HyST: LLM-Powered Hybrid Retrieval over Semi-Structured Tabular Data

Jihyeon Park; Jiyoon Myung; Joohyung Han

REVIEW 3 major objections 4 minor 1 cited by

HyST: LLM-Powered Hybrid Retrieval over Semi-Structured Tabular Data

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read HyST claims hybrid recommendation queries are best answered by splitting them: an LLM turns hard constraints into metadata filters applied before embedding-based search, and the split wins every metric on an adapted STaRK benchmark.

desk verdict The claimed gains are likely inflated by the curated candidate pool, and the core method is prior work. read the letter →

arxiv 2508.18048 v1 pith:EHEK4JLT submitted 2025-08-25 cs.IR cs.AI

Jiyoon Myung , Jihyeon Park , Joohyung Han This is my paper

classification cs.IRcs.AI

keywords hybridretrievalsemi-structureddatametadatafilteringLLMquerydecompositionvectordatabaserecommendersystemssemanticsearchSTaRKbenchmark

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

HyST claims that realistic recommendation queries — which mix hard, checkable conditions like brand or category with subjective preferences like 'great firepower' or 'user-friendly' — are better answered by splitting them apart before retrieval. An LLM converts the hard conditions into metadata filters that a vector database applies first, and the leftover subjective phrasing drives embedding-based search over product descriptions and reviews. On 76 hybrid queries adapted from the STaRK Amazon benchmark, the framework reports the best score on all five metrics, with the largest P@5 and P@10 advantages over the common alternative of flattening every field into one text string and embedding the whole thing. The paper's own ablation shows the load-bearing piece is the structured filter, not the query-refinement step; refinement even costs a little precision and is presented as optional. A sympathetic reader would care because the design uses only standard components — an LLM, an embedding model, and a vector database's native filtering — so the win is plausibly portable to production pipelines.

What carries the argument

The carrying mechanism is query decomposition enforced as a pre-filter. The LLM is prompted with the target schema (columns, allowable values, operators like $eq, $in, $lt, $between) to produce a JSON filter that the vector database applies before similarity search; the remaining query text is embedded and matched by cosine similarity against product text embeddings. The key property this exploits is that exact constraints are enforced structurally by the database rather than being left to the embedding model to infer. A second, smaller component — query refinement, isolating the subjective residue — turns out in the paper's own ablation to be non-essential and is recommended as an optional

What would settle it

Run HyST and the linearized semantic baseline on the full uncurated 1M-product STaRK catalog with the same 76 queries, then on a larger query sample including price and rating ranges. The central claim survives if HyST still shows large P@5/P@10 advantages; it is undercut if the gaps shrink or reverse. A second check targets the bottleneck: sample hundreds of queries and audit the LLM's extracted filters for hallucinated or schema-mismatched attributes — one failure in 76 queries is small, but a materially higher rate would make filter accuracy, not retrieval architecture, the limiting factor.

Watch

Extended reading notes

Core claim

HyST's central claim is that a hybrid user query should not be embedded as one blob of text. Instead, an LLM prompted with the schema and its allowable values emits a vector-database metadata filter (brand $eq 'Martin', category $in ['archery']), and the residual subjective phrasing ('user-friendly and highly precise') drives dense retrieval over the filtered candidates. On 76 hybrid queries adapted from the STaRK Amazon benchmark, this decomposition reports the best Precision@1/5/10, Recall@20, and MRR among all baselines, and the case studies show why: purely semantic retrieval returns a similar item from the wrong brand, while HyST enforces the constraint exactly. The ablation adds nuance

Load-bearing premise

The evaluation rests on a hand-curated corpus: 76 manually selected queries and only the 3,335 products that are either ground-truth answers or match the brands and categories those queries name (out of 1M+ entities). If that curation happens to favor methods that enforce exact constraints, or the 76 queries are unrepresentative, the reported superiority may not carry over to the full catalog.

Editorial extensions

If this is right

On hybrid product queries, pre-retrieval metadata filtering raises precision at deeper ranks most: HyST reports +14.0 points in P@5 and +20.5 points in P@10 over the linearized semantic baseline, with roughly equal Recall@20.
The approach needs no SQL translation and no row-by-row LLM inference over candidates, so it inherits the scalability of a vector database's native filter-then-search path.
Retrieval quality is bounded by the two components: an LLM's filter-extraction accuracy (one hallucinated attribute seen in 76 queries) and the embedding model's semantic expressiveness.
Query refinement is optional: when catalog text already carries structured cues, keeping the full query in the embedding step can beat the refined version.

