AnTenA: Actionable and Explainable Tensor Analysis System with Large Language Models

Auder Der; Dawon Ahn; Evangelos E. Papalexakis

arxiv: 2606.28708 · v1 · pith:4WXRIHBUnew · submitted 2026-06-27 · 💻 cs.CL · cs.LG

AnTenA: Actionable and Explainable Tensor Analysis System with Large Language Models

Dawon Ahn , Auder Der , Evangelos E. Papalexakis This is my paper

Pith reviewed 2026-06-30 10:13 UTC · model grok-4.3

classification 💻 cs.CL cs.LG

keywords tensor decompositionlarge language modelsexplainable AIhuman narrativesco-clusteringlatent patternsinference taskspattern explanation

0 comments

The pith

AnTenA uses large language models to explain latent patterns in human narratives extracted via tensor decomposition.

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

The paper proposes AnTenA as a method to explain hidden patterns in multi-aspect data like human narratives without relying on labels or auxiliary metadata, which are often inaccurate or unavailable. It extracts co-clustered latent patterns through tensor decomposition and then applies task-agnostic and task-specific prompts to large language models to generate explanations. These explanations are evaluated by testing the models on forward and backward inference tasks. A sympathetic reader would care because this approach turns complex tensor outputs into actionable insights from data where traditional supervision falls short.

Core claim

AnTenA leverages the knowledge of large language models to explain the hidden patterns in human narratives. It uses task-agnostic and task-specific prompts to explain extracted co-clustered latent patterns from tensor decomposition. To evaluate these explanations, the system tests the LLMs on forward and backward inference tasks.

What carries the argument

Tensor decomposition to identify co-clustered latent patterns, combined with LLM prompting for generating explanations of those patterns.

If this is right

Explanations for data patterns become available even when labels are nonstandard, inconsistent, or missing.
Time-dependent recordings can be interpreted without depending on static tabular metadata.
Tensor analysis outputs move from opaque clusters to testable explanations via inference performance.
The method supports generating actionable insights directly from the decomposition results.

Where Pith is reading between the lines

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

The approach could apply to any tensor-representable dataset where human-interpretable descriptions are needed, such as user behavior logs or sensor streams.
Success on inference tasks might serve as a proxy for explanation quality that avoids the need for separate human validation studies.
If the prompts generalize, the system reduces the cost of making latent factor models interpretable across domains.

Load-bearing premise

The LLM-generated explanations accurately capture the meaning of the tensor-derived patterns, as measured by success on the inference tasks.

What would settle it

An experiment showing that LLM responses based on the tensor patterns produce no gain or a loss in accuracy on the forward and backward inference tasks relative to prompts that ignore the tensor output.

Figures

Figures reproduced from arXiv: 2606.28708 by Auder Der, Dawon Ahn, Evangelos E. Papalexakis.

**Figure 1.** Figure 1: Overview of AnTenA using DBLP dataset and a CP model. We explain our demo system using a DBLP dataset as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

read the original abstract

Accurately explaining hidden patterns in multi-aspect data has typically been done by leveraging labels and/or accompanying auxiliary metadata. However, labels and auxiliary data may be inaccurate (e.g. nonstandard, inconsistent), insufficient (e.g. static tabular metadata for time-dependent recordings), or unavailable. % We propose \fullmethod (\method), which leverages the knowledge of large language models (LLMs) to explain the hidden patterns in human narratives. \method uses task-agnostic and task-specific prompts to explain extracted co-clustered latent patterns from tensor decomposition. To evaluate these explanations, we test the LLMs on forward and backward inference tasks. % Our demo system is available at https://github.com/dawonahn/ECML_PKDD_AnTenA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AnTenA sketches an LLM-based explainer for tensor factors on narrative data but gives no evidence the explanations track the factors or that the inference tests validate them.

read the letter

The main thing to know is that this paper proposes a system called AnTenA that runs tensor decomposition on multi-aspect narrative data, then feeds the resulting co-clustered patterns to an LLM via task-agnostic and task-specific prompts to generate explanations, with evaluation planned through forward and backward inference tasks.

