An Ensemble Embedding Approach for Improving Semantic Caching Performance in LLM-based Systems

Mohammad Akbari; Shervin Ghaffari; Zohre Bahranifard

REVIEW 4 major objections 5 minor 31 references

An Ensemble Embedding Approach for Improving Semantic Caching Performance in LLM-based Systems

T0 review · 4 major / 5 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 Fusing two low-correlated embedding models through a trained meta-encoder detects semantically equivalent queries better than any single model, yielding a 92% cache hit ratio and 85% correct rejection of non-equivalent queries.

desk verdict The ensemble meta-encoder idea is a reasonable extension of existing meta-embedding work, but the caching metrics are internally inconsistent and the evaluation doesn't simulate a real cache. read the letter →

arxiv 2507.07061 v1 pith:DW5KCOMA submitted 2025-07-08 cs.LG

Shervin Ghaffari , Zohre Bahranifard , Mohammad Akbari This is my paper · ORCID

classification cs.LG

keywords semanticcachingensembleembeddingsmeta-encoderLLMinferencecostcontrastivelearningduplicatequerydetectionQuoraQuestionPairscachehitratio

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 claims that semantic caching in LLM-based systems is held back by relying on one embedding model, and that fusing two complementary embedding models through a small trained meta-encoder gives a more reliable signal for deciding when two queries are the same question. If correct, an LLM service could serve repeated or paraphrased queries from a cache instead of running expensive inference, lowering both latency and token costs. The paper reports a 92% cache hit ratio on semantically equivalent query pairs and 85% accuracy at rejecting non-equivalent pairs, with average gains of 10.3 percentage points in hit ratio and 7.5 points in miss rejection over the strongest single model. These numbers are measured on the Quora Question Pairs benchmark, with a binary equivalence label standing in for the cache-hit decision.

What carries the argument

The load-bearing mechanism is the meta-encoder: a multi-layer neural network (three hidden layers with residual connections, batch normalization, LeakyReLU, dropout, then an L2-normalized projection) that takes the concatenated embeddings of two low-correlated sentence-transformer models and learns to fuse them under a contrastive loss so that duplicate question pairs are pulled together and non-duplicates are pushed apart. It is paired with a similarity threshold (0.80) chosen on validation by F1 score, so the downstream cache decision is just cosine similarity against the fused embedding of the cached query. The second design choice doing real work is the selection of two base models whose similarity-score vectors are least correlated (0.74), which is what makes the ensemble information complementary rather than redundant.

What would settle it

Run the trained encoder on a real LLM query log where users judge whether a cached answer actually answers the new query: the claim collapses if hits drop to single-model levels or if a large share of served cached responses are judged wrong.

Watch

Extended reading notes

Core claim

The central claim is that an ensemble of embedding models, combined by a trainable meta-encoder, captures semantic equivalence between user queries more accurately than any single embedding model, and that this directly improves the hit/miss decisions of a semantic cache. The authors select two base models with the lowest pairwise correlation among five candidates, concatenate their normalized embeddings, and train a three-layer residual encoder with contrastive loss to produce a fused unit-length embedding. Cosine similarity on the fused embedding, with an optimized threshold of 0.80, decides cache hit or miss. On the QQP test split the fused encoder reaches 86% classification accuracy, 92% hit ratio on duplicate pairs, and 85% correct rejection of non-duplicates; at the system level this translates to roughly 20% token savings and a measured drop from 2.7 to 0.3 seconds per request in a populated 1,000-query cache.

Load-bearing premise

The load-bearing premise is that Quora's duplicate-question labels capture what 'same query' means in a real semantic cache, so a fusion encoder trained on those labels will make correct hit/miss decisions on actual user query streams.

Editorial extensions

If this is right

A semantic cache using the fused encoder will serve paraphrased equivalents without rerunning the LLM, which the paper measures as a drop from 2.7 to 0.3 seconds per request in a populated 1,000-query cache.
Token costs fall by roughly 20% on duplicate workloads, because both prompt and completion tokens are skipped when a response is served from cache.
The ensemble is an upgrade to the representation module of an existing semantic cache: storage, eviction, and response-generation policies remain unchanged.
The average 10.3 percentage-point hit-ratio gain and 7.5 percentage-point miss-rejection gain over the best single model are consistent across five dataset sizes, indicating the benefit is not tied to one scale.
The 0.80 similarity threshold balances precision and recall on validation, keeping both false hits and false misses low at the reported operating point.

