Pith. sign in

REVIEW 2 major objections 5 minor 29 references

TGMS claims that an LLM agent should only plan operator calls and verbalize results, while the database owns time, arithmetic, grounding, and verification — a split that answers belief-state questions no baseline can.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 07:19 UTC pith:L4TA6HCG

load-bearing objection Careful, honest systems paper with real engineering — the empirical headline is real but conditional on a co-designed workload, and the paper says so itself. the 2 major comments →

arxiv 2607.10265 v2 pith:L4TA6HCG submitted 2026-07-11 cs.DB

TGMS: An Agent-Native Bi-Temporal Graph Management System

Xiaofei Zhang This is my paper
classification cs.DB
keywords bi-temporal graphtransaction timevalid timeLLM agenttemporal graph question answeringoperator contractsclaim verificationbelief state
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central claim is that an LLM agent answering temporal-graph questions should not compute at all; it should only plan calls to a fixed set of deterministic, bounded, bi-temporal operators and verbalize the returned results. The database system executes those operators, records content-addressed traces annotated with completeness metadata, and gates the final answer against those traces. On a frozen 94-task test split of a real communication network, this design reaches 0.408 normalized typed-answer accuracy with a 14B open-source model, versus 0.064–0.152 for vector retrieval, static-graph retrieval, and text-to-Cypher baselines. On probe questions asking what the system believed before a later correction — questions that require transaction time — TGMS scores 0.897 while latest-state baselines score exactly zero. A sympathetic reader would care because the experiments identify two interface requirements — declared output contracts on operators and completeness-aware verification — that apply beyond temporal graphs to any agent-facing database interface.

Core claim

TGMS is a bi-temporal property graph system, meaning every fact is versioned along two axes: valid time (when the fact held in the world) and transaction time (when the system came to believe it), with an append-only event log that replays to identical state. The discovery is that exposing this store to an LLM through thirteen typed, deterministic, bounded operators — with the model kept strictly outside the trusted boundary — yields accurate answers to temporal-graph questions, including belief-state questions such as 'as of transaction time T, what did the system believe?' that no latest-state snapshot or retrieval pipeline can answer. The paper's evidence: on a frozen 94-task test split o

What carries the argument

The central mechanism is the fixed operator algebra — thirteen operator contracts (O1–O13) with typed input and output schemas, deterministic results, pre-execution cost guards, and bi-temporal 'as of transaction time' semantics — combined with a content-addressed, completeness-annotated execution trace and a static plan verifier plus claim verifier. Plans are JSON DAGs; the static verifier rejects references to nonexistent output fields, ungrounded literal identifiers, invalid temporal arguments, and over-budget cost before execution. The executor produces SHA-256 result digests and propagates truncation taint through dependent steps; the claim verifier re-checks counts, values, entities, a

Load-bearing premise

The load-bearing premise is that the 17 co-designed task templates and the 13 operators jointly capture what people actually ask of temporal graphs; the paper states the workload and operator set were co-designed and that failing instances were discarded, so an independent question set could shrink the reported gaps.

What would settle it

Run TGMS and the three baselines on a temporal-graph question set collected from users who never saw the operator list (and on a non-communication domain), with gold answers produced by an independent bi-temporal engine or human annotation; if TGMS's margins over the baselines largely vanish, or its correction-probe accuracy falls well below 0.897, the separation-of-concerns claim is limited to the co-designed workload.

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

If this is right

  • Belief-state questions — 'what did the system believe as of transaction time T?' — become answerable; on correction probes TGMS scores 0.897 (CollegeMsg) and 0.846 (email-EU) while latest-state baselines score exactly zero.
  • Output contracts on operators are load-bearing: disabling them at 14B lowers accuracy from 0.409 to 0.318 on the development split, while first-attempt plan 'validity' misleadingly rises because the validator checks less.
  • Evidence-completeness tracking prevents false full support: with truncation taint disabled, 15/15 truncated-page counts would be accepted as fully supported; with tainting, 0/15.
  • The operator-backed interface makes better use of larger models than retrieval baselines: full-precision accuracy rises from 0.138 (7B) to 0.628 (32B) while baselines stay below 0.277.
  • The main result is not an artifact of limited retrieval breadth: a long-context vector-RAG control with 20 chunks scored 0.021 against TGMS's 0.362 in the same run.