What is new is the concrete pairing of tensor co-clustering with LLM prompting to handle label-scarce settings in human narratives. The abstract correctly notes that labels and metadata are often missing, inconsistent, or static while the data itself is time-dependent.

The paper does a reasonable job stating the problem and the high-level architecture. It identifies a real gap where standard tensor methods produce patterns that are hard to interpret without auxiliary information.

The soft spots are the lack of any methods, results, metrics, baselines, or error analysis. The stress-test concern is on target: success on the inference tasks could come from the LLM's pre-trained knowledge alone, and nothing in the description shows controls such as ablations that remove the tensor-derived explanations or comparisons against generic prompts. Without those, the claim that the explanations are accurate or actionable has no grounding.

This is for people already working on tensor methods for text or narrative data who want to experiment with LLM wrappers for interpretability. A reader in that niche might pick up the prompt strategy as a starting idea, but there is nothing here to cite or build on directly.

I would not send it for peer review in this form. It needs the actual experiments with proper controls and reported outcomes before it can be evaluated seriously.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes AnTenA, a system that uses large language models (LLMs) with task-agnostic and task-specific prompts to generate explanations for co-clustered latent patterns extracted via tensor decomposition from multi-aspect human narrative data. These explanations are evaluated by testing LLM performance on forward and backward inference tasks. The abstract notes that labels and auxiliary metadata are often unavailable or insufficient and positions the approach as a way to produce actionable explanations without them; a demo is referenced on GitHub.

Significance. If the central claim is substantiated, the work could offer a practical route to interpretable tensor analysis in domains where labels are absent or unreliable, bridging tensor methods with LLM reasoning for narrative data. The approach is novel in its use of prompt engineering for post-decomposition explanation, but the absence of any reported methods, metrics, baselines, or controls in the abstract leaves the significance unassessable at present.

major comments (2)

[Abstract] Abstract: The claim that LLM-generated explanations are evaluated via forward and backward inference tasks is load-bearing for the central contribution, yet the abstract supplies no description of the tasks themselves, the metrics used, the baselines or ablations (e.g., prompts without tensor-derived patterns, random patterns, or no explanations), or any human fidelity checks. Without such controls it is impossible to determine whether task success can be attributed to the quality of the explanations rather than the LLM's pre-existing knowledge.
[Abstract] Abstract: No information is given on the tensor decomposition method, the definition of co-clustered patterns, the data domain or size, or how task-agnostic versus task-specific prompts are constructed. These omissions make it impossible to assess reproducibility or to evaluate whether the inference-task results actually test fidelity to the tensor-derived patterns.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful comments. We address each major comment below and have revised the abstract to provide additional high-level information on the evaluation tasks, methods, and controls while respecting length constraints.

read point-by-point responses

Referee: [Abstract] Abstract: The claim that LLM-generated explanations are evaluated via forward and backward inference tasks is load-bearing for the central contribution, yet the abstract supplies no description of the tasks themselves, the metrics used, the baselines or ablations (e.g., prompts without tensor-derived patterns, random patterns, or no explanations), or any human fidelity checks. Without such controls it is impossible to determine whether task success can be attributed to the quality of the explanations rather than the LLM's pre-existing knowledge.

Authors: We agree the abstract is concise and omits these specifics. The full manuscript describes the forward and backward inference tasks (predicting aspects from explanations and inferring missing aspects from explanations), the accuracy metric, and control conditions including ablations with prompts lacking tensor-derived patterns and random patterns. We have revised the abstract to briefly outline the evaluation tasks and mention the controls used to link performance to the explanations. revision: yes
Referee: [Abstract] Abstract: No information is given on the tensor decomposition method, the definition of co-clustered patterns, the data domain or size, or how task-agnostic versus task-specific prompts are constructed. These omissions make it impossible to assess reproducibility or to evaluate whether the inference-task results actually test fidelity to the tensor-derived patterns.

