Pith. sign in

REVIEW 5 major objections 6 minor 38 references

Towards Mitigation of Hallucination for LLM-empowered Agents: Progressive Generalization Bound Exploration and Watchdog Monitor

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read HalMit aims to catch hallucinations in black-box LLM agents using stored boundary queries.

desk verdict The 'generalization bound' is never defined; the watchdog is a retrieval classifier, so the core claim is unsupported despite a novel query-generation scheme. read the letter →

arxiv 2507.15903 v2 pith:PBHKUQPV submitted 2025-07-21 cs.LG cs.AI

Siyuan Liu , Wenjing Liu , Zhiwei Xu , Xin Wang , Bo Chen , Tao Li This is my paper · ORCID
classification cs.LGcs.AI
keywords hallucinationdetectionLLMagentsblack-boxmonitoringgeneralizationboundprobabilisticfractalsamplingsemanticentropyvectordatabasemulti-agentexploration
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

This paper tries to establish that hallucinations in LLM-powered agents can be detected and mitigated without any look inside the model and without checking the output against an external knowledge base. The proposed watchdog, HalMit, first explores each agent's per-domain 'generalization bound' — the semantic frontier beyond which the agent's answers become unreliable — by generating large numbers of probe queries with a probabilistic fractal sampling scheme. Hallucinated query-response pairs discovered along the way are stored in a vector database, and at run time a new input is flagged as hallucination-prone when it sits close to those stored boundary points or shows higher semantic entropy than the nearest stored example. If the claim holds, closed commercial agents that expose only an API can still receive persistent black-box hallucination monitoring, and the paper reports experiments in which HalMit outperforms baseline monitors across multiple domains and model backbones.

What carries the argument

The central object is a probabilistic Iterated Function System applied to natural-language queries, where deduction, analogy, and induction act as affine transformations that generate increasingly boundary-challenging probes. A multi-agent system runs these transformations in parallel under a core agent, and a reinforcement-learned policy sets each transformation's probability using rewards tied to semantic-entropy gain. The discovered boundary is materialized as a vector database of hallucinated query-response pairs; the monitor embeds each new query, computes cosine similarity against stored vectors, and flags a hallucination when the top three similar vectors exceed a threshold and the query is close to their centroid, otherwise falling back to a semantic-entropy comparison. This machinery converts an abstract 'generalization bound' into a concrete, queryable database.

What would settle it

In a new domain, collect true queries and their responses, then measure how often a non-hallucinated query is closer to the nearest stored hallucinated boundary point than to the nearest stored non-hallucinated point; if that overlap is large, the vector-database boundary will misclassify, and detection accuracy should fall toward the majority baseline at the paper's threshold of 0.8.

Watch

Extended reading notes

Core claim

HalMit's central claim is that the boundary between reliable and hallucinated behavior of an LLM agent can be captured by a finite collection of hallucinated query-response pairs in a domain, and that this collection acts as an empirical generalization bound. The paper argues that semantic entropy distributions are stable within a domain but differ across domains, so the bound must be learned per agent and per domain rather than set globally. Exploration is driven by three fractal affine transformations over language — deduction, analogy, and induction — applied in parallel by multiple query generators, with a reinforcement-learning policy adjusting which transformation to favor based on changes in semantic entropy. Queries that provoke hallucinations enter a vector database; monitoring then compares an incoming query's embedding against stored boundary vectors, using a centroid-based cosine-similarity test with a semantic-entropy fallback. The paper states this is the first hallucination monitor that operates without internal model access or cross-verification algorithms.

Load-bearing premise

The load-bearing premise is that within a fixed domain, hallucinated and non-hallucinated queries occupy separable regions of embedding space, so a finite set of stored hallucinated examples can stand in for the entire generalization boundary.

Editorial extensions

If this is right

  • API-only LLM agents can be monitored permanently without weight access, because the watchdog relies only on input queries and stored boundary records.
  • Each new agent or domain needs only a fresh boundary-exploration pass, after which monitoring is just retrieval plus a similarity test.
  • Detection is cheap at run time: one embedding lookup and a few cosine computations, with no repeated sampling or confidence calibration.
  • Domain-appropriate probing should transfer the approach to agents built on other model backbones, as the experiments show across several open-weight models.