Where Pith is reading between the lines

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

  • Because the workload and operator set were co-designed (the paper says so), the measured gaps are upper bounds for this closed world; an independently collected temporal-graph question set could shrink them, so the durable claim is about the belief-state capability and the contract/verifier mechanisms, not about general temporal-graph QA.
  • The architecture suggests a general recipe for agent-native database interfaces: expose a small set of deterministic, contract-typed, bounded operators and verify final claims against execution provenance; relationally or document-oriented stores with transaction-time history could adopt the same pattern.
  • The zero-unsupported-claim guarantee is partly bought by abstention (coverage 0.706, conditional accuracy 0.548), so deployers should weigh refusal rates; the next test is extending gating to pattern claims and approximate operators to see whether reliability survives wider coverage.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper presents TGMS, a bi-temporal property graph management system that exposes thirteen typed, deterministic, bounded temporal operators as agent tools. An LLM is used only as a planner and reporter; the system performs all graph computation, static plan verification, execution with content-addressed traces, and claim verification. The central claim is that this separation (LLM outside the trusted boundary) yields accurate answers to temporal-graph questions, including belief-state questions that require transaction-time reasoning. The evaluation on frozen splits of CollegeMsg, email-EU, and a synthetic suite reports TGMS accuracy of 0.408 vs. 0.106/0.064/0.152 for vector-RAG, static-graph RAG, and text-to-Cypher, and 0.897 on correction probes vs. zero for latest-state baselines. The paper also reports ablations supporting output contracts and completeness-aware verification, plus honest zero coverage on some tasks.

Significance. If the results hold, the architecture is a valuable contribution: treating the LLM as an untrusted planner/reporter with deterministic operators and trace-based claim checking is a promising direction for reliable temporal graph QA. The paper has strong experimental hygiene: frozen splits with SHAs, thresholds recorded in a dated decision log, three seeds with paired bootstrap CIs, brute-force operator validation on 500 randomized cases each, byte-identical replay across two backends, and explicit disclosure of abstention and coverage trade-offs. The honest reporting of the T2 zero-coverage and the unsupported-claim rates is commendable. However, the co-designed workload and abstract inaccuracies temper the external validity of the headline numbers.

major comments (2)
  1. [§5.2, §8] The evaluation workload is co-designed with the operator set: the 17 semantic templates were written to align with the 13 operators, gold answers are computed by TGMS's own engine executing oracle plans, and instances whose oracle execution fails or degenerates are discarded (§5.2). The paper admits this in §8. Because the baselines are compared on a task suite selected for expressibility by TGMS's operators, the headline gaps (Table 4: 0.408 vs 0.106–0.152; 0.897 vs 0) cannot be taken as evidence of general temporal-graph QA ability. The central claim remains plausible, but the external validity is not established. Please either evaluate on an independently collected or human-labeled temporal-graph QA set, or substantially temper the abstract/conclusion claims to state that accuracy is measured on a co-designed workload.
  2. [Abstract vs §5.3, Table 4] The abstract reports '0.409 exact match' and '0.67 exact match on correction probes', while Table 4 reports normalized typed-answer accuracy of 0.408 and 0.897 on probes; the abstract also says 'three 14B baselines score zero' on probes, but the body reports vector-RAG at 0.154 (Table 4). These are not minor typos: they misstate the comparison and must be corrected. Also, the metric is not exact match because interval answers use an IOU threshold of 0.5 (§5.2); the abstract should use the paper's own term 'normalized typed-answer accuracy'.
minor comments (5)
  1. [§2] The text says 'uses 262 as the open-ended upper bound'—presumably a typo for 2^62. Please fix.
  2. [Table 4, Table 6] The model name is written as 'A WQ' (with a space) in several places; should be 'AWQ'.
  3. [References] Reference [15] has 'F ActScore'—formatting artifact; should be 'FactScore'. Similar LaTeX artifacts appear in the text (e.g., 'K` uzu' in §2).
  4. [Figure 1] The label 'asoftt' inside the trusted boundary appears to be a typo for 'as of tt'.
  5. [Abstract] The phrase 'thirteen verified temporal operators' is stronger than the evidence: §3 describes randomized brute-force testing on 500 cases, not formal verification. Suggest 'validated' or 'tested'.