Authors: The abstract prioritizes brevity and therefore omits these details, which are provided in the manuscript (tensor decomposition approach and co-clustered pattern definition in the methods, data domain as multi-aspect human narratives with experimental sizes in the experiments section, and prompt construction for task-agnostic vs. task-specific variants in the prompt engineering section). This supports reproducibility and fidelity assessment. We have revised the abstract to include high-level references to the data domain, decomposition for co-clustering, and the two prompt types. revision: yes

Circularity Check

0 steps flagged

No derivation chain or equations present; system description only

full rationale

The provided abstract and description contain no equations, no fitted parameters, no predictions of derived quantities, and no self-citations invoked to justify a central premise. AnTenA is presented as a prompting-based system that feeds tensor co-clusters into LLMs for explanations, then evaluates via forward/backward inference tasks. No step reduces a claimed result to its own inputs by construction, self-definition, or load-bearing self-citation. The work is a methodological proposal without a mathematical derivation chain, so circularity does not apply; the content is self-contained as a high-level system outline.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The proposal rests on the assumption that LLMs possess sufficient general knowledge to interpret tensor-derived patterns in narratives; no free parameters, invented entities, or additional axioms are stated in the abstract.

axioms (1)

domain assumption Large language models contain knowledge sufficient to generate accurate explanations of latent patterns in human narratives
The system is built around prompting LLMs for explanations without additional supervision.

pith-pipeline@v0.9.1-grok · 5663 in / 1191 out tokens · 38903 ms · 2026-06-30T10:13:41.080581+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages · 1 internal anchor

[1]

All-at-once Optimization for Coupled Matrix and Tensor Factorizations

Acar, E., Kolda, T.G., Dunlavy, D.M.: All-at-once optimization for coupled matrix and tensor factorizations. arXiv preprint arXiv:1105.3422 (2011)

work page internal anchor Pith review Pith/arXiv arXiv 2011
[2]

In: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management

Ahn, D., Saini, U.S., Papalexakis, E.E., Payani, A.: Neural additive tensor de- composition for sparse tensors. In: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. p. 14–23. CIKM ’24, As- sociation for Computing Machinery, New York, NY, USA (2024).https://doi. org/10.1145/3627673.3679833,https://doi.org/10.114...

work page doi:10.1145/3627673.3679833 2024
[3]

In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management

Ahn, D., Son, S., Kang, U.: Gtensor: Fast and accurate tensor analysis system using gpus. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management. pp. 3361–3364 (2020)

2020
[4]

LLMs for explainable AI: A compre- hensive survey,

Bilal, A., Ebert, D., Lin, B.: Llms for explainable ai: A comprehensive survey. arXiv preprint arXiv:2504.00125 (2025)

work page arXiv 2025
[5]

SIAM review 51(3), 455–500 (2009)

Kolda, T.G., Bader, B.W.: Tensor decompositions and applications. SIAM review 51(3), 455–500 (2009)

2009
[6]

arXiv preprint arXiv:2308.01157 (2023)

Lengerich, B.J., Bordt, S., Nori, H., Nunnally, M.E., Aphinyanaphongs, Y., Kellis, M.,Caruana,R.:Llmsunderstandglass-boxmodels,discoversurprises,andsuggest repairs. arXiv preprint arXiv:2308.01157 (2023)

work page arXiv 2023
[7]

arXiv preprint arXiv:2410.15268 (2024)

Pan, B., Xiong, Z., Wu, G., Zhang, Z., Zhang, Y., Zhao, L.: Tagexplainer: Narrat- ing graph explanations for text-attributed graph learning models. arXiv preprint arXiv:2410.15268 (2024)

work page arXiv 2024
[8]

In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management

Park, N., Jeon, B., Lee, J., Kang, U.: Bigtensor: Mining billion-scale tensor made easy. In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. pp. 2457–2460 (2016)

2016
[9]

IEEE Transactions on signal processing65(13), 3551–3582 (2017)

Sidiropoulos,N.D.,DeLathauwer,L.,Fu,X.,Huang,K.,Papalexakis,E.E.,Falout- sos, C.: Tensor decomposition for signal processing and machine learning. IEEE Transactions on signal processing65(13), 3551–3582 (2017)

2017
[10]

Usable XAI: 10 strategies to- wards exploiting explainability in the LLM era,

Wu, X., Zhao, H., Zhu, Y., Shi, Y., Yang, F., Liu, T., Zhai, X., Yao, W., Li, J., Du, M., et al.: Usable xai: 10 strategies towards exploiting explainability in the llm era. arXiv preprint arXiv:2403.08946 (2024)

work page arXiv 2024
[11]

