Pith. sign in

REVIEW 3 major objections 6 minor 50 references

MALM: A Multi-Information Adapter for Large Language Models to Mitigate Hallucination

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MALM inserts a multi-edge graph attention layer between a transformer's last block and its output head, and claims this jointly reduces Input-, Context-, and Fact-Conflicting hallucination across seven base LLMs.

desk verdict MALM's controlled RAG results and ablations are worthwhile, but its headline seven-LLM adaptability claim is undermined by a knowledge-access confound. read the letter →

arxiv 2506.12483 v1 pith:T4WEBENT submitted 2025-06-14 cs.AI cs.CL

Ao Jia , Haiming Wu , Guohui Yao , Dawei Song , Songkun Ji , Yazhou Zhang This is my paper
classification cs.AIcs.CL
keywords hallucinationmitigationlargelanguagemodelgraphneuralnetworksattentionnetworkquestionansweringretrieval-augmentedgenerationinput-conflictingfact-conflicting
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 show that the three named kinds of LLM hallucination, Input-, Context-, and Fact-Conflicting, are interdependent and can be mitigated together by letting the model attend simultaneously to the user's query, the text it has already produced, and an external knowledge passage. To do this it proposes MALM, a plug-in multi-layer graph attention adapter placed between the last transformer block and the output head, with directed edges that carry query information into the partial output, link output tokens to each other under a causal mask, and pull in knowledge tokens. On HaluEval, TruthfulQA, Natural Questions, and TriviaQA, the authors report consistent improvements for all seven tested base models, including a 132.01% relative ROUGE-2 gain for LLaMA-2 on TruthfulQA, and state-of-the-art RAG results when the knowledge is supplied by BM25, Spider, or DPR. The central claim is that the graph's edge topology, not just the presence of extra knowledge, is what lowers hallucination; the paper's type-specific counts show the non-hallucinated answer rate rising from 44% to 64%.

What carries the argument

The central object is a multi-layered graph attention network whose vertices are tokens and whose edges encode three information channels: input connections (unidirectional from query tokens to partial-output tokens), context connections (masked fully connected edges among output tokens, so future tokens cannot leak backwards), and knowledge connections (unidirectional from fact tokens to every output token). Within each layer, input tokens and knowledge tokens are also fully connected among themselves. The update uses multi-head GAT-style attention scores $\alpha^{h,l}_{pq}$ to aggregate neighbors, and the adapter is trained with cross-entropy while the base LLM is fine-tuned with LoRA; the final token distribution is a weighted mixture of the graph-based prediction and the LLM's own prediction. The graph's job is to propagate factual and query information onto the next-token representation during decoding, so that generation is simultaneously anchored to user intent, internal consistency, and supplied knowledge.

What would settle it

Run MALM and a no-graph baseline on the same test split with identical retrieved knowledge concatenated into the prompt, holding the base model and LoRA setup fixed; if the baseline matches or exceeds MALM's ROUGE and FEQA scores, the three-edge graph topology is not what drives the reported gains. Sharper still, randomly permute the input, context, and knowledge edge sets in MALM's adjacency matrix; if hallucination rates do not change, the specific wiring carries no information.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that hallucination types co-move: improving one supports the other two, and a mechanism that models the joint state of query, partial output, and external knowledge can reduce all three at once. MALM realizes this as a directed graph whose vertices are tokens from the input $T$, the partial output $X_{<i}$, and the knowledge $K$; input edges run from $T$ to $X_{<i}$, context edges form a masked fully connected graph among output tokens with $A_{p,q}=1$ for $p \geq q$, and knowledge edges run from $K$ to every output node. A graph attention update aggregates the non-hallucinated information from these neighbors, and the final prediction is a weighted residual $\hat{y}_i = \operatorname{Softmax}(\lambda y^g_i + (1-\lambda) y^o_i)$ combining the graph readout with the original LLM prediction. The paper reports that this adapter improves every tested base model, cuts Fact-Conflicting hallucination by 23.8%, Context-Conflicting by 66.7%, and Input-Conflicting by 72.7% in its type-specific analysis, and that GPT-4 and human judges prefer MALM's answers in 79.4% and 65.6% of pairings.