Reading between the lines

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

  • Editorial inference: if the boundary database is updated online with newly detected hallucinations, the watchdog could track drift in an agent's behavior rather than freezing a snapshot of its bound.
  • Editorial inference: the three fractal transformations bias exploration toward deductive, analogical, and inductive variations; adding temporal, causal, or negation transformations could either widen coverage or add noise, which is a testable extension.
  • Editorial inference: because the monitor relies on embedding proximity, it inherits the embedding model's notion of similarity; a domain where embeddings do not track truthfulness would require a different representation before the same method could work.
  • Editorial inference: the strongest new test would be adversarial — queries deliberately placed far from all stored boundary points in embedding space — since those are precisely where a finite database could miss a novel hallucination type.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper presents HalMit, a black-box hallucination watchdog for LLM-empowered agents. It proposes a multi-agent exploration system with probabilistic fractal-based query generation to collect query-response pairs that are judged hallucinated by an evaluator, and stores them in a vector database interpreted as the agent's 'generalization bound'. At inference, an input query is flagged if its embedding is close to stored examples or its semantic entropy is high relative to retrieved neighbors. Experiments on MedQuAD and SQuAD subsets with six LLMs claim consistent improvements over PP, ICL, and SelfCheckGPT baselines.

Significance. The problem of black-box hallucination monitoring is practically important, and the paper makes a genuine attempt at a domain-specific, training-free watchdog. The exploration-as-fractal metaphor is interesting, and the empirical comparison spans multiple models and domains. However, the central contribution is not a generalization bound in any formal sense: no theoretical quantity is derived or estimated. The method is essentially a nearest-neighbor classifier over stored hallucinated examples, and its success hinges on an unexamined local-constancy assumption in embedding space. The claim of independence from cross-verification is also contradicted by the use of GPT-4 as the labeler. These issues undermine the stated novelty, though the engineering system may still be useful.

major comments (5)
  1. [Section 3.1, 3.2, Algorithm 1] The paper never defines the 'generalization bound' formally; the object stored in the vector database is simply the set of query-response pairs that the evaluator deemed hallucinated (Section 3.1.1, step 3), and detection is thresholded cosine similarity against these points (Algorithm 1, lines 14-27). No statistical, PAC, or coverage property is stated or proven, so the abstract's claim that HalMit 'models the generalization bound' is unsubstantiated. At best this is a retrieval-based detector, and the theory framing is currently decorative.
  2. [Algorithm 1, Section 3.1.1] Algorithm 1 implicitly assumes that hallucination status is locally constant at scale epsilon around each stored example and that queries with lower semantic entropy than the nearest stored example are safe. Section 3.1.1's fractal transformations generate candidate queries but no argument or experiment establishes that these candidates approximate the actual boundary or that the stored set is representative of the agent's failure region. The New York City results in Table 1, where HalMit is outperformed by SelfCheckGPT, illustrate that the method is sensitive to topic coverage; without a coverage analysis, the method cannot detect novel hallucination types.
  3. [Section 1 vs. Section 4.2] Section 1 states that HalMit is 'the first hallucination monitoring approach that operates without access to internal model knowledge or reliance on cross-verification algorithms.' Section 4.2, however, says 'GPT 4 is used to judge whether each response of the target LLM has hallucinations' while constructing the bound. Because those GPT-4 labels define the stored boundary, the system does rely on an external verifier. This is a direct contradiction of a headline contribution.
  4. [Section 3.1.2, Eqs. (2)-(3)] Equation (2) is not a well-defined reward. The notation 'prod_{k=1}^K sig(a_i^tau(k)) tau != 0' is malformed; reading it literally, the condition selects the Delta-H branch when all responses are non-hallucinated and the 1/R_{i-1} branch when at least one response is hallucinated, which is opposite to what one would expect for boundary-seeking exploration. Equation (3) then defines p_j = R_j / sum_k R_k without specifying how R_j is aggregated from the per-query R_i values of transformation j. This makes the RL training procedure unreproducible.
  5. [Tables 1 and 2] Tables 1 and 2 are not readable as printed. For example, in the Treatment block of Table 1, the ICL row lists eight numbers without clear delimiters for the two backbone columns, so the four metrics cannot be unambiguously assigned to Llama2 versus Llama3.1. Table 2 has similar alignment issues, with values run together. Because the paper's central empirical claim is that HalMit 'significantly outperforms existing approaches', this reporting flaw prevents verification of the experimental results.
minor comments (6)
  1. [Section 3.1.1] There is a typo in step 1: 'to cover a broader semantic space with in the bound' should read 'within the bound'.
  2. [Section 1, contributions list] The fourth contribution is duplicated: 'A unique hallucination mitigation technology is provided to enable a more dependable monitoring and monitoring of potential hallucinations' appears twice, with the second instance containing a redundant 'monitoring'.
  3. [Section 3.1.1 vs. Section 4.2] The notation for thresholds is inconsistent: gamma denotes the hallucination ratio in Section 3.1.1, while epsilon is used for the similarity threshold in Algorithm 1; in Section 4.2, gamma is set to 0.6 and epsilon to 0.8, but the text in Section 3.1.1 says 'Once gamma becomes larger than an empirical threshold epsilon', conflating the two.
  4. [Section 4.2] The text says 'the semantic entropy defined in Formula (2) in the Appendix', but Formula (2) is the reward function, not the semantic entropy definition; this should refer to the correct equation in Appendix B.
  5. [Algorithm 1, line 23] The algorithm computes semantic entropy H(Q_v) for an input query, but semantic entropy is defined for responses, not queries; the paper should clarify how query-level entropy is obtained.
  6. [Figure 2] The boxplots in Figure 2 are labeled by 'Group' on the x-axis without explaining what a group is; the caption should define the grouping and the aggregation procedure.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'generalization bound' is the stored set of hallucinated examples, and the monitor is thresholded similarity to that set; the central theoretical claim reduces by construction to labeled retrieval.

  1. self definitional [Section 3.1.1 (step 3), Section 3.2, Algorithm 1]
    "In case a hallucination is reported, the CA embeds the QA pair and the context information into a vector database as a point of the generalization bound of agent τ ... If the query closely resembles the retrieved records in the vector base, it is considered near the boundary, and the response corresponding to the input query is flagged as a potential hallucination."

    The generalization bound is never defined as a formal object; operationally it is exactly the set of QA pairs that are labeled hallucinated and stored. Monitoring then flags a query iff its cosine similarity to those stored points exceeds a threshold (Algorithm 1: results[3] > ε; S_C ≥ ε), or iff its semantic entropy exceeds the maximum entropy among the stored points. Thus 'being beyond the bound' and 'being similar to a stored hallucinated example' are the same predicate by construction. The stored examples are the labels, and the similarity rule is the classifier; the bound contributes no independent constraint. The fractal/RL exploration only selects which examples to store; it does not derive or upper-bound any generalization error.

  2. fitted input called prediction [Section 4.2; Section 1]
    "During modeling the generalization bound, Qwen-max is used to generate queries, while GPT 4 is used to judge whether each response of the target LLM has hallucinations. ... To the best of our knowledge, this is the first hallucination monitoring approach that operates without access to internal model knowledge or reliance on cross-verification algorithms."

    The bound database is constructed from GPT-4's hallucination judgments on generated queries, and the monitor then labels a query by proximity to those same labeled judgments. The first-mover claim of no cross-verification is contradicted by the construction: the database is populated by an external evaluator. The monitoring output is forced by the GPT-4 labels plus a similarity kernel; the only agent-specific information learned is where GPT-4 said hallucinations occur. The evaluation on MedQuAD/SQuAD therefore tests the transfer of a fitted retrieval rule, not an independently estimated generalization bound.

full rationale

HalMit's empirical component is evaluated on MedQuAD and SQuAD against PP, ICL, and SelfCheckGPT, and those benchmark comparisons are independent evidence that the retrieval rule can work; I do not treat the experimental results as fabricated. The circularity is concentrated in the paper's theoretical packaging. The 'generalization bound' is never defined; operationally it is exactly the set of QA pairs that the EA/GPT-4 labels as hallucinated and stores (Section 3.1.1, step 3). The monitor (Section 3.2, Algorithm 1) decides by cosine similarity to those stored points, or by semantic entropy relative to those stored points. Hence the claimed prediction 'this query is beyond the bound' is, by construction, the same as 'this query is close to a stored hallucinated example.' That is a supervised nearest-neighbor rule with a local-constancy assumption, not a derived bound, and the Section 1 claim that the approach avoids cross-verification is contradicted by the GPT-4 labeling used to build the database. The fractal/RL exploration selects which examples to store but does not bound generalization error. No load-bearing self-citation was found. Because real external benchmarks exist, the paper is not wholly vacuous, but the central theoretical claim reduces to labeled retrieval; score 6.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The framework depends on several unproven domain assumptions (hallucination-bound correspondence, semantic entropy validity, embedding-space geometry) and on hand-set thresholds. No theoretically grounded derivation is provided; the 'generalization bound' is a label-derived storage. This counts as the honest measure of what the paper postulates versus what it derives.

free parameters (6)
  • similarity threshold epsilon = 0.8
    Chosen via ablation in Section 4.5 to maximize accuracy; detection rule depends on it directly (Algorithm 1).
  • hallucination ratio threshold gamma = 0.6
    Used to decide when exploration stops; ablation shows insensitivity but the value is still hand-set.
  • normalization parameter omega = not reported
    Appears in state definition in Eq. (4); no value or tuning procedure given.
  • number of repeated responses K = not reported
    Used in the reward computation in Eq. (2); K is never specified.
  • confidence threshold for manual review = 60%
    GPT-4 judgments below 60% confidence are manually reviewed; threshold is arbitrary.
  • initial number of queries per domain = 10
    The core agent randomly initializes ten queries per domain (Section 4.2).
assumptions (5)
  • domain assumption Hallucinated responses correspond to the agent's generalization bound
    Foundational premise stated in the introduction and Section 2, never proven.
  • domain assumption Semantic entropy is a valid proxy for hallucination likelihood
    Used to define rewards and detection; the paper itself notes in PS3 that fixed entropy thresholds are insufficient, weakening this axiom.
  • domain assumption Cosine similarity in embedding space reflects proximity to the generalization boundary
    Monitoring in Algorithm 1 flags queries based on embedding similarity to stored hallucinated queries.
  • ad hoc to paper The three fractal affine transformations (deduction, analogy, induction) cover the relevant semantic space
    No evidence that these three patterns suffice to explore the boundary; they are presented as a design choice.
  • domain assumption GPT-4 judgments of hallucination are accurate for building the boundary database
    The boundary database is constructed from GPT-4 evaluations with partial manual review; label noise propagates into detection.
invented entities (1)
  • Generalization bound in semantic space
    purpose: A theoretical object that HalMit claims to model and use for hallucination detection
    The 'bound' is never defined formally or derived; it is operationalized as a set of hallucinated QA pairs in a vector database. No independent evidence of its existence or shape is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Mitigation of Hallucination for LLM-empowered Agents: Progressive Generalization Bound Exploration and Watchdog Monitor." pith.science (2026). https://pith.science/paper/PBHKUQPV

@misc{pith2026250715903,
  author       = {Pith},
  title        = {Pith review of: Towards Mitigation of Hallucination for LLM-empowered Agents: Progressive Generalization Bound Exploration and Watchdog Monitor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PBHKUQPV}},
  note         = {Machine review of arXiv:2507.15903}
}
read the original abstract

Empowered by large language models (LLMs), intelligent agents have become a popular paradigm for interacting with open environments to facilitate AI deployment. However, hallucinations generated by LLMs-where outputs are inconsistent with facts-pose a significant challenge, undermining the credibility of intelligent agents. Only if hallucinations can be mitigated, the intelligent agents can be used in real-world without any catastrophic risk. Therefore, effective detection and mitigation of hallucinations are crucial to ensure the dependability of agents. Unfortunately, the related approaches either depend on white-box access to LLMs or fail to accurately identify hallucinations. To address the challenge posed by hallucinations of intelligent agents, we present HalMit, a novel black-box watchdog framework that models the generalization bound of LLM-empowered agents and thus detect hallucinations without requiring internal knowledge of the LLM's architecture. Specifically, a probabilistic fractal sampling technique is proposed to generate a sufficient number of queries to trigger the incredible responses in parallel, efficiently identifying the generalization bound of the target agent. Experimental evaluations demonstrate that HalMit significantly outperforms existing approaches in hallucination monitoring. Its black-box nature and superior performance make HalMit a promising solution for enhancing the dependability of LLM-powered systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    Augmenting LLMs with Knowledge: A survey on hallucination prevention

    K. Andriopoulos and J. Pouwelse. Augmenting llms with knowledge: A survey on hallucination prevention.arXiv preprint arXiv:2309.16459, 2023

    work page Pith review arXiv 2023

  2. [2]

    Ben Abacha and D

    A. Ben Abacha and D. Demner-Fushman. A question-entailment ap- proach to question answering.BMC Bioinf, 20, 2019

    work page 2019

  3. [3]

    Casper, X

    S. Casper, X. Davies, C. Shi, T. K. Gilbert, J. Scheurer, J. Rando, R. Freedman, T. Korbak, D. Lindner, P. Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feed- back.arXiv preprint arXiv:2307.15217, 2023

    arXiv 2023

  4. [4]

    Y . Chen, Q. Fu, Y . Yuan, Z. Wen, G. Fan, D. Liu, D. Zhang, Z. Li, and Y . Xiao. Hallucination detection: Robustly discerning reliable answers in large language models. InCIKM, 2023

    work page 2023

  5. [5]

    Elasticsearch

    Elastic. Elasticsearch. https://www.elastic.co/guide/en/elasticsearch/ reference/current/elasticsearch-intro-what-is-es.html, 2023. Accessed: 2023-10-10

    work page 2023

  6. [6]

    Farquhar, J

    S. Farquhar, J. Kossen, L. Kuhn, and Y . Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017), 2024

    work page 2024

  7. [7]

    Freyer, I

    O. Freyer, I. C. Wiest, J. N. Kather, and S. Gilbert. A future role for health applications of large language models depends on regulators en- forcing safety standards.Lancet Digit Health, 6(9), 2024

    work page 2024

  8. [8]

    D. Gao, Z. Li, X. Pan, W. Kuang, Z. Ma, B. Qian, F. Wei, W. Zhang, Y . Xie, D. Chen, L. Yao, H. Peng, Z. Y . Zhang, L. Zhu, C. Cheng, H. Shi, Y . Li, B. Ding, and J. Zhou. Agentscope: A flexible yet robust multi- agent platform.CoRR, abs/2402.14034, 2024

    arXiv 2024

Show all 38 references
  1. [9]

    C. M. Greco and A. Tagarelli. Bringing order into the realm of transformer-based language models for artificial intelligence and law. Artif Intell Law, 2023

    2023

  2. [10]

    J. Han, J. Kossen, M. Razzak, L. Schut, S. A. Malik, and Y . Gal. Seman- tic entropy probes: Robust and cheap hallucination detection in llms. In ICML, 2024

    2024

  3. [11]

    J. He, Y . Gong, Z. Lin, C. Wei, Y . Zhao, and K. Chen. Llm factoscope: Uncovering llms’ factual discernment through measuring inner states. InACL Findings, 2024

    2024

  4. [12]

    B. Hou, Y . Zhang, J. Andreas, and S. Chang. A probabilistic frame- work for llm hallucination detection via belief tree propagation.arXiv preprint arXiv:2406.06950, 2024

    2024 arXiv

  5. [13]

    Huang, W

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, et al. A survey on hallucination in large lan- guage models: Principles, taxonomy, challenges, and open questions. ACM Trans Inf Syst, 2023

    2023

  6. [14]

    Z. Ji, D. Chen, E. Ishii, S. Cahyawijaya, Y . Bang, B. Wilie, and P. Fung. Llm internal states reveal hallucination risk faced with a query.arXiv preprint arXiv:2407.03282, 2024

    2024 arXiv

  7. [15]

    Kasneci, K

    E. Kasneci, K. Seßler, S. Küchemann, M. Bannert, D. Dementieva, F. Fischer, U. Gasser, G. Groh, S. Günnemann, E. Hüllermeier, et al. Chatgpt for good? on opportunities and challenges of large language models for education.Learn Individ Differ, 103, 2023

    2023

  8. [16]

    L. Kuhn, Y . Gal, and S. Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. arXiv preprint arXiv:2302.09664, 2023

    2023 arXiv

  9. [17]

    S. Lin, J. Hilton, and O. Evans. Truthfulqa: Measuring how models mimic human falsehoods. InACL, 2022

    2022

  10. [18]

    Lotfi, M

    S. Lotfi, M. Finzi, Y . Kuang, T. G. Rudner, M. Goldblum, and A. G. Wilson. Non-vacuous generalization bounds for large language models. arXiv preprint arXiv:2312.17173, 2023

    2023 arXiv

  11. [19]

    Lotfi, Y

    S. Lotfi, Y . Kuang, B. Amos, M. Goldblum, M. Finzi, and A. G. Wil- son. Unlocking tokens as data points for generalization bounds on larger language models.arXiv preprint arXiv:2407.18158, 2024

    2024 arXiv

  12. [20]