Circularity Check

0 steps flagged

No significant circularity; disclosed workload co-design is an external-validity limitation, not a derivation that reduces to its inputs.

full rationale

The central empirical claim is that TGMS's architecture (LLM plans and reports; deterministic system computes and verifies) yields accurate answers on frozen bi-temporal graph tasks. The derivation chain is design -> operator algebra -> static/claim verification -> measured accuracy. No equation is defined in terms of the claimed result, and no fitted parameter is renamed as a prediction. The closest issue is benchmark construction: §5.2 states 'Tasks are program-generated with engine-computed gold: oracle plans are executed by TGMS itself,' and §8 admits 'the tasks and the operator set were co-designed; the study does not yet establish coverage of independently collected temporal-graph questions.' This is a genuine external-validity limitation, but it is not circularity: the operator engine is independently checked against brute force (§3: 'Each operator is compared with an independent brute-force implementation on 500 randomized combinations'), the test split was frozen before measurement with a dated decision log, and the measured accuracy is not forced (0.408, with abstention on the T2 family and a 0.897 probe score). Baselines are scored against the same gold, so the relative gap is a real observation on this workload. The paper explicitly reports coverage/abstention trade-offs and the co-design limitation rather than hiding them. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled via citation. The co-design concern belongs to correctness/external-validity risk, not to circularity; under the requested scale, this is a 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The paper's claims rest on standard bi-temporal semantics (Snodgrass), standard temporal-graph definitions (time-respecting paths, δ-temporal motifs), and one ad-hoc-to-paper assumption: the 13-operator/17-template co-design that defines the evaluation. No numeric parameter is fit to data in the sense of a calibrated model; the hand-chosen constants are metric thresholds (interval IOU 0.5; float tolerance 1e-9) and operator bounds (k≤3, k≤20, page size 100). The invented artifacts (operator algebra, completeness taint, content-addressed digests) all carry falsifiable handles inside the paper (brute-force oracles, the 15/15-vs-0/15 taint test, replay byte-identity).