Load-bearing premise

The central claim stands on the assumption that the supplied external knowledge contains the correct answer and that the graph's knowledge edges let the model attend to it; the authors themselves accept in §4.8 that when the knowledge lacks the answer, Fact-Conflicting hallucination persists.

Editorial extensions

If this is right

  • If the graph topology is what does the work, then any open-source LLM that exposes hidden states can be retrofitted with MALM without retraining the base model; the paper demonstrates this on GPT-2, BLOOM, Qwen-2, Qwen-2.5, Falcon, Vicuna, and LLaMA-2.
  • MALM can be layered on top of existing retrieval systems: using the same retrieved documents from BM25, Spider, or DPR, the adapter raises ROUGE-1 by 17.57% to 37.49% over the strongest RAG baselines in the paper on HaluEval, NQ, and TriviaQA.
  • Because the paper's type-specific counts show Input-Conflicting hallucination falling by 72.7% and Context-Conflicting by 66.7%, the adapter's largest benefits are alignment with the query and internal consistency, not just factual recall.
  • The two-layer setting is the paper's reported optimum; deeper graphs lose performance, which the authors attribute to over-smoothing, implying that stack depth is a practical hyperparameter to tune per task rather than something to maximize.
  • If the knowledge passage does not contain the answer, Fact-Conflicting hallucination persists, so MALM is a knowledge-utilization device, not a knowledge source; the paper states this limitation explicitly in §4.8.

Reading between the lines

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

  • A testable implication the authors do not run: feed the identical retrieved knowledge to a no-graph baseline (for example, concatenated into the prompt) and to MALM; if the gap vanishes, the improvement is attributable to knowledge access rather than to the three-edge topology.
  • The masked context fully-connected graph is a form of intra-response attention, so comparing MALM against an adapter that reuses the base model's own causal attention over the same three inputs would isolate whether a separate graph module is needed.
  • The authors' own numerical-failure cases suggest that the mechanism is feature mixing rather than arithmetic or multi-hop reasoning; a natural extension is to pair MALM with a calculator or symbolic module, or to test it on datasets that require exact quantities.
  • Since the graph needs hidden states, MALM is confined to open-weight models; a closed-API analogue would need to route information through token-level prompts, which would test whether the graph's benefit survives without exposed hidden features.
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

3 major / 6 minor

Summary. The paper proposes MALM, a Multi-Information Adapter that inserts a multi-layer graph attention network between the last transformer block and the output head of a foundation LLM. The graph has three subgraphs corresponding to input tokens, partial output (context) tokens, and external knowledge tokens, with directed edges from input to output, masked fully connected edges among output tokens, and knowledge-to-output edges. The adapter is trained with cross-entropy and combined with the original prediction via a weighted residual. The authors evaluate adaptability on HaluEval and TruthfulQA with seven LLMs, generalizability to RAG on HaluEval, NQ, and TriviaQA with three retrievers, plus ablations, layer analysis, GPT-4 and human preference evaluations, and a type-specific hallucination analysis. They report significant improvements of MALM over baselines, particularly for LLaMA-2 (+132.01% ROUGE-2 on TruthfulQA) and over strong RAG systems (up to +37.49% ROUGE-1 on NQ with DPR).

Significance. The paper proposes a novel adapter that uses a multi-layer graph attention network to connect input, context, and external knowledge at the token level, inserted into the frozen LLM at the final transformer layer, with a weighted residual combination. The experimental scope is broad: seven base LLMs, two hallucination-focused datasets, three retrievers on three QA datasets, plus ablation, layer analysis, GPT-4 and human evaluation. If the central causal claim is established, the method would be a flexible and parameter-efficient way to reduce hallucination by explicitly modeling interactions among the three hallucination types, and it would complement existing RAG systems. The RAG experiments (Table 4) are well controlled (all models see the same retrieved documents) and show consistent gains across datasets and retrievers; the ablation study (Table 7) isolates the three edge types; and the layer analysis (Table 8) shows a plausible over-smoothing trend. However, the headline adaptability and type-specific mitigation claims currently rest on uncontrolled comparisons in which MALM receives external knowledge that the baselines do not.

major comments (3)
  1. [Sec. 4.2, Tables 2 and 3] The comparison between MALM and its w/o MALM baselines does not control for access to external knowledge K. Section 4.1.1 states that HaluEval "provides relevant external knowledge information for each input query" and the authors "directly use the knowledge information included in the dataset," but the description of the fine-tuning protocol (Sec. 4.2.1) mentions only questions as inputs for the baselines. Since MALM encodes K as a separate parallel stream (Sec. 3.2) and the gold K often contains the answer text, the reported gains (e.g., LLaMA-2 on TruthfulQA: +132.01% ROUGE-2) could be due to supplying the answer to MALM only, not to the graph topology. Please add a control where the w/o MALM baselines receive the same K (e.g., as a prefix in the prompt, or via retrieval) and re-report the adaptability and type-specific results. Without this control, the headline claims of adaptability across seven LLMs and of mitigating all three hallucination types are not supported.
  2. [Sec. 4.8, Table 11] The type-specific mitigation analysis compares LLaMA-2 (which apparently did not see K) against MALM (which did), and the counts are small (n=100 total; e.g., Context-Conflicting counts are 3 vs. 1). The claimed reductions (72.7% input, 66.7% context, 23.8% fact) therefore reflect both the presence of K and the graph structure, and are not statistically quantified. This table is load-bearing for the paper's central claim that MALM mitigates all three types, so it should be re-run with the knowledge-access control and with an appropriate significance test.
  3. [Secs. 4.6 and 4.7] The GPT-4 and human preference evaluations ask judges to choose between responses from MALM (which had access to K during generation) and from LLaMA-2 (which apparently did not). The evaluation prompt in Table 9 explicitly provides "relevant knowledge" to the judge, so the judges can simply prefer answers that use that knowledge, regardless of the graph mechanism. Please either generate the LLaMA-2 baseline with the same K in its prompt, or conduct a blind evaluation where both options are generated under identical knowledge access, before interpreting the 79.4% and 65.6% preference rates as evidence for the graph topology.
minor comments (6)
  1. [Abstract and Sec. 3.3] There are typos in the abstract ("to mitigate the them") and in Sec. 3.3 ("fact-confliciting hallucination").
  2. [Sec. 4.2.2] The statement "All the improvements over the best baseline are statistically significant (two-tailed t-test, p-value=0.004<0.01)" does not specify which comparison, how many seeds, or whether the test is paired; please provide per-metric and per-baseline details.
  3. [Sec. 4.8] The random selection of 25 samples per dataset yields very small per-type counts; please report confidence intervals or an exact test.
  4. [Sec. 4.7] The human evaluation uses only three volunteers and 30 items; this is acceptable as a pilot but should be described as such, and the inter-rater reliability would be stronger with more participants.
  5. [General] The paper does not mention whether code or trained models will be released; for reproducibility, please add an availability statement.
  6. [Sec. 4.3.1] In the description of RECOMP, the authors say they replace Flan-UL2-20B with LLaMA-2-7B for fairness; this is a reasonable choice but should be reported as a potential limitation since RECOMP was not designed for this backbone.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: MALM's gains are empirical outcomes on external benchmarks, not predictions forced by construction or by load-bearing self-citation.

full rationale

MALM is an empirical adapter method trained with cross-entropy on HaluEval QA data and evaluated on external benchmarks; no prediction is derived from fitted parameters. The three-type taxonomy is attributed to Zhang et al. (2023b), a self-cited survey, but it is used as a premise rather than as support for the claimed improvements, and the graph adapter is a new construction inspired by but not reduced to M3GAT (Zhang et al., 2023a). The causal claim that the edge topology mitigates hallucination is supported by ablations (removing input, knowledge, or context connections) and by RAG experiments in which all compared models receive the same retrieved documents (Table 4), so the central result does not reduce by construction to a self-citation or to a fitted input. The uncontrolled comparison in Tables 2-3, where MALM receives external knowledge K while the 'w/o MALM' baselines are not described as receiving K, is a potential confound about what the comparison measures; however, this is an experimental-control concern, not a circularity in the derivation, and the paper's own limitation section concedes that when K lacks the answer the model still hallucinates. Under the stated rules, no step qualifies as circular, so the paper is self-contained against external benchmarks and the score is 0.

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

The method introduces no new entities. Its performance depends on several hand-set hyperparameters and on two modeling assumptions: that the knowledge is relevant, and that the three edge types correspond to the three hallucination types. The most consequential hidden assumption is the baseline experiment's treatment of knowledge access, which is a reporting gap rather than an axiom.

free parameters (4)
  • residual weight lambda = 0.2
    Chosen by hand in §4.1.3; balances graph prediction and original LLM prediction.
  • number of graph layers L = 2 (but L=1 better with BM25 in Table 8)
    Set to 2 in §4.1.3 and used in main experiments, while the layer sweep in §4.5 shows L=1 yields higher ROUGE-1/ROUGE-L/FEQA for MALM with BM25 on HaluEval.
  • number of attention heads H = 8
    Hyperparameter in §4.1.3.
  • GAT dropout = 0.1
    Hyperparameter in §4.1.3.
assumptions (4)
  • domain assumption The three-type hallucination taxonomy (Input, Context, Fact) is a valid decomposition and the three edge types in MALM correspond one-to-one to these types.
    Invoked in §1.1 and §3.3; the paper asserts the mapping without experimental isolation of each edge-hallucination pair beyond a small 100-sample analysis in §4.8.
  • domain assumption The external knowledge K is relevant and contains the correct answer for the query.
    The knowledge connections in §3.3 rely on K; §4.8 explicitly states that when K lacks the correct answer, Fact-Conflicting hallucination persists.
  • domain assumption Hidden states from the last transformer block can serve as node features, and a separate parallel LLM pass over K yields uncontaminated knowledge features.
    Stated in §3.2; the dual-stream design is not ablated, so its necessity is assumed.
  • standard math Standard GAT update equations (Veličković et al., 2018) are applicable to the directed, masked adjacency matrix defined in Eqs. 2-4.
    Eqs. 3-6 are taken from the GAT literature without proof; the directed and masked edges are a modification that is assumed to preserve convergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MALM: A Multi-Information Adapter for Large Language Models to Mitigate Hallucination." pith.science (2026). https://pith.science/paper/T4WEBENT

@misc{pith2026250612483,
  author       = {Pith},
  title        = {Pith review of: MALM: A Multi-Information Adapter for Large Language Models to Mitigate Hallucination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4WEBENT}},
  note         = {Machine review of arXiv:2506.12483}
}
read the original abstract

Large language models (LLMs) are prone to three types of hallucination: Input-Conflicting, Context-Conflicting and Fact-Conflicting hallucinations. The purpose of this study is to mitigate the different types of hallucination by exploiting the interdependence between them. For this purpose, we propose a Multi-Information Adapter for Large Language Models (MALM). This framework employs a tailored multi-graph learning approach designed to elucidate the interconnections between original inputs, contextual information, and external factual knowledge, thereby alleviating the three categories of hallucination within a cohesive framework. Experiments were carried out on four benchmarking datasets: HaluEval, TruthfulQA, Natural Questions, and TriviaQA. We evaluated the proposed framework in two aspects: (1) adaptability to different base LLMs on HaluEval and TruthfulQA, to confirm if MALM is effective when applied on 7 typical LLMs. MALM showed significant improvements over LLaMA-2; (2) generalizability to retrieval-augmented generation (RAG) by combining MALM with three representative retrievers (BM25, Spider and DPR) separately. Furthermore, automated and human evaluations were conducted to substantiate the correctness of experimental results, where GPT-4 and 3 human volunteers judged which response was better between LLaMA-2 and MALM. The results showed that both GPT-4 and human preferred MALM in 79.4% and 65.6% of cases respectively. The results validate that incorporating the complex interactions between the three types of hallucination through a multilayered graph attention network into the LLM generation process is effective to mitigate the them. The adapter design of the proposed approach is also proven flexible and robust across different base LLMs.

Figures

Figures reproduced from arXiv: 2506.12483 by the authors.

Figure 1
Figure 1. Examples of three types of hallucination [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the proposed approach. Ao Jia et al.: Preprint submitted to Elsevier Page 25 of 24 [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. An example to better explain the mechanism of MALM. The red, purple, and blue tokens represent input nodes, partial output nodes, and knowledge nodes in the graph, respectively. And the red, purple, and blue arrows represent input connection, partial output connection, and knowledge connection, respectively. Please note that not all edges are drawn for the sake of simplicity. Ao Jia et al.: Preprint submitted to Els… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The performance of foundation models w/o and w/ MALM. Ao Jia et al.: Preprint submitted to Elsevier Page 27 of 24 [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]
Figure 5
Figure 5. Figure 5: Performance curves with respect to different layers [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: The result of automated evaluation. Ao Jia et al.: Preprint submitted to Elsevier Page 28 of 24 [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]
Figure 7
Figure 7. Figure 7: The confusion matrix of human evaluation. Ao Jia et al.: Preprint submitted to Elsevier Page 29 of 24 [PITH_FULL_IMAGE:figures/full_fig_p029_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 36 canonical work pages

  1. [1]

    The falcon series of open language models.arXiv preprint arXiv:2311.16867

    Almazrouei,E.,Alobeidli,H.,Alshamsi,A.,Cappelli,A.,Cojocaru,R.,Debbah,M.,Goffinet,É.,Hesslow,D.,Launay,J.,Malartic,Q.,etal.(2023). The falcon series of open language models.arXiv preprint arXiv:2311.16867. Asai,A.,Wu,Z.,Wang,Y.,Sil,A.,andHajishirzi,H.(2024). Self-RAG:Learningtoretrieve,generate,andcritiquethroughself-reflection. InThe Twelfth Internationa...

    arXiv 2023

  2. [2]

    Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024). A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology, 15(3):1–45

    work page 2024

  3. [3]

    Chen, A., Pasupat, P., Singh, S., Lee, H., and Guu, K. (2023). Purr: Efficiently editing language model hallucinations by denoising language model corruptions.arXiv preprint arXiv:2305.14908

    arXiv 2023

  4. [4]

    Chen, J. (2024). An entity-guided text summarization framework with relational heterogeneous graph neural network.Neural Computing and Applications, 36(7):3613–3630

    work page 2024

  5. [5]

    Chen, J., Lin, H., Han, X., and Sun, L. (2024a). Benchmarking large language models in retrieval-augmented generation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17754–17762. Chen,X.,Wang,C.,Zhang,N.,Xue,Y.,xiaoyanyang,Shen,Y.,Gu,J.,andChen,H.(2024b). Unifiedhallucinationdetectionformultimodallarge language models. I...

    work page 2024

  6. [6]

    Cheng, X., Luo, D., Chen, X., Liu, L., Zhao, D., and Yan, R. (2023). Lift yourself up: Retrieval-augmented text generation with self-memory. In Advances in Neural Information Processing Systems, volume 36, pages 43780–43799

    work page 2023

  7. [7]

    Dhuliawala, S., Komeili, M., Xu, J., Raileanu, R., Li, X., Celikyilmaz, A., and Weston, J. E. (2024). Chain-of-verification reduces hallucination in large language models. InICLR 2024 Workshop on Reliable and Responsible Foundation Models, pages 1–19

    work page 2024

  8. [8]

    Durmus, E., He, H., and Diab, M. (2020). FEQA: A question answering evaluation framework for faithfulness assessment in abstractive summarization. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5055–5070

    work page 2020

Show all 50 references
  1. [9]

    Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., and Larson, J. (2024). From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130

    2024 arXiv

  2. [10]

    T., Fan, Y., Zhao, V., Lao, N., Lee, H., Juan, D., and Guu, K

    Gao, L., Dai, Z., Pasupat, P., Chen, A., Chaganty, A. T., Fan, Y., Zhao, V., Lao, N., Lee, H., Juan, D., and Guu, K. (2023). RARR: Researching and revising what language models say, using language models. InProceedings of the 61st Annual Meeting of the Association for Computat...

    2023

  3. [11]

    Gori, M., Monfardini, G., and Scarselli, F. (2005). A new model for learning in graph domains. InProceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729–734

    2005

  4. [12]

    Hamilton, W., Ying, Z., and Leskovec, J. (2017). Inductive representation learning on large graphs. InAdvances in Neural Information Processing Systems, volume 30, pages 1–11

    2017

  5. [13]

    Hao, Q., Wang, C., Xiao, Y., and Lin, H. (2024). Simplices-based higher-order enhancement graph neural network for multi-behavior recommendation.Information Processing & Management, 61(5):103790

    2024

  6. [14]

    He, R., Liu, L., Ye, H., Tan, Q., Ding, B., Cheng, L., Low, J., Bing, L., and Si, L. (2021). On the effectiveness of adapter-based tuning for pretrainedlanguagemodeladaptation. InProceedingsofthe59thAnnualMeetingoftheAssociationforComputationalLinguisticsandthe11th Internation...

    2021

  7. [15]

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and Gelly, S. (2019). Parameter-efficient transfer learning for NLP. InProceedings of the 36th International Conference on Machine Learning, volume 97, pages 2790–2799

    2019

  8. [16]

    J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., yelong shen, Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. (2022). LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, pages 1–13. Ao Jia et al.:Preprint submitted to ElsevierPage 21 of 24...

    2022

  9. [17]

    Huang, X., Han, K., Yang, Y., Bao, D., Tao, Q., Chai, Z., and Zhu, Q. (2024b). Can gnn be good adapter for llms? InProceedings of the ACM on Web Conference 2024, pages 893–904

    2024

  10. [18]

    and Grave, E

    Izacard, G. and Grave, E. (2021). Leveraging passage retrieval with generative models for open domain question answering. InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 874–880. Ji,Z.,Lee,N.,Fri...

    2021

  11. [19]

    Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., and Zhang, S. (2024). Hallucination augmented contrastive learningformultimodallargelanguagemodel. InProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition,pages 27036–27046....

    2024

  12. [20]

    Kipf, T. N. and Welling, M. (2017). Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations, pages 1–14. Kwiatkowski,T.,Palomaki,J.,Redfield,O.,Collins,M.,Parikh,A.,Alberti,C.,Epstein,D.,Polosukhin,I.,Devlin,J.,L...