Reading between the lines

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

Extension: if the paper's mechanism is right, the same recipe can be retargeted to code generation, chat, or summarization by training the small fusion head on task-specific equivalence pairs; the paper only demonstrates question pairs, so this is an extrapolation.
Extension: the fixed 0.80 threshold is a system parameter, not a law of the method; a live deployment would want to set it by the relative cost of serving a wrong cached answer versus paying for an extra LLM call, which the paper does not quantify.
Extension: the low-correlation selection step suggests a cheap screening procedure for any pool of candidate encoders: compute similarity-score correlations, pick the two most complementary models, and train only the fusion layer.

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

Quick take: the ensemble meta-encoder idea is a reasonable extension of existing meta-embedding work to semantic caching, and the classification numbers are plausible, but the headline caching metrics (92% hit, 85% miss, 20% token savings) don't cohere with the paper's own Table 1 or the workload described. I'd want the authors to reconcile those numbers before trusting any of the caching claims.

What's actually new: combining two sentence encoders with a trained contrastive meta-encoder for semantic cache lookups is a straightforward application of Kiela et al.'s meta-embeddings; the paper is honest about that lineage. The QQP evaluation setup is standard: stratified split, threshold selection on validation, FAISS flat search. The classification report in Table 1 (86% accuracy, balanced precision/recall) is credible and shows the encoder does learn something beyond simple concatenation.

Where it falls apart: the caching-specific metrics. Under the paper's own definitions, hit ratio for duplicate queries should equal duplicate recall (0.88 at threshold 0.80), and miss ratio should equal non-duplicate recall (0.83). The reported 92.0% hit and 85.1% miss are different from those numbers, and no protocol is given to explain the gap. The token savings claim is worse: with 1,000 duplicate pairs and the cache pre-populated, if 92% of requests are hits, the savings should be close to 92% of tokens, not 20%. The paper never explains what the denominator includes. On top of that, the evaluation is QQP pair classification, not a real cache with request streams, capacity, eviction, and varied duplication rates. The authors do include an eviction-policy section, but it's qualitative and not connected to the main benchmarks. Base model selection uses a 100k-pair sample from QQP with no statement that the test portion was excluded, which is a potential test-information leak.

Bottom line: the method is plausible but the evidence as reported does not support the headline claims. This is the kind of paper where the internal arithmetic needs to be fixed and the evaluation needs to actually simulate a cache before it makes claims about caching performance. I wouldn't cite it and I don't think it's ready for a serious referee without major revision. If the authors clean up the metric definitions and rerun a proper cache simulation, there might be a small useful contribution here.

Referee Report

4 major / 5 minor

Summary. The paper proposes an ensemble embedding approach for semantic caching in LLM-based systems. It selects two sentence-embedding models with low pairwise correlation, concatenates their embeddings, and trains a meta-encoder with a contrastive loss on the Quora Question Pairs (QQP) dataset. The authors report cache hit ratios, miss ratios, token savings, and response times, claiming a 92% cache hit ratio for semantically equivalent queries and 85% accuracy for rejecting non-equivalent queries. The central idea is that combining complementary embedding spaces improves the duplicate-detection decision at the heart of semantic caching.

Significance. If the reported results were reliable, the ensemble meta-encoder would be a practical contribution to semantic caching, a topic of current interest for reducing LLM inference cost and latency. The use of QQP as a proxy for cache-hit decisions is reasonable, the contrastive training setup is understandable, and the general direction of combining multiple embedders is plausible. However, the paper's headline numbers are internally inconsistent with its own classification table and token-saving calculation, and the experimental protocol for the caching metrics is not fully specified. As written, the evidence does not establish the central claims, despite the potential interest of the approach.

major comments (4)

[Section 5.2.1, Table 1 vs. Sections 5.2.3 and Abstract] The reported 92.0% cache hit ratio for semantically equivalent queries conflicts with Table 1, where the duplicate class recall is 0.88. Under Eq. (1) and the protocol described in Section 5.2.4, a cache populated with true duplicate first questions and queried with their paired duplicates should yield a hit ratio equal to the duplicate recall, i.e., approximately 88%, not 92.0%. Similarly, the 85.1% miss ratio for non-equivalent queries does not match the non-duplicate recall of 0.83 in Table 1, and the 86% overall accuracy is not the average of 92.0% and 85.1%. No experimental detail reconciles these numbers, so the central quantitative claim in the abstract is not supported by the paper's own data.
[Section 5.2.5, Eq. (2)] The approximately 20% token savings claim is inconsistent with the reported hit ratio in the same experimental setting. Section 5.2.5 states that all 1,000 paired questions in Section 5.2.4 are duplicates and that on cache hits both input and output tokens are completely avoided. If 92% of these requests are served from cache, Eq. (2) would give token savings close to 92%, not approximately 20%. The manuscript gives no token-mix or partial-cache accounting that would explain the 20% figure, so this efficiency result is unexplained.
[Section 5.1.1, Base Model Selection] The correlation analysis used to select the two base models is performed on a balanced sample of 100,000 query pairs from QQP, but the paper does not state that this sample is restricted to the training partition defined in Section 4.4. Since the data is split into 70% training, 15% validation, and 15% test, sampling from the full dataset could allow test-pair information to influence model selection. The authors should either confirm that the test partition was excluded or rerun the selection on the training set only; otherwise the reported comparisons against single models are not clean.
[Sections 4.5.3 and 5.2.3, Cache Hit and Miss Ratio Protocol] The experimental protocol for the cache hit and miss ratios in Table 2 and Figures 5 and 6 is not described. Eq. (1) defines hit ratio as cache hits divided by total requests, but it is not stated how the cache is populated, in what order queries are issued, what fraction of requests are duplicates, or whether the 15,000-pair test set is used as-is or reorganized into a request stream. Without this protocol, the reported 92.0% hit ratio and 85.1% miss ratio cannot be reproduced or interpreted, especially because a balanced QQP test set contains only 50% duplicate pairs.

minor comments (5)

[Section 4.3.1] The architecture description mentions one residual connection between Layer 1 and Layer 2, while Section 4.4 refers to 'three residual blocks'; please align the terminology.
[Section 5.2.4] The response-time experiment uses 1,000 randomly selected paired queries, but no random seed, variance, or confidence interval is reported; a single point estimate is not sufficient to support the 2.7-to-0.3 second claim.
[Table 2] The relative improvement percentages are presented without the underlying absolute hit and miss ratios for the baselines, so the reader cannot verify the reported +8.9% to +12.4% improvements.
[References] Several references appear unrelated to the surrounding claims (e.g., [27], [28], [30], [31]); these should be replaced with appropriate citations or removed.
[Section 4.5.2] The threshold of 0.80 is selected on the validation set by maximizing F1, but no sensitivity analysis around this threshold is provided; given that caching performance depends directly on the threshold, a small threshold variation study would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the ensemble is trained and evaluated on separate QQP splits; the reported inconsistencies are correctness or external-validity concerns, not circular reductions.

full rationale

The paper's central claim is an empirical evaluation, not a derivation that reduces to its inputs. The meta-encoder is trained on QQP training pairs and evaluated on a held-out test set (Section 4.4: 'We perform a stratified split of the dataset into training (70%), validation (15%), and test (15%)'), and the similarity threshold is fit on the validation set (Section 4.5.2: 'Based on validation set performance, we select an optimal threshold of 0.80'). Reporting test-set hit/miss ratios under this protocol is standard supervised evaluation. No load-bearing self-citation appears: the cited meta-embedding works are external, and the ensemble approach is trained on QQP labels rather than imported by ansatz. The QQP-to-caching mapping in Section 3 is an external-validity assumption, not a circular step. The main problems are numerical inconsistencies and possible leakage, not circularity: the 92% hit ratio in Table 2 is not reconciled with the 88% duplicate recall in Table 1; the ~20% token savings is not reconciled with a 92% hit ratio under Equation (2); and base-model selection uses a 100,000-pair QQP sample (Section 5.1.1) without an explicit statement that it excludes the test split. These are correctness or reporting concerns, and the paper does not exhibit a reduction that makes any reported prediction equivalent to a fitted parameter by construction.

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

The paper introduces no new physical or conceptual entities. Its free parameters are the threshold and architecture choices; the QQP proxy assumption is the main domain assumption.

free parameters (4)

Similarity threshold = 0.80
Selected on validation set by maximizing F1 score, then used for all reported metrics.
Meta-encoder hidden dimensions = 1024, 512, 384
Architecture hyperparameters chosen by the authors; no sensitivity analysis is provided.
Contrastive loss margin = not reported
The loss in Section 4.3.2 uses a predefined margin, but the value is not stated.
Number of base models = 2
Two models are selected by lowest pairwise correlation; this choice is not derived from first principles.

assumptions (4)

domain assumption QQP duplicate labels are a valid proxy for whether a query should be a cache hit or miss.
Section 3 equates label 1 with cache hits and label 0 with misses, but real caching involves request streams, cache content, and threshold trade-offs.
domain assumption Sentence transformer embeddings capture the semantic dimensions relevant to caching.
Section 4.1 assumes pretrained embeddings are sufficient for similarity search.
ad hoc to paper The two models with lowest pairwise correlation provide optimal complementarity.
Section 5.1.2 selects these two; this is a choice not derived from first principles.
standard math Contrastive learning with cosine distance is a valid training objective.
Section 4.3.2 applies a standard contrastive loss to train the meta-encoder.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Ensemble Embedding Approach for Improving Semantic Caching Performance in LLM-based Systems." pith.science (2026). https://pith.science/paper/DW5KCOMA

@misc{pith2026250707061,
  author       = {Pith},
  title        = {Pith review of: An Ensemble Embedding Approach for Improving Semantic Caching Performance in LLM-based Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DW5KCOMA}},
  note         = {Machine review of arXiv:2507.07061}
}

read the original abstract

Semantic caching enhances the efficiency of large language model (LLM) systems by identifying semantically similar queries, storing responses once, and serving them for subsequent equivalent requests. However, existing semantic caching frameworks rely on single embedding models for query representation, which limits their ability to capture the diverse semantic relationships present in real-world query distributions. This paper presents an ensemble embedding approach that combines multiple embedding models through a trained meta-encoder to improve semantic similarity detection in LLM caching systems. We evaluate our method using the Quora Question Pairs (QQP) dataset, measuring cache hit ratios, cache miss ratios, token savings, and response times. Our ensemble approach achieves a 92\% cache hit ratio for semantically equivalent queries while maintaining an 85\% accuracy in correctly rejecting non-equivalent queries as cache misses. These results demonstrate that ensemble embedding methods significantly outperform single-model approaches in distinguishing between semantically similar and dissimilar queries, leading to more effective caching performance and reduced computational overhead in LLM-based systems.

Figures

Figures reproduced from arXiv: 2507.07061 by the authors.

**Figure 2.** Architecture of the proposed trainable encoder-based meta-encoder. The model dynamically integrates concatenated embeddings from multiple base models through a multi-layer neural network featuring residual connections, dimensionality reduction, and L2 normalization to produce task-aligned semantic representations [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Cosine similarity distribution on the test set. The vertical dashed line indicates the selected threshold of 0.80. of generating new ones. This includes both prompt and completion tokens. Token Saving Ratio = Tokens Served by Cache Total Tokens Processed × 100 (2) Response Time: This metric evaluates the average time required to serve a request. Lower response times indicate faster, more efficient caching performanc… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Correlation matrix of base embedding models showing pairwise cosine similarity correlations across the evaluated models. 5.1.3 Baseline Comparisons Our experimental evaluation includes several baseline approaches to provide comprehensive performance context: • Individ…

**Figure 5.** Figure 5: Performance comparison of the proposed trainable encoder against individual embedding models [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Detailed comparison of our encoder against fusion strategies. 5.2.4 Response Time As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 8.** Figure 8: illustrates the impact of cache size—expressed as a percentage of 10,000 total queries—on the hit ratio for LRU and LFU algorithms [29]. LFU generally outperforms LRU across different query scales, though the optimal choice depends on request patterns and embedding qu…

**Figure 7.** Figure 7: Comparison of response times, demonstrating a reduction from 2.7 seconds to 0.3 seconds using semantic caching with the encoder-based ensemble method. 5.2.5 Token Savings The system achieves approximately 20% token savings, calculated using the same 1,000 duplicate pa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 26 canonical work pages

[1]

Brown, B

T. Brown, B. Mann, N. Ryder, et al., Language models are few-shot learners, in: Advances in Neural Infor- mation Processing Systems (NeurIPS), V ol. 33, 2020, pp. 1877–1901

work page 2020
[2]

Huang, Y

Z. Huang, Y . Liang, H. Liu, et al., A survey on efficient inference for large language models, arXiv preprint arXiv:2404.14294 (2024)

arXiv 2024
[3]

Detecting critical treatment effect bias in small subgroups

Z. Wang, et al., Privacy-aware semantic cache for large language models, arXiv preprint arXiv:2404.18905 (2024)

work page Pith review arXiv 2024
[4]

Reimers, I

N. Reimers, I. Gurevych, Sentence-bert: Sentence em- beddings using siamese bert-networks, in: EMNLP- IJCNLP, 2019, pp. 3982–3992

work page 2019
[5]

Deerwester, S

S. Deerwester, S. T. Dumais, G. W. Furnas, T. K. Lan- dauer, R. Harshman, Indexing by latent semantic anal- ysis, Journal of the American Society for Information Science 41 (6) (1990) 391–407

work page 1990
[6]

J. Tu, Y . Qu, C. Li, H. Chen, Enhancing semantic text similarity with functional semantic knowledge, Jour- nal of Informetrics 18 (1) (2024) 101482

work page 2024
[7]

Y . Bang, S. Cahyawijaya, N. Lee, et al., Gptcache: An open-source semantic cache for llm applications, in: EMNLP System Demonstrations, 2023, pp. 286–295

work page 2023
[8]

Kiela, C

D. Kiela, C. Wang, K. Cho, Sentence meta- embeddings for unsupervised semantic textual simi- larity, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 6289–6294

work page 2020

Show all 31 references

[9]

W. Yin, H. Sch¨utze, Learning word meta-embeddings, in: ACL (Long Papers), V ol. 1, 2016, pp. 1351–1360

2016
[10]

S. Iyer, N. Dandekar, K. Csernai, First quora dataset release: Question pairs, quora Engineering Blog. (2017)

2017
[11]

Y . Sun, C. Wang, J. Huang, T. Zhang, X. He, Seman- tic approaches for query expansion: taxonomy, chal- lenges and opportunities in big data era, Journal of Big Data 12 (1) (2025) 45

2025
[12]

J. Li, C. Xu, F. Wang, et al., Scalm: Towards se- mantic caching for automated chat services with large language models, arXiv preprint arXiv:2406.00025 (2024)

2024 arXiv
[13]

Iyengar, A

A. Iyengar, A. Kundu, R. Kompella, S. Krish- namurthy, A generative caching system for large language models, arXiv preprint arXiv:2503.17603 (2025)

2025 arXiv
[14]

Akg ¨ul, S

S. Akg ¨ul, S. M. E. Din, H. Ferhatosmanoglu, En- hancing semantic code search with deep graph match- ing and retrieval, Information Retrieval Journal 27 (2) (2024) 504–529

2024
[15]

Chua, Y .-Y

B. Chua, Y .-Y . Ahn, S. Kojaku, Comparing seman- tic representation methods for keyword analysis in scientometrics, Journal of Informetrics 18 (3) (2024) 101552

2024
[16]

Douze, A

M. Douze, A. Guzhva, C. Deng, et al., The faiss li- brary, arXiv preprint arXiv:2401.08281 (2024)

2024 arXiv
[17]

T. Gao, X. Yao, D. Chen, Simcse: Simple contrastive learning of sentence embeddings, in: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021, pp. 6894–6910

2021
[18]

Z. Li, Y . Yang, R. E. Schapire, Consistent penalizing field loss for zero-shot image retrieval with seman- tic embedding, Information Retrieval Journal 26 (3) (2023) 1–28

2023
[19]

K. S. Jones, A statistical interpretation of term speci- ficity and its application in retrieval, Journal of Docu- mentation 28 (1) (1972) 11–21

1972
[20]

Zhang, J

Z. Zhang, J. Wang, J. Xu, K. Wei, Y . Liu, Textcg: A text classification framework based on information fu- sion of pre-trained language model and graph neural network, Information Sciences 685 (2025) 122413

2025
[21]

X. Wang, Y . Qu, C. Li, H. Chen, Word embedding for social sciences: an interdisciplinary survey, Sciento- metrics 129 (8) (2024) 4371–4420

2024
[22]

Robertson, H

S. Robertson, H. Zaragoza, The probabilistic rele- vance framework: Bm25 and beyond, Foundations and Trends in Information Retrieval 3 (4) (2009) 333– 389. 9

2009
[23]

G. Mai, K. Janowicz, Y . Hu, S. Gao, Improving infor- mation retrieval through correspondence analysis in- stead of latent semantic analysis, Journal of Machine Learning Research 25 (1) (2024) 1–42

2024
[24]

K. Song, X. Tan, T. Qin, et al., Mpnet: Masked and permuted pre-training for language understanding, in: Advances in Neural Information Processing Systems (NeurIPS), V ol. 33, 2020, pp. 16857–16867

2020
[25]

Y . Liu, M. Ott, N. Goyal, et al., Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019)

2019 arXiv
[26]

Q. Liu, J. Lu, F. Liu, A. Porter, H. Chen, G. Zhang, Academic collaboration recommendation based on graph neural network with application to covid-19 re- search, Journal of Informetrics 18 (2) (2024) 101401

2024
[27]

M. A. Taiye, A. H. A. Hashim, N. Salim, Algorithmic recommendations enabling and constraining informa- tion practices among young people, Journal of Docu- mentation 80 (1) (2024) 1–23

2024
[28]

Callon, J

M. Callon, J. Law, A. Rip, Practice architectures for bridging the semantic gap in museum information sys- tems, Journal of Documentation 80 (6) (2024) 1626– 1644

2024
[29]

L. A. Belady, A study of replacement algorithms for a virtual-storage computer, IBM Systems Journal 5 (2) (1966) 78–101

1966
[30]

Robbins, G

C. Robbins, G. Boesch, S. Forrest, The artificial intel- ligence patent dataset (aipd): Tracing innovation in ai, Scientometrics 129 (4) (2024) 2145–2180

2024
[31]

C. Chen, Y . Chen, M. Horowitz, H. Hou, Z. Hu, K. Liu, Topical evolution and thematic progression of research frontiers in information science, Knowledge Organization 52 (1) (2025) 1–25. 10

2025

Pith tools

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

[1] [1]

Brown, B

T. Brown, B. Mann, N. Ryder, et al., Language models are few-shot learners, in: Advances in Neural Infor- mation Processing Systems (NeurIPS), V ol. 33, 2020, pp. 1877–1901

work page 2020

[2] [2]

Huang, Y

Z. Huang, Y . Liang, H. Liu, et al., A survey on efficient inference for large language models, arXiv preprint arXiv:2404.14294 (2024)

arXiv 2024

[3] [3]

Detecting critical treatment effect bias in small subgroups

Z. Wang, et al., Privacy-aware semantic cache for large language models, arXiv preprint arXiv:2404.18905 (2024)

work page Pith review arXiv 2024

[4] [4]

Reimers, I

N. Reimers, I. Gurevych, Sentence-bert: Sentence em- beddings using siamese bert-networks, in: EMNLP- IJCNLP, 2019, pp. 3982–3992

work page 2019

[5] [5]

Deerwester, S

S. Deerwester, S. T. Dumais, G. W. Furnas, T. K. Lan- dauer, R. Harshman, Indexing by latent semantic anal- ysis, Journal of the American Society for Information Science 41 (6) (1990) 391–407

work page 1990

[6] [6]

J. Tu, Y . Qu, C. Li, H. Chen, Enhancing semantic text similarity with functional semantic knowledge, Jour- nal of Informetrics 18 (1) (2024) 101482

work page 2024

[7] [7]

Y . Bang, S. Cahyawijaya, N. Lee, et al., Gptcache: An open-source semantic cache for llm applications, in: EMNLP System Demonstrations, 2023, pp. 286–295

work page 2023

[8] [8]

Kiela, C

D. Kiela, C. Wang, K. Cho, Sentence meta- embeddings for unsupervised semantic textual simi- larity, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 6289–6294

work page 2020