free parameters (4)
  • interval-answer IOU threshold = 0.5
    Metric acceptance rule (§5.2): interval answers count as correct at intersection-over-union ≥ 0.5. Chosen by hand; loosens the "exact match" label used in the abstract.
  • float comparison tolerance = 1e-9
    Count/value answers must agree within 1e-9 after float canonicalization (§5.2). Metric constant, not fit to data.
  • operator cost limits = k≤3 (O2), k≤20 (O5), page size 100
    Bounded-execution constants in the operator registry (§3). Hand-chosen design bounds that define what the planner can express; the page-size-100 setting drives the truncation-taint finding (§5.4).
  • burst-detection thresholds (O9) = rolling z-score or trailing-median rule; thresholds not stated
    Operator-design constants (§ Table 1). Not fitted to the evaluation data, but unspecified in the paper.
axioms (5)
  • domain assumption Bi-temporal model: every version carries half-open valid-time [vt_s, vt_e) and transaction-time [tt_s, tt_e) intervals; snapshot G(t, tt) contains versions whose intervals contain t and tt (§2).
    Inherited from Snodgrass/bitemporal database literature [24]; the belief-state capability depends on it.
  • domain assumption Temporal semantics adopt time-respecting paths [27] and δ-temporal motifs [17] as ground truth for operators O4–O7.
    These definitions, with exact multi-label search for max-wait reachability (§3), are adopted rather than derived; a task set with different temporal semantics would mis-score operator answers.
  • ad hoc to paper The 13-operator algebra covers the expressible task space; the 17 semantic templates were co-designed with it.
    Disclosed in §5.2 and §8: "the operator set and the workload were co-designed." This workload-alignment assumption is the main external-validity risk.
  • domain assumption Gold answers are correct because the operator engine is validated against independent brute-force implementations (500 randomized store/argument combinations per operator, §3).
    The evaluation oracle is TGMS's own engine; correctness rests on the brute-force tests and property-based invariants (pairwise-disjoint valid intervals, bi-temporal immutability).
  • domain assumption The dated decision log and SHA-frozen splits suffice to prevent test-set overfitting; the log is "an internal dated record, not an external registry" (footnote, §5.1).
    Weaker than external pre-registration; the acceptance-threshold discipline is self-administered and cannot be independently audited from the paper alone.
invented entities (3)
  • TGMS operator algebra O1–O13 (typed, deterministic, bounded, bi-temporal operator interface) independent evidence
    purpose: The fixed computation surface the LLM planner composes; replaces free-form query generation and serialized graph context.
    Each operator is tested against an independent brute-force oracle on 500 randomized cases and the algebra is the load-bearing artifact behind the accuracy claims; a falsifiable handle exists (brute-force oracles in the repo).
  • Completeness metadata / truncation taint on result pages independent evidence
    purpose: Marks evidence as incomplete when a step reads a truncated page, so dependent claims cap below full support.
    The controlled generator provides a falsifiable handle: 15/15 truncated-page counts refused full support with tainting, 0/15 without (§5.4).
  • Content-addressed trace digests (SHA-256 result digests) independent evidence
    purpose: Give claims stable evidence identifiers for the claim verifier and pin as-of-tt byte-identical immutability.
    The replay-across-backends byte-identity test (§2) is an external check; digests are also used by the verifier's re-execution comparison.

pith-pipeline@v1.3.0-alltime-deepseek · 11855 in / 21005 out tokens · 214643 ms · 2026-08-02T07:19:57.070169+00:00 · methodology

0 comments
read the original abstract

Temporal graph questions require reliable handling of time, identifiers, and arithmetic. Large language model (LLM) agents often fail on these tasks, especially when a graph records both ordinary evolution and later corrections. We present TGMS, a bi-temporal property graph management system that exposes thirteen verified temporal operators as agent tools. Each operator is typed, deterministic, bounded, cost-guarded, and bi-temporal by default. The LLM plans operator calls and writes the final response, while the system performs all graph computation. Numeric, entity, ordering, and pattern claims are checked against the content-addressed execution trace. TGMS separates valid time from transaction time. It can therefore answer belief-state questions such as ``as of transaction time $T$, what did the system believe?'' Standard latest-state snapshots and retrieval pipelines do not preserve enough information to answer such questions. On a development benchmark built from a real communication network, TGMS with a 14B open-source model reaches 0.409 exact match. Vector-RAG, static-graph RAG, and text-to-Cypher reach 0.045--0.182 under the same serving setup. TGMS reaches 0.67 exact match on correction probes, while the three 14B baselines score zero. The claim verifier detects all 500 injected count and entity errors with no false positives on the clean answers. Two implementation findings were especially important. First, operator output contracts prevent plans from referring to fields that do not exist. Second, verification must track whether the cited evidence is complete, because correct arithmetic over a truncated result is still misleading. The code, benchmark, and trace viewer are open source under Apache-2.0.

Figures

Figures reproduced from arXiv: 2607.10265 by Xiaofei Zhang.

Figure 1
Figure 1. Figure 1: TGMS architecture. The LLM plans and reports. Operators perform the graph compu [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 5 linked inside Pith

  1. [1]

    Model context protocol.https://modelcontextprotocol.io, 2024

    Anthropic. Model context protocol.https://modelcontextprotocol.io, 2024

    2024

  2. [2]

    Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W. Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks.Trans- actions on Machine Learning Research, 2023

    2023

  3. [3]

    A model and query language for temporal graph databases.The VLDB Journal, 30:825–858, 2021

    Ariel Debrouvier, Eliseo Parodi, Mat ´ ıas Perazzo, Valeria Soliani, and Alejandro Vaisman. A model and query language for temporal graph databases.The VLDB Journal, 30:825–858, 2021

    2021

  4. [4]

    Chain-of-verification reduces hallucination in large language models, 2023

    Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. Chain-of-verification reduces hallucination in large language models, 2023. arXiv:2309.11495

    Pith/arXiv arXiv 2023

  5. [5]

    From local to global: A graph RAG approach to query-focused summarization, 2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. From local to global: A graph RAG approach to query-focused summarization, 2024. arXiv:2404.16130

    Pith/arXiv arXiv 2024

  6. [6]

    K` uzu graph database management system

    Xiyang Feng, Guodong Jin, Ziyi Chen, Chang Liu, and Semih Saliho˘ glu. K` uzu graph database management system. InConference on Innovative Data Systems Research (CIDR), 2023

    2023

  7. [7]

    Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu

    Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Y. Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. RARR: Researching and revising what language models say, using language models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), 2023

    2023

  8. [8]

    PAL: Program-aided language models

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. PAL: Program-aided language models. InInternational Conference on Machine Learning (ICML), 2023

    2023

  9. [9]

    HippoRAG: Neurobiologically inspired long-term memory for large language models

    Bernal Jim´ enez Guti´ errez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. HippoRAG: Neurobiologically inspired long-term memory for large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  10. [10]

    Temporal networks.Physics Reports, 519(3):97–125, 2012

    Petter Holme and Jari Saram¨ aki. Temporal networks.Physics Reports, 519(3):97–125, 2012

    2012

  11. [11]

    Aeong: An efficient built-in temporal support in graph databases.Proceedings of the VLDB Endowment, 17(6):1515–1527, 2024

    Jiamin Hou, Zhanhao Zhang, Zhouyu Wang, Yongjun Zhang, Wei Lu, Anqun Pan, and Xi- aoyong Du. Aeong: An efficient built-in temporal support in graph databases.Proceedings of the VLDB Endowment, 17(6):1515–1527, 2024. 14

    2024

  12. [12]

    Tem- poral graph benchmark for machine learning on temporal graphs

    Shenyang Huang, Farimah Poursafaei, Jacob Danovitch, Matthias Fey, Weihua Hu, Emanuele Rossi, Jure Leskovec, Michael Bronstein, Guillaume Rabusseau, and Reihaneh Rabbany. Tem- poral graph benchmark for machine learning on temporal graphs. InAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023

    2023

  13. [13]

    Mahoney, Kurt Keutzer, and Amir Gholami

    Sehoon Kim, Suhong Moon, Ryan Tabrizi, Nicholas Lee, Michael W. Mahoney, Kurt Keutzer, and Amir Gholami. An LLM compiler for parallel function calling. InInternational Conference on Machine Learning (ICML), 2024

    2024

  14. [14]

    ToolGate: Contract-grounded and verified tool execution for LLMs

    Yanming Liu, Xinyue Peng, Jiannan Cao, Xinyi Wang, Songhang Deng, Jintao Chen, Jianwei Yin, and Xuhong Zhang. ToolGate: Contract-grounded and verified tool execution for LLMs. arXiv preprint arXiv:2601.04688, 2026

    arXiv 2026

  15. [15]

    F ActScore: Fine-grained atomic evaluation of factual precision in long form text generation

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. F ActScore: Fine-grained atomic evaluation of factual precision in long form text generation. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023

    2023

  16. [16]

    Pietro Panzarasa, Tore Opsahl, and Kathleen M. Carley. Patterns and dynamics of users’ behavior and interaction: Network analysis of an online community.Journal of the American Society for Information Science and Technology, 60(5):911–932, 2009

    2009

  17. [17]

    Benson, and Jure Leskovec

    Ashwin Paranjape, Austin R. Benson, and Jure Leskovec. Motifs in temporal networks. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining (WSDM), pages 601–610, 2017

    2017

  18. [18]

    Duckdb: An embeddable analytical database

    Mark Raasveldt and Hannes M¨ uhleisen. Duckdb: An embeddable analytical database. In Proceedings of the 2019 International Conference on Management of Data (SIGMOD), pages 1981–1984, 2019

    2019

  19. [19]

    Zep: A temporal knowledge graph architecture for agent memory, 2025

    Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A temporal knowledge graph architecture for agent memory, 2025. arXiv:2501.13956

    Pith/arXiv arXiv 2025

  20. [20]

    Distributed temporal graph analytics with GRADOOP.The VLDB Journal, 31:375–401, 2022

    Christopher Rost, Kevin Gomez, Matthias T¨ aschner, Philip Fritzsche, Lucas Schons, Lukas Christ, Timo Adameit, Martin Junghanns, and Erhard Rahm. Distributed temporal graph analytics with GRADOOP.The VLDB Journal, 31:375–401, 2022

    2022

  21. [21]

    Question answering over tem- poral knowledge graphs

    Apoorv Saxena, Soumen Chakrabarti, and Partha Talukdar. Question answering over tem- poral knowledge graphs. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021

    2021

  22. [22]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess ` ı, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  23. [23]

    Cultivating multidisciplinary research and education on gpu infrastructure for mid-south institutions at the university of memphis: Practice and challenge, 2025

    Mayira Sharif, Guangzeng Han, Weisi Liu, and Xiaolei Huang. Cultivating multidisciplinary research and education on gpu infrastructure for mid-south institutions at the university of memphis: Practice and challenge, 2025

    2025

  24. [24]

    Snodgrass.Developing Time-Oriented Database Applications in SQL

    Richard T. Snodgrass.Developing Time-Oriented Database Applications in SQL. Morgan Kaufmann, 1999. 15

    1999

  25. [25]

    From agent traces to trust: A survey of evidence tracing and execution provenance in LLM agents.arXiv preprint arXiv:2606.04990, 2026

    Yiqi Wang, Jiaqi Zhang, Taotao Cai, Zirui Liu, Qingqiang Sun, Zequn Sun, Zhangkai Wu, Man- qing Dong, Mingkai Zheng, Xuefei Yin, and Yanming Zhu. From agent traces to trust: A survey of evidence tracing and execution provenance in LLM agents.arXiv preprint arXiv:2606.04990, 2026

    Pith/arXiv arXiv 2026

  26. [26]

    TOKI: A bitemporal operator algebra for contradiction resolution in LLM-agent persistent memory.arXiv preprint arXiv:2606.06240, 2026

    Ziming Wang. TOKI: A bitemporal operator algebra for contradiction resolution in LLM-agent persistent memory.arXiv preprint arXiv:2606.06240, 2026

    Pith/arXiv arXiv 2026

  27. [27]

    Path problems in temporal graphs.Proceedings of the VLDB Endowment, 7(9):721–732, 2014

    Huanhuan Wu, James Cheng, Silu Huang, Yiping Ke, Yi Lu, and Yanyan Xu. Path problems in temporal graphs.Proceedings of the VLDB Endowment, 7(9):721–732, 2014

    2014

  28. [28]

    ReAct: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  29. [29]

    Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task

    Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (...

    2018