Reading between the lines

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

I infer the hardest untested regime is numeric and multi-condition filtering: the adapted STaRK subset has no price or rating range queries, even though the prompt defines $gt and $between operators. A natural probe is to add range queries to this catalog and compare the gap against the linearized baseline — the gap should widen if constraint enforcement is really the mechanism.
I infer the curated corpus (3,335 of 1M+ products, only ground-truth answers plus queried brands and categories) may remove exactly the distractors that would stress LLM schema alignment at scale; rerunning on the full catalog or on a random query sample would settle whether the reported margins transfer.
I infer a cheap adaptive rule from the ablation: skip query refinement whenever the unstructured fields repeat brand or category tokens (descriptions often do), and use it only when the non-structured text is semantically disjoint from the schema.

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.

Desk Editor's Note

read the letter

Read this if you want an example of how easy it is to design an evaluation that makes your method look good. HyST is an LLM-plus-vector-DB pipeline: extract structured filters from the query, apply them as metadata filters, then do embedding search on the remaining text. That exact recipe is already in their reference [7] (Poliakov & Shvai), so the novelty is only the product-domain application and the specific benchmark adaptation. The paper is clear and the pipeline is reasonable, and I credit the authors for including an ablation that shows their query-refinement step does not help—that takes some honesty.

The soft spot is the evaluation. They manually selected 76 queries, then built a candidate pool of 3,335 products by keeping only ground-truth answers and products matching the brands/categories named in those queries. That means for any given query, all true answers are guaranteed to be in the pool, and unrelated products whose brands/categories don't appear in any query are excluded. Baselines like linearized semantic retrieval search over the whole 3,335-item pool, which includes products relevant to other queries' constraints, so they have a harder retrieval task. The paper acknowledges in the limitations that this 'reduces realism,' but it doesn't address how much this inflates HyST's margins. On 76 queries with no error bars or significance tests, the +2.6 P@1 difference over linearized retrieval is noise; the larger P@5 and P@10 gaps might be real or might be artifacts of the curation.

There's also a smaller issue: the ablation shows that using the original full query instead of the refined query gives better P@5 and P@10, which is a direct hit to one of the three pipeline stages. The paper handles this by recommending the refinement be optional, but it weakens the 'query refinement' selling point.

Overall, this is a workshop-level paper with a simple method and a flawed benchmark. The central idea is not new and the empirical evidence isn't reliable as reported. But the authors are transparent about many limitations, and the prompt in the appendix might be useful to someone building a similar pipeline. I would not cite the numbers, but I wouldn't desk-reject it either; send it to a reviewer with the request to re-run on an unfiltered sample and add significance tests. If they release the query list and subset, it could become a useful data point for the community.

Referee Report

3 major / 4 minor

Summary. The paper proposes HyST, a retrieval pipeline for semi-structured tabular data that uses an LLM to extract structured metadata filters (e.g., brand, category) from a natural-language query and applies those filters in a vector database before performing embedding-based semantic search over the remaining unstructured query. The authors adapt the STaRK Amazon benchmark to a flat table, manually select 76 hybrid queries, curate a 3,335-product subset, and compare HyST against BM25, DPR, ANCE, ColBERTv2, BM25+DPR, BM25+OpenAI embeddings, and linearized semantic retrieval. They report that HyST achieves the best scores on Precision@1/5/10, Recall@20, and MRR (Table 2), and include an ablation of the query-refinement step and a small case study.

Significance. If the empirical result were robust, HyST would be a useful practical contribution: it demonstrates a simple, modular way to combine LLM-generated structured filters with dense retrieval using native vector-database filtering, and it compares against a strong linearized semantic baseline that uses the same embedding model. The paper is transparent about limitations and includes a useful ablation and error analysis. However, the central claim of consistent superiority over baselines rests on a small, manually curated evaluation set and a single run with no uncertainty quantification, so the magnitude of the reported advantages is not yet established.

major comments (3)