    Manakul, A

    P. Manakul, A. Liusie, and M. Gales. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. InEMNLP, 2023

    2023

  13. [21]

    Quevedo, J

    E. Quevedo, J. Yero, R. Koerner, P. Rivas, and T. Cerny. Detecting hallucinations in large language model generation: A token probability approach.arXiv preprint arXiv:2405.19648, 2024

    2024 arXiv

  14. [22]

    Rajpurkar, J

    P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang. SQuAD: 100,000+ questions for machine comprehension of text. InEMNLP, 2016

    2016

  15. [23]

    N. Riemer. Remetonymizing metaphor: Hypercategories in semantic extension. 2002

    2002

  16. [24]

    Sriramanan, S

    G. Sriramanan, S. Bharti, V . S. Sadasivan, S. Saha, P. Kattakinda, and S. Feizi. Llm-check: Investigating detection of hallucinations in large language models.Adv Neural Inf Process Syst, 37:34188–34216, 2024

    2024

  17. [25]

    D. Su, X. Li, J. Zhang, L. Shang, X. Jiang, Q. Liu, and P. Fung. Read before generate! faithful long form question answering with machine reading. InACL Foundings, 2022

    2022

  18. [26]

    Udekwe, O

    D. Udekwe, O. ofe Ajayi, O. Ubadike, K. Ter, and E. Okafor. Com- paring actor-critic deep reinforcement learning controllers for enhanced performance on a ball-and-plate system.Expert Syst Appl, 245, 2024. ISSN 0957-4174

    2024

  19. [27]

    Y . Wang, Q. Sun, and S. He. M3E: Moka Massive Mixed Embedding Model, 2023

    2023

  20. [28]

    J. Wei, Y . Yao, J.-F. Ton, H. Guo, A. Estornell, and Y . Liu. Measur- ing and reducing llm hallucination without gold-standard answers via expertise-weighting.arXiv preprint arXiv:2402.10412, 2024

    2024 arXiv

  21. [29]

    Xu and C

    S. Xu and C. Zhang. Misconfidence-based demonstration selection for llm in-context learning.arXiv preprint arXiv:2401.06301, 2024

    2024 arXiv

  22. [30]

    H. Yang, H. Lu, W. Lam, and D. Cai. Exploring compositional general- ization of large language models. InNAACL-HLT, pages 16–24, 2024

    2024

  23. [31]

    J. Yang, H. Jin, R. Tang, X. Han, Q. Feng, H. Jiang, S. Zhong, B. Yin, and X. Hu. Harnessing the power of llms in practice: A survey on chatgpt and beyond.ACM Trans Knowl Discov Data, 18(6), 2024

    2024

  24. [32]

    C. Zhang. User-controlled knowledge fusion in large language models: Balancing creativity and hallucination.arXiv preprint arXiv:2307.16139, 2023

    2023 arXiv

  25. [33]

    Zhang, S

    C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals. Understand- ing deep learning (still) requires rethinking generalization.Communi- cations of the ACM, 64(3):107–115, 2021

    2021

  26. [34]

    Zhang, S

    Y . Zhang, S. Li, J. Liu, P. Yu, Y . R. Fung, J. Li, M. Li, and H. Ji. Knowledge overshadowing causes amalgamated hallucination in large language models.arXiv preprint arXiv:2407.08039, 2024

    2024 arXiv

  27. [35]

    H. Zhao, Z. Liu, Z. Wu, Y . Li, T. Yang, P. Shu, S. Xu, H. Dai, L. Zhao, G. Mai, et al. Revolutionizing finance with llms: An overview of appli- cations and insights.arXiv preprint arXiv:2401.11641, 2024

    2024

  28. [36]

    Z. Zhao, B. Wang, L. Ouyang, X. Dong, J. Wang, and C. He. Be- yond hallucinations: Enhancing lvlms through hallucination-aware di- rect preference optimization.arXiv preprint arXiv:2311.16839, 2023

    2023 arXiv

  29. [37]

    X. Zhou, M. Zhang, Z. Lee, W. Ye, and S. Zhang. Hademif: Hallucina- tion detection and mitigation in large language models. InICLR

  30. [38]

    D. Zhu, D. Chen, Q. Li, Z. Chen, L. Ma, J. Grossklags, and M. Fritz. Pollmgraph: Unraveling hallucinations in large language models via state transition dynamics.arXiv preprint arXiv:2404.04722, 2024

    2024 arXiv

Pith tools

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