Pith. sign in

REVIEW 3 major objections 6 minor 28 references

THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics

T0 review · 3 major / 6 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 Fine-tuning embeddings on theme–stock pairs and then on recent returns produces embeddings that retrieve thematically aligned stocks more accurately than general-purpose LLMs and improve backtested portfolio Sharpe ratios.

desk verdict Retrieval results are genuinely strong; the portfolio claim hangs on an unstated temporal split that needs a clean answer before Table 2 means anything. read the letter →

arxiv 2508.16936 v2 pith:56TZ25K3 submitted 2025-08-23 q-fin.PM cs.AIcs.CLcs.IR

Hoyoung Lee , Wonbin Ahn , Suhwan Park , Jaehoon Lee , Minjae Kim , Sungdong Yoo , Taeyoon Lim , Woohyung Lim
show 1 more author
Yongjae Lee
This is my paper
classification q-fin.PMcs.AIcs.CLcs.IR
keywords thematicinvestinghierarchicalcontrastivelearningstockembeddingsinformationretrievalportfolioconstructiontemporaldynamicsLoRARepresentationSet
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

Thematic investing requires finding stocks that belong to a cross-sectoral theme and tracking how that membership shifts over time. The paper argues that general-purpose LLM embeddings fail at this because they encode general text semantics rather than the financial meaning of a company's business. THEME is a two-stage fine-tuning procedure: first, it uses contrastive learning to pull a stock's textual profile toward the text description of any theme it belongs to; second, a lightweight adapter fuses that semantic embedding with recent daily returns, trained to put higher-forward-return stocks closer to the theme anchor. On retrieval benchmarks, the tuned embeddings outscore GPT-4.1 and several domain-specific models; in a one-year backtest they raise Sharpe ratios, for example for gte-Qwen2-7B from 0.50 to 0.76 at top-3. The paper's practical claim is that theme-aware embeddings, not bigger general models, are the path to dynamic thematic portfolios.

What carries the argument

The central object is the hierarchy-aware, two-stage contrastive framework: (1) a LoRA-tuned embedding model with theme-anchored contrastive loss (Eq. 2); (2) a two-layer temporal adapter that fuses the semantic embedding with the last 60 days of daily returns and is trained by triplet loss (Eq. 4) where positive/negative are decided by 14-day forward returns. The Thematic Representation Set (TRS)—expanded from 1,153 ETFs with sector taxonomies and news-derived themes to about 200 themes—provides the supervision. Retrieval is cosine similarity between the theme/query embedding and precomputed fusion embeddings.

What would settle it

Re-run the Stage 2 training with a documented cutoff (e.g., all training return windows end on or before April 22, 2024) and verify that the top-3 Sharpe for gte-Qwen2-7B-instruct stays near 0.7592 rather than collapsing; alternatively, inspect the rolling-window generation code to confirm each training sample's 14-day forward window ends before the test start.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the two sources of supervision—the hierarchical theme–stock relation and the relative forward-return ranking within a theme—are jointly sufficient to turn a frozen general embedding model into a financially meaningful stock embedding space. The first stage (thematic alignment) uses LoRA-adapted embeddings with a contrastive objective that anchors on the theme description, pulling constituent stocks closer and pushing non-constituents away. The second stage (temporal refinement) trains a two-layer adapter on triplet loss with the theme anchor and positive/negative stocks defined by whether their 14-day forward return is higher. The authors claim

Load-bearing premise

The two years of historical U.S. market data used to train the return-refinement stage must end before the April 2024-to-April 2025 portfolio test window; if any forward-return label overlaps the evaluation period, the reported Sharpe and cumulative-return gains are the model fitting labels it was trained on.