[Section 4.1 / Table 2] The evaluation corpus is constructed by keeping only ground-truth answers and products matching the brands/categories mentioned in the 76 manually selected queries, reducing the STaRK pool from over 1M to 3,335 products. The paper itself states in Section 6 that this reduces realism. Since the central claim is empirical superiority over all baselines, this curation is load-bearing. I do not agree with the stronger version of the stress-test concern: removing products whose brand/category never appear in any query could actually help the unfiltered baselines by deleting irrelevant distractors, so it is not obvious that this curation inflates HyST's margin. However, the external validity of the reported margins cannot be assessed without an evaluation on an unfiltered or randomly sampled corpus. Please add such an experiment, or clearly re-frame the claims as results on a purpose-built sub
[Section 5.1 / Table 2] With 76 queries, the reported differences are small on key metrics: HyST's P@1 advantage over Linearized Semantic Retrieval is 0.9211 vs 0.8947, i.e., roughly 2 queries, and Recall@20 is essentially tied (0.8063 vs 0.8019, less than one query). No error bars, bootstrap/permutation intervals, or per-query variance are reported. The claim that HyST 'consistently outperforms' baselines on all metrics is therefore unsupported at the margins. Please report uncertainty estimates, and temper the 'best on all metrics' wording where differences are within noise.
[Section 3.2 / Section 5.2] The ablation in Table 3 shows that removing the query-refinement step yields higher Precision@5 (0.8586 vs 0.8349) and Precision@10 (0.8167 vs 0.8022), with comparable P@1 and Recall@20. This means the full HyST pipeline is not better than its own no-refinement variant on half of the reported metrics. The authors acknowledge this and suggest treating refinement as optional, but the presentation still describes query refinement as a core stage of the framework. Please clarify which variant is the 'HyST' reported in Table 2 and consider making the no-refinement variant the primary method, or presenting both as equally valid configurations.

minor comments (4)

[Abstract / Section 1] Typo: 'tradtional' should be 'traditional'.
[Section 4.1] The manual selection of 76 queries is not described with enough detail to assess representativeness. Please provide the selection criteria or a reference to the full list of chosen query IDs.
[Section 4.3] DPR, ANCE, and ColBERTv2 are used as pretrained models without domain adaptation, while HyST and the Linearized Semantic Retrieval baseline use OpenAI's text-embedding-3-small. The comparison is still informative, but the paper should explicitly note that the pretrained dense models are not fine-tuned on the STaRK domain.
[Section 6] The limitations paragraph is candid about the restricted filtering logic and the cleaner subset of schema values. These points should also be stated in the Experimental Setup section, not only as future work.

Circularity Check

1 steps flagged · score 4.0 of 10

Benchmark curation in §4.1 overlaps with HyST's own brand/category filter, creating a structural advantage for HyST; final ranking is not fully forced.

self definitional [Section 4.1 (Dataset construction), interacting with Section 3.1 and 3.3]
"To build a focused yet challenging evaluation set, we included only products that either appeared as ground-truth answers to the selected queries or matched the brands or categories specified within them."

The curated evaluation corpus is defined by the brands/categories mentioned in the queries. HyST's first stage (§3.1) extracts exactly those brand/category constraints and applies them as metadata filters (§3.3). Therefore, for each query, the post-filter candidate set is precisely the subset of the curated corpus that the curation rule already singled out; the filter cannot be 'wrong' at the corpus level, and it systematically removes constraint-violating products that unfiltered baselines must search. This makes the comparison of filter-based vs. non-filter-based retrieval partly a consequence of the benchmark's inclusion criterion rather than an independent measurement. The paper itself acknowledges in §6 that limiting to a cleaner subset 'reduces realism.' The final ranking over the fi

full rationale

This is an empirical systems paper with no mathematical derivation chain, so most classical circularity patterns (self-citation, uniqueness-import, ansatz-smuggling, renaming) are absent. The only significant structural overlap is in the evaluation design: the §4.1 corpus construction keeps products matching the queries' brands/categories, and HyST's core mechanism is to filter on exactly those brands/categories. This does not make the final P@k/MRR scores equal to the filter by construction (HyST still must rank, and its scores are imperfect), but it does guarantee that HyST's filtering stage is aligned with the benchmark's inclusion rule while baselines receive no such alignment. The paper's own limitation statement (§6: 'it also reduces realism') supports the concern. No load-bearing self-citation exists, and the LLM filter generation is tested against a real QA benchmark, so the central claim retains independent empirical content. A moderate score of 4 reflects this partial structural circularity without overstating it as a fully forced result.