Transactions of the Association for Computational Linguistics13, 357–375 (2025)

Yang, X., Zhao, H., Phung, D., Buntine, W., Du, L.: Llm reading tea leaves: Au- tomatically evaluating topic models with large language models. Transactions of the Association for Computational Linguistics13, 357–375 (2025)

2025
[12]

In: The Thirty-eighth Annual Conference on Neural Information Processing Systems

Yang, X., Wang, X.: Language model as visual explainer. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems

[1] [1]

All-at-once Optimization for Coupled Matrix and Tensor Factorizations

Acar, E., Kolda, T.G., Dunlavy, D.M.: All-at-once optimization for coupled matrix and tensor factorizations. arXiv preprint arXiv:1105.3422 (2011)

work page internal anchor Pith review Pith/arXiv arXiv 2011

[2] [2]

In: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management

Ahn, D., Saini, U.S., Papalexakis, E.E., Payani, A.: Neural additive tensor de- composition for sparse tensors. In: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. p. 14–23. CIKM ’24, As- sociation for Computing Machinery, New York, NY, USA (2024).https://doi. org/10.1145/3627673.3679833,https://doi.org/10.114...

work page doi:10.1145/3627673.3679833 2024

[3] [3]

In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management

Ahn, D., Son, S., Kang, U.: Gtensor: Fast and accurate tensor analysis system using gpus. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management. pp. 3361–3364 (2020)

2020

[4] [4]

LLMs for explainable AI: A compre- hensive survey,

Bilal, A., Ebert, D., Lin, B.: Llms for explainable ai: A comprehensive survey. arXiv preprint arXiv:2504.00125 (2025)

work page arXiv 2025

[5] [5]

SIAM review 51(3), 455–500 (2009)

Kolda, T.G., Bader, B.W.: Tensor decompositions and applications. SIAM review 51(3), 455–500 (2009)

2009

[6] [6]

arXiv preprint arXiv:2308.01157 (2023)

Lengerich, B.J., Bordt, S., Nori, H., Nunnally, M.E., Aphinyanaphongs, Y., Kellis, M.,Caruana,R.:Llmsunderstandglass-boxmodels,discoversurprises,andsuggest repairs. arXiv preprint arXiv:2308.01157 (2023)

work page arXiv 2023

[7] [7]

arXiv preprint arXiv:2410.15268 (2024)

Pan, B., Xiong, Z., Wu, G., Zhang, Z., Zhang, Y., Zhao, L.: Tagexplainer: Narrat- ing graph explanations for text-attributed graph learning models. arXiv preprint arXiv:2410.15268 (2024)

work page arXiv 2024

[8] [8]

In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management

Park, N., Jeon, B., Lee, J., Kang, U.: Bigtensor: Mining billion-scale tensor made easy. In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. pp. 2457–2460 (2016)

2016

[9] [9]

IEEE Transactions on signal processing65(13), 3551–3582 (2017)

Sidiropoulos,N.D.,DeLathauwer,L.,Fu,X.,Huang,K.,Papalexakis,E.E.,Falout- sos, C.: Tensor decomposition for signal processing and machine learning. IEEE Transactions on signal processing65(13), 3551–3582 (2017)

2017

[10] [10]

Usable XAI: 10 strategies to- wards exploiting explainability in the LLM era,

Wu, X., Zhao, H., Zhu, Y., Shi, Y., Yang, F., Liu, T., Zhai, X., Yao, W., Li, J., Du, M., et al.: Usable xai: 10 strategies towards exploiting explainability in the llm era. arXiv preprint arXiv:2403.08946 (2024)

work page arXiv 2024

[11] [11]

Transactions of the Association for Computational Linguistics13, 357–375 (2025)

Yang, X., Zhao, H., Phung, D., Buntine, W., Du, L.: Llm reading tea leaves: Au- tomatically evaluating topic models with large language models. Transactions of the Association for Computational Linguistics13, 357–375 (2025)

2025

[12] [12]

In: The Thirty-eighth Annual Conference on Neural Information Processing Systems

Yang, X., Wang, X.: Language model as visual explainer. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems