and +3.3 points for GPT-OSS-120B (Holm-adjusted p=0.14), while merely exposing the state taxonomy gives significant gains of +9.17 and +5.00 points. The paper's broader aim is to promote an audit protocol that separates dataset shortcuts, bundled prompt components, answer-associated labels, and provider failures before crediting any structured intermediate output with an improvement.

What carries the argument

The mechanism that carries the argument is the frozen controlled counterfactual set: 160 synthetic examples organized into 40 scenario families, each with four matched variants whose task text is byte-identical and whose reference policies are derived deterministically from structured semantic slots before any natural-language rendering. The matched prompt ablation varies one component at a time (policy-only, taxonomy-only, state-output, state-output-with-rationale, deterministic state routing) so that accuracy differences can be pinned to a single change; the label-conditioning diagnostic then supplies correct or deliberately wrong state labels to test whether routing is sensitive to answer-associated information. These instruments, plus family-cluster statistics, seed-stability checks, and a rule-based error taxonomy, are what let the paper attribute the development-set gain to dataset shortcuts and bundled changes rather than to the state field itself.

What would settle it

An independent re-annotation of a naturalistic (non-synthetic) memory-policy corpus with the same matched four-way family structure that shows an isolated state-output field producing a large, clustered, statistically significant accuracy gain would overturn the central claim. Short of that, a replication where forced-correct-state routing approaches perfect accuracy while naturally emitted state accuracy stays low would localize the bottleneck to state inference rather than the state-to-policy mapping.

Watch

Extended reading notes

Core claim

The paper's central claim is a dissociation: telling a model what the four memory states mean helps its policy decisions, but requiring it to emit a state label in its JSON output does not. On the frozen 160-example controlled set, the matched state-output ablation is statistically null for Llama and non-significant for GPT-OSS, and family-level analysis shows that models almost never solve all four counterfactual variants of a scenario, so example-level accuracy overstates consistency. Supplying benchmark-implied state labels shifts routing, but because the four states deterministically map to the four policies, the paper interprets this as label-conditioning sensitivity rather than evidence of a faithful internal state mechanism. The conclusion is that explicit state elicitation requires more than adding a field to the prompt schema, and that apparent prompt gains must survive a controlled audit before being attributed to the intermediate representation.

Load-bearing premise

The load-bearing premise is that the rule-derived reference policies on the frozen 160-example set are the correct ground truth for every accuracy number; the paper concedes these labels 'may not capture every reasonable human judgment' and that annotator agreement 'partly reflects the construction,' so if the rules mis-encode the intended abstraction for any family, the null state-output result, the taxonomy gain, and the family-level statistics are all measured against a disputable standard.

Editorial extensions

If this is right

  • Memory-agent builders should not expect an explicit state-output field to improve policy routing; the isolated-field gain is null for Llama and non-significant for GPT-OSS.
  • Taxonomy exposure alone gives modest but significant accuracy gains, yet because the state wording mirrors the policy wording, the gain is at least partly a restatement effect rather than typed knowledge.
  • Supplied state labels can shift routing by up to about 12 points, and a conflicting supplied label costs roughly 15 points relative to a correct one, so deployed systems need state-quality checks, abstention, or clarification rather than state fields alone.
  • The state field redistributes errors between policies — Llama shifts toward Use (+17.5 pp) and away from Ask (-14.2 pp) — so an overall null result can hide opposing class-level effects.
  • Family-level counterfactual consistency is the right audit unit: complete four-way family success is rare (0% for Llama; at most 5.0% for GPT-OSS), so marginal accuracy improvements are not the same as tracking the intended semantic relation.

Reading between the lines

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

  • If the result generalizes, the bottleneck for memory agents is the decision mapping from history and task to policy, not information availability; the paper's own normalized-input probe, which made relations explicit and hurt accuracy, points the same direction.
  • The five-stage audit protocol transfers to any structured intermediate output — rationales, plans, tags — claimed to improve agent behavior; the same shortcut, isolation, and answer-association checks would apply, although the paper instantiates the protocol only on memory-policy classification.
  • A clean test of the open question would use evidence fields with no deterministic mapping to the final answer; the paper's semantic-evidence follow-up is negative for Llama but unusable for GPT-OSS because of provider-side JSON validation failures, so that condition remains untested.
  • The forced-wrong-state result suggests a concrete safety design: if a model routes on supplied state, an erroneous upstream state extractor could silently shift policy decisions by roughly 8 points relative to no state at all, arguing for confidence thresholds or abstention on state outputs.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper studies memory-policy classification for personalized agents, where a retrieved user memory must be routed to Use, Ignore, Update, or Ask. It first shows that a 480-example synthetic development set is lexically separable and lacks positive Ignore cases, so its apparent gains from a state-structured prompt bundle are treated as motivation rather than evidence. The authors then construct a frozen 160-example controlled counterfactual set with 40 matched four-way families and rule-derived reference policies, and run a five-stage audit protocol: dataset shortcuts, prompt isolation, answer-associated label conditioning, decomposed semantic evidence, and provider-level execution failures. On the frozen set, exposing the four state definitions improves accuracy over policy-only prompting, but the paper claims that an isolated explicit state-output field does not significantly improve policy accuracy (Llama +0.6 pp, p=0.8084; GPT-OSS +3.3 pp, raw p=0.0710). A separate label-conditioning diagnostic shows that supplying benchmark-associated state labels shifts routing, and family-level analyses show that complete four-way family success is rare. The paper concludes that explicit state elicitation is not a reliable mechanism for improving memory-policy routing and that development-set bundle gains do not survive isolation.

Significance. If the central result holds, the paper makes a useful negative contribution: it provides a reusable audit protocol for structured intermediate outputs and demonstrates, with unusually transparent execution accounting, that a state-output field is not a reliable route to better policy classification. The design has real strengths: the controlled-set arms were frozen before hosted evaluation; parse failures count as incorrect; uncertainty is quantified with family-cluster bootstrap and permutation tests with Holm adjustment; label-conditioning conditions are explicitly separated from mechanism evidence; and the appendices contain detailed cost, failure, and per-policy reporting. These practices are exactly what the field needs for prompt-intervention claims. However, the headline quantitative claim about an "isolated" state-output field is not the contrast that was actually tested, as detailed below. The broader qualitative conclusion is defensible, but the reported numbers and the claimed dissociation need substantive reanalysis before the paper can be accepted.

major comments (3)
  1. [Abstract; §7; Table 1; Appendix F (Table 4)] The abstract and Section 7 describe the state-output result as an "isolated explicit state-output field" effect, but the primary matched comparison (State_Output_No_Rationale versus Clean_Policy_Only) changes both taxonomy exposure and state output simultaneously. The correct control for the effect of the output field is Taxonomy_No_State_Output, which supplies the same state definitions without requiring a state field. Appendix F, Table 4 provides the needed numbers: for Llama, state-output accuracy is 44.6% versus taxonomy-only 53.1%, giving -8.5 percentage points; for GPT-OSS, state-output is 58.1% versus taxonomy-only 59.8%, giving -1.7 percentage points. Section 7 states that the design "dissociates taxonomy exposure from state emission," but no cluster interval or p-value is reported for the state-output-minus-taxonomy-only contrast. The abstract's +0.6 and +3.3 figures therefore conflate the taxonomy exposure effect with the state emission effect. This is an internal mismatch between the headline claim and the experimental contrast, not a stylistic issue. The paper should either report the dissociation contrast as the isolated state-output effect or rewrite the abstract and Section 7 so that the tested contrast is described accurately.
  2. [§8; §7 (taxonomy-only result)] The reported taxonomy-only gain (+9.17 pp for Llama, +5.00 pp for GPT-OSS) is treated as a finding that "exposing the four state definitions improves accuracy," but Section 8 concedes that the rendered taxonomy prompt lists the four states in the same order as the four policies and closely paraphrases them, so the mapping is recoverable without being stated. Because the benchmark states map deterministically to the reference policies, the taxonomy-only condition may simply restate the answer structure. The paper acknowledges this confound in prose, but the main-text result is still presented as if taxonomy exposure were a separable treatment. A control that permutes or rewords the state definitions, or that uses semantically decomposed evidence fields not isomorphic to the policy labels, would be needed to substantiate the claim that the taxonomy gain is due to useful typed knowledge rather than definitional leakage. At minimum, the taxonomy-only result should be labeled as confounded with definition restatement throughout the abstract and results sections.
  3. [§5; Limitations; Appendix A] Every accuracy number in the paper is measured against rule-derived reference policies generated deterministically from structured semantic slots before rendering. The paper correctly discloses that labels are rule-derived and that annotator agreement "partly reflects the construction," but this limitation is load-bearing for the central null result and for the taxonomy gain: if the rule set mis-encodes the intended abstraction for any of the 40 families, the state-output null, the taxonomy-only gain, and the family-level statistics are all measured against a disputable standard. I do not think this requires rejection, because the paper is transparent about the limitation and the controlled counterfactual design is a reasonable audit standard. However, the authors should add at least one concrete robustness check, such as showing the rule-derived labels for all 40 families in an appendix or reporting results under an alternative annotation or labeling rule, so readers can assess how much of the central result depends on the specific rule encoding.
minor comments (6)
  1. [§1] The phrase "positive Ignorereference class" is missing a space between "Ignore" and "reference".
  2. [§7; Table 1 caption] The caption's explanation of the Holm adjustment is hard to follow: "Holm adjustment is over the four arm-versus-clean contrasts shown for taxonomy-only and state-output across the two endpoints; the state-output Holm values are unchanged under this expanded family." Please restate exactly which hypotheses form the family and what the expanded family contains.
  3. [§7; Table 2] Table 2 uses "Policy-only" for a baseline from a different protocol than the matched ablation in Table 1. The text notes this in passing, but the two tables appear side by side and a reader could easily compare their policy-only rows as if they were the same condition. Please add a clear per-table label, such as "Policy-only (label-conditioning protocol)".
  4. [Appendix D] The text says "Exact artifact IDs are stored in the raw-result metadata" and "Appendix D maps them to reader-facing names," but Appendix D lists only paper-facing names and artifact IDs, not machine IDs. Please state explicitly whether machine IDs are in a separate artifact, and include one concrete example mapping if possible.
  5. [Appendix F.2] The third-endpoint replication reports a large positive delta for llama-3.1-8b-instant but also 18/480 parse failures in the state-output arm, and the text appropriately warns against reading it as a clean gain. The warning is good, but the phrase "reproduces the direction of the state-output effect" could still be misleading; consider calling it a direction-only, high-parse-failure sensitivity check in the main text.
  6. [Appendix N] The semantic-evidence follow-up reports grouped word TF-IDF decoders reaching 55.6% accuracy and character TF-IDF 62.5%, which are far above the 25.0% majority baseline, and the text dismisses these as "design constraints." Please clarify in one sentence whether these diagnostics indicate residual lexical separability in the semantic-evidence set or whether they are expected given the template structure, since the current phrasing is open to misinterpretation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the conclusions come from pre-specified matched counterfactual runs on a frozen set, with the paper explicitly flagging its own label-construction and taxonomy-restatement confounds.

full rationale

The paper's central negative result is empirical rather than definitional. The frozen 160-example controlled set is constructed before evaluation, reference policies are rule-derived from structured semantic slots before rendering, and the matched ablation compares pre-specified prompt arms on that frozen set. The paper repeatedly and explicitly disclaims over-interpretation: it states that annotator agreement 'partly reflects the construction,' that labels 'remain rule-derived and may not capture every reasonable human judgment,' and that the deterministic state-to-policy mapping makes the label-conditioning diagnostic 'a label-conditioning diagnostic rather than evidence of a faithful internal mechanism.' The taxonomy-only gain is likewise not presented as an independent derivation; the paper itself says the taxonomy prompt 'lists the four states in the same order as the four policies and closely paraphrases them' and that 'the taxonomy-only gain is confounded with definition restatement.' That is an acknowledged validity limitation, not a circular derivation. No load-bearing self-citation or imported uniqueness theorem appears; the cited prior work is contextual and not used to force the paper's conclusions. The one notable issue is a reporting mismatch: the abstract calls the state-output versus policy-only contrast an 'isolated explicit state-output field,' although that contrast changes both taxonomy exposure and the output field, whereas the truly isolated field contrast (state-output minus taxonomy-only) is not the headline delta and is negative in direction. This is an internal-validity/interpretation concern about which comparison answers the stated question, not a case where a claimed prediction is equivalent to its inputs by construction. Accordingly, no circular step meeting the evidentiary standard is identified, and the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No parameters are fitted to produce the reported claims; statistics are pre-specified cluster bootstrap and permutation tests, and the wrong-state assignment uses a fixed arbitrary seed rather than a fitted value. No invented physical or ontological entities are introduced; the four state labels are a benchmark-local routing abstraction that the paper explicitly declines to present as a universal or internal mechanism. The load-bearing assumptions are the domain assumptions listed above, all stated and hedged by the authors.

assumptions (4)
  • domain assumption The four-policy ontology Use/Ignore/Update/Ask and the four benchmark states are a meaningful abstraction for memory-policy routing.
    Section 3 states these are 'a benchmark-specific routing abstraction, not a complete, universal, or mutually exhaustive ontology'; every accuracy figure is measured against this taxonomy.
  • domain assumption Rule-derived reference policies, generated from structured semantic slots before rendering, are the correct labels for the 160 controlled examples.
    Section 5 and Limitations: 'labels remain rule-derived and may not capture every reasonable human judgment'; the deterministic rule set is the ground truth for all reported deltas and p-values.
  • domain assumption The deterministic state-to-policy mapping is Active to Use, Stale-or-Irrelevant to Ignore, Conflicting-or-Superseded to Update, and Underspecified to Ask.
    Appendix E states the mapping; it is the premise that makes supplied states answer-associated and forces the label-conditioning, not mechanism, interpretation.
  • standard math The 40 scenario families are exchangeable units for bootstrap and permutation inference.
    Section 6 and Appendix L resample families for 95% intervals and sign-flip p-values; validity assumes family-level independence of the matched variants.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification." pith.science (2026). https://pith.science/paper/JJ4SQYAJ

@misc{pith2026260817247,
  author       = {Pith},
  title        = {Pith review of: Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JJ4SQYAJ}},
  note         = {Machine review of arXiv:2608.17247}
}
read the original abstract

Personalized agents must decide whether retrieved user memory should be used, ignored, updated, or queried before it affects a current task. We use this setting to develop an empirical audit protocol for structured intermediate outputs: first audit dataset shortcuts, then isolate bundled prompt changes, check whether intermediate labels are answer-associated, test decomposed semantic evidence, and audit provider-level execution failures. A 480-example synthetic development set initially suggested large gains from a state-structured prompt bundle, but TF-IDF diagnostics showed lexical separability and no positive standalone Ignore cases. We therefore construct a frozen 160-example controlled counterfactual set with 40 matched four-way families and rule-derived reference policies. On this set, exposing the four state definitions improves accuracy, but an isolated explicit state-output field does not significantly improve policy accuracy for Llama-3.3-70B and gives only a marginal, non-significant gain for GPT-OSS-120B. Supplying benchmark-associated state labels shifts policy predictions, but because those labels deterministically map to policies, this is a label-conditioning diagnostic rather than evidence of a faithful internal mechanism. Family-level and seed-stability analyses further show that example-level accuracy overstates counterfactual consistency: complete four-way family success is rare. An exploratory follow-up that elicits decomposed semantic evidence also fails to improve routing for the cleanly evaluated endpoint; the corresponding GPT-OSS condition was unavailable because of provider-side request validation. We evaluate policy classification only, not downstream responses, tool actions, or memory-store mutation.

Figures

Figures reproduced from arXiv: 2608.17247 by the authors.

Figure 1
Figure 1. Five-stage audit protocol instantiated in this paper. The label-conditioning diagnostic [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    and Stoica, Ion and Gonzalez, Joseph E

    Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , journal =. 2023 , url =

    2023

  2. [2]

    2025 , url =

    Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , booktitle =. 2025 , url =

    2025

  3. [3]

    2025 , url =

    Wu, Di and Wang, Hongwei and Yu, Wenhao and Zhang, Yuwei and Chang, Kai-Wei and Yu, Dong , booktitle =. 2025 , url =

    2025

  4. [4]

    LaMP: When Large Language Models Meet Personalization

    Salemi, Alireza and Mysore, Sheshera and Bendersky, Michael and Zamani, Hamed , booktitle =. 2024 , month = aug, address =. doi:10.18653/v1/2024.acl-long.399 , url =

    work page doi:10.18653/v1/2024.acl-long.399 2024

  5. [5]

    PersonaLLM: Investigating the Ability of Large Language Models to Express Personality Traits

    Jiang, Hang and Zhang, Xiajie and Cao, Xubo and Breazeal, Cynthia and Roy, Deb and Kabbara, Jad , booktitle =. 2024 , month = jun, address =. doi:10.18653/v1/2024.findings-naacl.229 , url =

    work page doi:10.18653/v1/2024.findings-naacl.229 2024

  6. [6]

    World Wide Web , volume =

    When Large Language Models Meet Personalization: Perspectives of Challenges and Opportunities , author =. World Wide Web , volume =. 2024 , doi =

    work page 2024

  7. [7]

    Retrieval-Augmented Generation for Knowledge-Intensive

    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 =

    2020

  8. [8]

    Knowledge Conflicts for LLMs: A Survey

    Xu, Rongwu and Qi, Zehan and Guo, Zhijiang and Wang, Cunxiang and Wang, Hongru and Zhang, Yue and Xu, Wei , booktitle =. Knowledge Conflicts for. 2024 , month = nov, address =. doi:10.18653/v1/2024.emnlp-main.486 , url =

    work page doi:10.18653/v1/2024.emnlp-main.486 2024

Show all 46 references
  1. [9]

    2024 , url =

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , booktitle =. 2024 , url =

    2024

  2. [10]

    Asking Clarification Questions in Knowledge-Based Question Answering , author =. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages =. 2019 , mo...

    2019 doi

  3. [11]

    2022 , url =

    Kuhn, Lorenz and Gal, Yarin and Farquhar, Sebastian , journal =. 2022 , url =

    2022

  4. [12]

    USENIX Security Symposium , pages =

    Extracting Training Data from Large Language Models , author =. USENIX Security Symposium , pages =. 2021 , url =

    2021

  5. [13]

    International Conference on Learning Representations , year =

    Beyond Memorization: Violating Privacy Via Inference with Large Language Models , author =. International Conference on Learning Representations , year =

  6. [14]

    International Conference on Learning Representations , year =

    Learning Differentially Private Recurrent Language Models , author =. International Conference on Learning Representations , year =

  7. [15]

    Evaluating Very Long-Term Conversational Memory of

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , booktitle =. Evaluating Very Long-Term Conversational Memory of. 2024 , month = aug, address =. doi:10.18653/v1/2024.acl-long.747 , url =

    2024 doi

  8. [16]

    Zhao, Siyan and Hong, Mingyi and Liu, Yang and Hazarika, Devamanyu and Lin, Kaixiang , booktitle =. Do. 2025 , url =

    2025

  9. [17]

    Grattafiori, Aaron and Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Vaughan, Alex and Yang, Amy and Fan, Angela and others , journal =. The. 2024 , url =

    2024

  10. [18]

    2024 , howpublished =

    2024

  11. [19]

    2025 , url =

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and Zheng, Chujie and Liu, Dayiheng and Zhou, Fan and Huang, Fei and Hu, Feng and Ge, Hao and Wei, Haoran and Lin, Huan an...

    2025

  12. [20]

    2025 , howpublished =

    2025

  13. [21]

    Advances in Neural Information Processing Systems , volume =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  14. [22]

    International Conference on Learning Representations , year =

    Least-to-Most Prompting Enables Complex Reasoning in Large Language Models , author =. International Conference on Learning Representations , year =

  15. [23]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =

    Measuring and Narrowing the Compositionality Gap in Language Models , author =. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages =. 2023 , publisher =. doi:10.18653/v1/2023.findings-emnlp.378 , url =

    2023 doi

  16. [24]

    arXiv preprint arXiv:2608.08942 , year =

    Same Question, Different Answer? Measuring and Mitigating Prompt Privilege for Equitable AI Access , author =. arXiv preprint arXiv:2608.08942 , year =. doi:10.48550/arXiv.2608.08942 , url =

    doi arXiv

  17. [25]

    2026 , eprint =

    Evidence-Unit Fairness and the Limits of Query-Adaptive Sparse-Dense Fusion in Financial Document Retrieval , author =. 2026 , eprint =

    2026

  18. [26]

    arXiv preprint arXiv:2606.28725 , year =

    DriftGuard: Safety-Aware Multi-Monitor Detection and Selective Adaptation for Evolving Toxicity Moderation , author =. arXiv preprint arXiv:2606.28725 , year =. doi:10.48550/arXiv.2606.28725 , url =

    doi arXiv

  19. [27]

    2026 , eprint =

    How Early Is Early Enough? Design-Dependent Observation-Window Sufficiency in Subscription Churn Prediction , author =. 2026 , eprint =. doi:10.48550/arXiv.2607.00473 , url =

    2026 doi arXiv

  20. [28]

    2026 , eprint =

    Class Weighting versus Amount Conditioning in Credit-Card Fraud Detection: A Dollar-Metric Study with a Temporal Explanation Audit , author =. 2026 , eprint =. doi:10.48550/arXiv.2607.14686 , url =

    2026 doi arXiv

  21. [29]

    2026 , eprint =

    Interpretable vs Learned Encoders for High-Cardinality Fraud Detection , author =. 2026 , eprint =. doi:10.48550/arXiv.2607.00477 , url =

    2026 doi arXiv

  22. [30]

    doi:10.48550/arXiv.2602.06714 , url =

    Li, Jialin and Chen, Zhenhao and Luo, Hanjun and Salam, Hanan , year =. doi:10.48550/arXiv.2602.06714 , url =. 2602.06714 , archivePrefix =

    doi

  23. [31]

    The Memory Curse: How Expanded Recall Erodes Cooperative Intent in

    Liu, Jiayuan and Li, Tianqin and Du, Shiyi and Luo, Xin and Zeng, Haoxuan and Tewolde, Emanuel and Lee, Tai Sing and Wang, Tonghan and Kingsford, Carl and Conitzer, Vincent , journal =. The Memory Curse: How Expanded Recall Erodes Cooperative Intent in. 2026 , eprint =. doi:10...

    2026 doi arXiv

  24. [32]

    Wu, Chuchu and Zhou, Zhiyin and Hu, Jingzhuo and You, Liang , year =

  25. [33]

    Agentic-

    Su, Yiyun and Zhu, Huiying and Tian, Yu and Zhao, Changruo and Peng, Zujun and Liu, Yuting and Zhang, Luyan and Fan, Liang and Li, Baihua , editor =. Agentic-. Advanced Intelligent Computing Technology and Applications , series =. 2026 , publisher =. doi:10.1007/978-981-92-343...