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

HyST itself has no fitted numerical parameters; the free parameters listed are evaluation choices and baseline settings. The main assumptions are that the LLM filters are reliable and that the curated benchmark is representative.

free parameters (4)

Number of selected queries = 76
Manually chosen by the authors from the STaRK QA set to require both structured filtering and semantic reasoning. The central evaluation depends on this hand-picked sample.
Product subset size = 3,335
Only products that are ground-truth answers or match brands/categories in the selected queries were kept, reducing the corpus from 1M+ to 3,335. This curation affects all methods and may bias the comparison.
BM25+DPR fusion weight lambda = 0.5
Set to 0.5 for the hybrid baseline without tuning; this affects the baseline comparison in Table 2.
LLM sampling temperature/top-p = 0.3 / 0.8
Chosen by hand for GPT-4o filter generation; results could vary with different settings.

assumptions (4)

domain assumption LLM-generated metadata filters are accurate enough for the evaluated queries
Section 5.4 reports only one hallucination, but filter quality is not measured independently. The framework's success depends on this.
domain assumption STaRK queries can be decomposed into structured and unstructured parts
Section 3 decomposes each query; the paper excludes multi-hop queries that do not fit this assumption.
domain assumption Embedding similarity measures subjective relevance
Section 3.3 ranks candidates by cosine similarity after filtering; no validation of embedding quality beyond final metrics.
domain assumption The curated subset preserves retrieval difficulty
Section 4.1 claims the subset preserves difficulty, but removing most negatives changes the task distribution.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HyST: LLM-Powered Hybrid Retrieval over Semi-Structured Tabular Data." pith.science (2026). https://pith.science/paper/EHEK4JLT

@misc{pith2026250818048,
  author       = {Pith},
  title        = {Pith review of: HyST: LLM-Powered Hybrid Retrieval over Semi-Structured Tabular Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHEK4JLT}},
  note         = {Machine review of arXiv:2508.18048}
}

read the original abstract

User queries in real-world recommendation systems often combine structured constraints (e.g., category, attributes) with unstructured preferences (e.g., product descriptions or reviews). We introduce HyST (Hybrid retrieval over Semi-structured Tabular data), a hybrid retrieval framework that combines LLM-powered structured filtering with semantic embedding search to support complex information needs over semi-structured tabular data. HyST extracts attribute-level constraints from natural language using large language models (LLMs) and applies them as metadata filters, while processing the remaining unstructured query components via embedding-based retrieval. Experiments on a semi-structured benchmark show that HyST consistently outperforms tradtional baselines, highlighting the importance of structured filtering in improving retrieval precision, offering a scalable and accurate solution for real-world user queries.

Figures

Figures reproduced from arXiv: 2508.18048 by the authors.

**Figure 1.** Overview of the HyST framework. Structured constraints are handled via metadata filtering, while unstructured [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Human Preference aligned Tabular Similarity
cs.LG 2026-07 conditional novelty 4.0 of 10

Downstream prediction metrics do not certify human-aligned tabular similarity rankings; a pairwise preference workflow and pilot show user-specific embedding quality gaps.

Reference graph

Works this paper leans on

18 extracted references · 10 canonical work pages · cited by 1 Pith paper

[7]

Mykhailo Poliakov and Nadiya Shvai. 2025. Multi-Meta-RAG: Improving RAG for Multi-hop Queries Using Database Filtering with LLM-Extracted Metadata . Springer Nature Switzerland, 334–342. https://doi.org/10.1007/978-3-031-81372- 6_25

work page doi:10.1007/978-3-031-81372- 2025
[1]

Gon- zalez, Carlos Guestrin, and Matei Zaharia

Asim Biswal, Liana Patel, Siddarth Jha, Amog Kamsetty, Shu Liu, Joseph E. Gon- zalez, Carlos Guestrin, and Matei Zaharia. 2024. Text2SQL is Not Enough: Unifying AI and Databases with TAG. arXiv:2408.14717 [cs.DB] https: //arxiv.org/abs/2408.14717

arXiv 2024
[2]

Cormack, Charles L A Clarke, and Stefan Buettcher

Gordon V. Cormack, Charles L A Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (Boston, MA, USA) (SIGIR ’09). Association for Computing Machinery, New York, NY, USA, 758–7...

arXiv 2009
[4]

Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. arXiv:2004.04906 [cs.CL] https://arxiv.org/abs/ 2004.04906

arXiv 2020
[5]

Semnani, Chen Jie Yu, and Monica S

Shicheng Liu, Jialiang Xu, Wesley Tjangnaka, Sina J. Semnani, Chen Jie Yu, and Monica S. Lam. 2024. SUQL: Conversational Search over Structured and Unstructured Data with Large Language Models. arXiv:2311.09818 [cs.CL] https://arxiv.org/abs/2311.09818

arXiv 2024
[6]

Barlas Oguz, Xilun Chen, Vladimir Karpukhin, Stan Peshterliev, Dmytro Okhonko, Michael Schlichtkrull, Sonal Gupta, Yashar Mehdad, and Scott Yih. 2022. UniK-QA: Unified Representations of Structured and Unstructured Knowledge for Open- Domain Question Answering. In Findings of the Association for Computational Linguistics: NAACL 2022 , Marine Carpuat, Mari...