    2017

  13. [21]

    N., Shoeybi, M., and Catanzaro, B

    Lee, N., Ping, W., Xu, P., Patwary, M., Fung, P. N., Shoeybi, M., and Catanzaro, B. (2022). Factuality enhanced language models for open-ended text generation.Advances in Neural Information Processing Systems, 35:34586–34599

    2022

  14. [22]

    Leiser, F., Eckhardt, S., Leuthe, V., Knaeble, M., Maedche, A., Schwabe, G., and Sunyaev, A. (2024). Hill: A hallucination identifier for large language models. InProceedings of the CHI Conference on Human Factors in Computing Systems, pages 1–13

    2024

  15. [23]

    Li, D., Gao, Y., Wang, Z., Qiu, H., Liu, P., Xiong, Z., and Zhang, Z. (2024). Homogeneous graph neural networks for third-party library recommendation.Information Processing & Management, 61(6):103831

    2024

  16. [24]

    X., Nie, J., and Wen, J

    Li, J., Cheng, X., Zhao, W. X., Nie, J., and Wen, J. (2023). Halueval: A large-scale hallucination evaluation benchmark for large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449–6464

    2023

  17. [25]

    Lin, C. (2004). Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81

    2004

  18. [26]

    Lin, S., Hilton, J., and Evans, O. (2022). Truthfulqa: Measuring how models mimic human falsehoods. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214–3252

    2022

  19. [27]

    Liu, S., Cheng, H., Wang, Y., He, Y., Fan, C., and Liu, Z. (2025). Evopath: Evolutionary meta-path discovery with large language models for complex heterogeneous information networks.Information Processing & Management, 62(1):103920

    2025

  20. [28]

    Maazallahi, A., Asadpour, M., and Bazmi, P. (2025). Advancing emotion recognition in social media: A novel integration of heterogeneous neural networks with fine-tuned language models.Information Processing & Management, 62(2):103974. Manakul,P.,Liusie,A.,andGales,M.(2023).Sel...

    2025

  21. [29]

    McDonald, D., Papadopoulos, R., and Benningfield, L. (2024). Reducing llm hallucination using knowledge distillation: A case study with mistral large and mmlu benchmark.TechRxiv

    2024

  22. [30]

    Meng, Y., Zong, S., Li, X., Sun, X., Zhang, T., Wu, F., and Li, J. (2022). GNN-LM: Language modeling based on global contexts via GNN. In ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing, pages 1–13. Mündler, N., He, J., Jenko, S., and Vechev, M. (...

    2022 arXiv

  23. [31]

    Papineni, K., Roukos, S., Ward, T., and Zhu, W. (2002). BLEU: a Method for Automatic Evaluation of Machine Translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318. Pfeiffer,J.,Piccinno,F.,Nicosia,M.,Wang,X.,Reid,M.,...

    2002

  24. [32]

    Rajpurkar, P., Zhang, J., Lopyrev, K., and Liang, P. (2016). SQuAD: 100,000+ Questions for Machine Comprehension of Text.arXiv preprint arXiv:1606.05250

    2016 arXiv

  25. [33]

    Ram, O., Levine, Y., Dalmedigos, I., Muhlgay, D., Shashua, A., Leyton-Brown, K., and Shoham, Y. (2023). In-Context Retrieval-Augmented Language Models.Transactions of the Association for Computational Linguistics, 11:1316–1331. Ram,O.,Shachaf,G.,Levy,O.,Berant,J.,andGloberson,...

    2023

  26. [34]

    and Zaragoza, H

    Robertson, S. and Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond.Information Retrieval, 3(4):333–389. Sarthi,P.,Abdullah,S.,Tuli,A.,Khanna,S.,Goldie,A.,andManning,C.D.(2024).RAPTOR:RecursiveAbstractiveProcessingforTree-Organized Retrieval. InThe Tw...

    2009

  27. [35]

    Belkada, Y., and Wolf, T. (2022). BLOOM: A 176B-Parameter Open-Access Multilingual Language Model.arXiv preprint arXiv:2211.05100. Shi,W.,Min,S.,Yasunaga, M.,Seo,M.,James,R.,Lewis,M.,Zettlemoyer, L.,andYih,W.-t.(2024). REPLUG:Retrieval-AugmentedBlack-Box Language Models. InPro...

    2022 arXiv

  28. [36]

    Sun, W., Shi, Z., Gao, S., Ren, P., de Rijke, M., and Ren, Z. (2023). Contrastive learning reduces hallucination in conversations. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 13618–13626. Touvron,H.,Lavril,T.,Izacard,G.,Martinet,X.,Lachaux...

    2023 arXiv

  29. [37]

    Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., and Scialom, T. (2023b). Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288

    2023 arXiv

  30. [38]

    Varshney, N., Yao, W., Zhang, H., Chen, J., and Yu, D. (2023). A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation.arXiv preprint arXiv:2307.03987. Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P.,...

    2023 arXiv

  31. [39]

    Wang, X., Wang, Z., Gao, X., Zhang, F., Wu, Y., Xu, Z., Shi, T., Wang, Z., Li, S., Qian, Q., et al. (2024). Searching for best practices in retrieval- augmented generation. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17716–17736

    2024

  32. [40]

    Wang, Y., Hu, X., Gan, Q., Huang, X., Qiu, X., and Wipf, D. (2025). Efficient Link Prediction via GNN Layers Induced by Negative Sampling. IEEE Transactions on Knowledge and Data Engineering, 37(1):253–264

    2025

  33. [41]

    M., and Gao, J

    Wu, Z., Chen, J., Al-Sabri, R., Oloulade, B. M., and Gao, J. (2025). Asymmetric augmented paradigm-based graph neural architecture search. Information Processing & Management, 62(1):103897

    2025

  34. [42]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Philip, S. Y. (2020). A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems, 32(1):4–24. Xing,B.andTsang,I.(2022). DARER:Dual-taskTemporalRelationalRecurrentReasoningNetwork...

    2020

  35. [43]

    Xu, F., Shi, W., and Choi, E. (2023). Recomp: Improving retrieval-augmented lms with compression and selective augmentation.arXiv preprint arXiv:2310.04408

    2023 arXiv

  36. [44]

    Yan, S.-Q., Gu, J.-C., Zhu, Y., and Ling, Z.-H. (2024). Corrective retrieval augmented generation.arXiv preprint arXiv:2401.15884

    2024 arXiv

  37. [45]

    Tu, J., Zhang, J., Ma, J., Yang, J., Xu, J., Zhou, J., Bai, J., He, J., Lin, J., Dang, K., Lu, K., Chen, K., Yang, K., Li, M., Xue, M., Ni, N., Zhang, P., Wang, P., Peng, R., Men, R., Gao, R., Lin, R., Wang, S., Bai, S., Tan, S., Zhu, T., Li, T., Liu, T., Ge, W., Deng, X., Zho...

    2024 arXiv

  38. [46]

    Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Wan, Y., Liu, Y., Cui, Z., Zhang, Z., and Qiu, Z. (2025). Qwen2.5 technical report.arXiv preprint arXiv:2412.15115

    2025 arXiv

  39. [47]

    and Zhong, G

    Yin, S. and Zhong, G. (2024). TextGT: A Double-View Graph Transformer on Text for Aspect-Based Sentiment Analysis. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19404–19412

    2024

  40. [48]

    Zhang, Y., Jia, A., Wang, B., Zhang, P., Zhao, D., Li, P., Hou, Y., Jin, X., Song, D., and Qin, J. (2023a). M3GAT: A Multi-Modal Multi-Task Interactive Graph Attention Network for Conversational Sentiment Analysis and Emotion Recognition.ACM Transactions on Information Systems...

    2023 arXiv

  41. [49]

    E., and Stoica, I

    Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., Zhang, H., Gonzalez, J. E., and Stoica, I. (2023). JudgingLLM-as-a-JudgewithMT-BenchandChatbotArena. InAdvancesinNeuralInformationProcessingSystems,volume36,pages 46595–4662...

    2023

  42. [50]

    Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang, L., Li, C., and Sun, M. (2020). Graph neural networks: A review of methods and applications.AI Open, 1:57–81. Ao Jia et al.:Preprint submitted to ElsevierPage 24 of 24 MALM: A Multi-Information Adapter for Large La...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.