    2026 doi

  26. [34]

    2026 , month = jun, eprint =

    Liu, Yuanzhe and Zhu, Jingyuan and Mo, Yuchen and Li, Gen and Cao, Xu and Jin, Jin and Shen, Yifan and Li, Zhengyuan and Yu, Tianjiao and Yuan, Wenzhen and Ding, Fangqiang and Lourentzou, Ismini , booktitle =. 2026 , month = jun, eprint =

    2026

  27. [35]

    2026 , month = apr, publisher =

    Fu, Rong and Wang, Yemin and Xu, Tianxiang and Liu, Yongtai and Tang, Weizhi and Wu, Wangyu and Ma, Xiaowen and Fong, Simon , booktitle =. 2026 , month = apr, publisher =. doi:10.1145/3774904.3792459 , url =. 2603.23512 , archivePrefix =

    2026

  28. [36]

    From Agent Traces to Trust: A Survey of Evidence Tracing and Execution Provenance in

    Wang, Yiqi and Zhang, Jiaqi and Cai, Taotao and Liu, Zirui and Sun, Qingqiang and Sun, Zequn and Wu, Zhangkai and Dong, Manqing and Zheng, Mingkai and Yin, Xuefei and Zhu, Yanming , year =. From Agent Traces to Trust: A Survey of Evidence Tracing and Execution Provenance in. d...

    doi arXiv

  29. [37]

    doi:10.48550/arXiv.2607.23929 , url =

    Li, Xiaoyang and Wang, Yiqi and Lu, Haohui and Chen, Zhi and Li, Mo and Song, Pingan and Zheng, Mingkai and Cai, Taotao , year =. doi:10.48550/arXiv.2607.23929 , url =. 2607.23929 , archivePrefix =

    doi arXiv

  30. [38]

    2026 , eprint =

    From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents , author =. 2026 , eprint =. doi:10.48550/arXiv.2608.10502 , url =

    2026 doi arXiv

  31. [39]

    doi:10.48550/arXiv.2608.10509 , url =

    Wang, Yiqi and Yan, Zihao and Zhang, Jiaqi and Wu, Zhangkai and Zheng, Mingkai and Sun, Zequn and Zhu, Yanming and Cai, Taotao , year =. doi:10.48550/arXiv.2608.10509 , url =. 2608.10509 , archivePrefix =

    doi arXiv

  32. [40]

    2026 , eprint =

    From Reaction to Anticipation: Proactive Failure Recovery through Agentic Task Graph for Robotic Manipulation , author =. 2026 , eprint =. doi:10.48550/arXiv.2605.11951 , url =

    2026 doi arXiv

  33. [41]

    Nature Machine Intelligence , volume =

    Shortcut Learning in Deep Neural Networks , author =. Nature Machine Intelligence , volume =. 2020 , doi =

    2020

  34. [42]

    International Conference on Learning Representations , year =

    Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations , author =. International Conference on Learning Representations , year =

  35. [43]

    and Roth, Dan , booktitle =

    Jiang, Bowen and Hao, Zhuoqun and Cho, Young-Min and Li, Bryan and Yuan, Yuan and Chen, Sihao and Ungar, Lyle and Taylor, Camillo J. and Roth, Dan , booktitle =. Know Me, Respond to Me: Benchmarking. 2025 , url =

    2025

  36. [44]

    2026 , publisher =

    Shen, Yiting and Li, Kun and Zhou, Wei and Hu, Songlin , booktitle =. 2026 , publisher =. doi:10.18653/v1/2026.acl-long.370 , url =

    2026 doi

  37. [45]

    International Conference on Learning Representations , year =

    Mass-Editing Memory in a Transformer , author =. International Conference on Learning Representations , year =

  38. [46]

    arXiv preprint arXiv:2207.05221 , year =

    Language Models (Mostly) Know What They Know , author =. arXiv preprint arXiv:2207.05221 , year =

    arXiv

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.