2022
[8]

Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109

1995
[9]

Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction. arXiv:2112.01488 [cs.IR] https://arxiv.org/abs/2112.01488

arXiv 2022

Show all 18 references

[10]

Ioannidis, Karthik Subbian, James Zou, and Jure Leskovec

Shirley Wu, Shiyu Zhao, Michihiro Yasunaga, Kexin Huang, Kaidi Cao, Qian Huang, Vassilis N. Ioannidis, Karthik Subbian, James Zou, and Jure Leskovec
[11]

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval. arXiv:2007.00808 [cs.IR] https://arxiv.org/abs/2007.00808 HyST: LLM-Powered ...

2020 arXiv
[13]

Identify structured components such as category , location , brand , price , or other attributes based on the given schema
[14]

$eq " for exact matches when the column type is

Use appropriate comparison operators : - " $eq " for exact matches when the column type is " single " ( e . g . , brand ) . - " $in " for multiple options when the column type is " multiple " or the query mentions a list of possible values . - " $lt " for less than comparisons...
[15]

If the query contains multiple conditions , combine them using AND logic
[16]

$in " operator for columns specified as

Always use the " $in " operator for columns specified as " multiple " in the schema , even if the input has only one value
[17]

Ignore subjective or descriptive language that cannot be directly used for structured filtering
[18]

Always match the categorical values strictly with the allowable values specified in the schema
[19]

CATEGORY

Answer only the filtering conditions in JSON format . ### Example Queries : Query 1: Find cozy Italian or French restaurants in New York with prices under $100 . Expected Output : { " CATEGORY ": {" $in ": [" Italian " , " French "]} , " LOCATION ": {" $eq ": " New York } , " ...
[2024]

arXiv:2404.13207 [cs.IR] https://arxiv.org/abs/2404.13207

STaRK: Benchmarking LLM Retrieval on Textual and Relational Knowledge Bases. arXiv:2404.13207 [cs.IR] https://arxiv.org/abs/2404.13207

arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [7]

Mykhailo Poliakov and Nadiya Shvai. 2025. Multi-Meta-RAG: Improving RAG for Multi-hop Queries Using Database Filtering with LLM-Extracted Metadata . Springer Nature Switzerland, 334–342. https://doi.org/10.1007/978-3-031-81372- 6_25

work page doi:10.1007/978-3-031-81372- 2025

[2] [1]

Gon- zalez, Carlos Guestrin, and Matei Zaharia

Asim Biswal, Liana Patel, Siddarth Jha, Amog Kamsetty, Shu Liu, Joseph E. Gon- zalez, Carlos Guestrin, and Matei Zaharia. 2024. Text2SQL is Not Enough: Unifying AI and Databases with TAG. arXiv:2408.14717 [cs.DB] https: //arxiv.org/abs/2408.14717

arXiv 2024

[3] [2]

Cormack, Charles L A Clarke, and Stefan Buettcher

Gordon V. Cormack, Charles L A Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (Boston, MA, USA) (SIGIR ’09). Association for Computing Machinery, New York, NY, USA, 758–7...

arXiv 2009

[4] [4]

Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. arXiv:2004.04906 [cs.CL] https://arxiv.org/abs/ 2004.04906

arXiv 2020

[5] [5]

Semnani, Chen Jie Yu, and Monica S

Shicheng Liu, Jialiang Xu, Wesley Tjangnaka, Sina J. Semnani, Chen Jie Yu, and Monica S. Lam. 2024. SUQL: Conversational Search over Structured and Unstructured Data with Large Language Models. arXiv:2311.09818 [cs.CL] https://arxiv.org/abs/2311.09818

arXiv 2024

[6] [6]

Barlas Oguz, Xilun Chen, Vladimir Karpukhin, Stan Peshterliev, Dmytro Okhonko, Michael Schlichtkrull, Sonal Gupta, Yashar Mehdad, and Scott Yih. 2022. UniK-QA: Unified Representations of Structured and Unstructured Knowledge for Open- Domain Question Answering. In Findings of the Association for Computational Linguistics: NAACL 2022 , Marine Carpuat, Mari...

2022

[7] [8]

Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995), 109

1995

[8] [9]

Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction. arXiv:2112.01488 [cs.IR] https://arxiv.org/abs/2112.01488

arXiv 2022