Editorial extensions

If this is right

  • If true, thematic screening can be improved without swapping the underlying large model: applying THEME's LoRA and adapter stages lets even a 33M-parameter model surpass vanilla 7B models and proprietary LLMs.
  • The temporal adapter shows a way to make embeddings sensitive to short-term return patterns while preserving long-term thematic meaning, a recipe that could transfer to other dynamic retrieval tasks where relevance shifts over time.
  • The portfolio backtests suggest a momentum-style signal can be fused into a semantic representation to lift Sharpe ratios, with the caveat that the temporal split between training and evaluation must be clean.
  • Because THEME is model-agnostic, gains should transfer across base embedding families; Table 1 shows consistent improvements from 33M- to 7B-scale models.
  • TRS enables coverage of niche themes beyond existing ETFs, so users can define their own themes and receive a ranked stock list without waiting for an ETF product to exist.

Reading between the lines

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

  • A natural extension is to test the same two-stage recipe on other asset classes or theme spaces (e.g., ESG, supply chains); the framework only needs a theme–member list and price history.
  • The paper's ablation compares training anchors, not query anchors; a reader could test whether a set of representative stocks as the query outperforms a textual description at inference time.
  • Because the temporal adapter is trained on 14-day forward returns, it likely behaves like a short-term momentum filter; a reader could check whether the adapter's rankings correlate with a simple price-momentum factor.
  • One manuscript note: reference [17] lists a placeholder URL (example.org), so the provenance of that cited prior work is not verifiable from the paper itself.
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 THEME, a two-stage hierarchical contrastive learning framework that fine-tunes general-purpose text embedding models for thematic stock retrieval and portfolio construction. Stage 1 aligns theme descriptions with constituent stock profiles using a LoRA-adapted backbone and a contrastive loss. Stage 2 trains a lightweight adapter that fuses the Stage-1 stock embedding with the stock's past L=60 days of returns, using a triplet loss whose positive/negative labels are based on forward returns over H=14 days. The authors evaluate retrieval on held-out thematic ETFs (Experiment 1), construct top-K equal-weighted portfolios over April 23, 2024 to April 29, 2025 (Experiment 2), and ablate the anchor strategy and training dataset (Experiment 3). The central claim is that THEME substantially improves both thematic retrieval accuracy and short-term portfolio profitability over vanilla embeddings and strong LLM baselines.

Significance. If the results hold, the contribution is practically valuable: a model-agnostic fine-tuning recipe that turns generic embeddings into finance-specific thematic retrieval systems, with consistent and large retrieval gains across many backbones (e.g., Linq-Embed-Mistral HR@3 from 0.5155 to 0.8196 in Table 1). The construction of the Thematic Representation Set (TRS) is also a useful resource. However, the portfolio claim, which is a headline contribution, rests on a temporal-separation assumption that the manuscript never states. The paper reports consistent retrieval improvements and informative ablations, but the backtest evidence for 'strong return potential' is not yet convincing because the Stage-2 training window may overlap the evaluation period and because the reported Sharpe ratios are not accompanied by any uncertainty quantification. The retrieval claim is more solid; the portfolio claim needs additional evidence.

major comments (3)
  1. [Section 3.2 / Eq. (4); Section 4.1; Section 4.3] The Stage-2 adapter is trained with triplet labels defined by forward returns over H=14 days (Eq. 4), using 'two years of historical U.S. market data' and a rolling window (Section 4.1). The portfolio backtest in Section 4.3 measures 14-day forward returns over April 23, 2024 to April 29, 2025. The paper never states the end date of the Stage-2 training data or otherwise demonstrates that training windows end before April 22, 2024. If any training window overlaps the backtest, the Sharpe-ratio and cumulative-return improvements in Table 2 (e.g., gte-Qwen2-7B-instruct SR@3 from 0.5014 to 0.7592) are the result of fitting the adapter to the same forward-return labels used in evaluation, not out-of-sample predictive skill. This is the load-bearing point for the portfolio claim. Please report the exact temporal split, retrain with an expanding-window or rolling-window scheme that strictly en
  2. [Section 4.3 / Table 2] The backtest methodology is ambiguous. The text says 'For each window' portfolios are held for 'the subsequent 14 trading days' and 'chaining the returns across all windows forms a continuous daily return series.' If windows start every day, the daily returns are heavily overlapping and serially correlated; the reported Sharpe ratio computed as mean/std with sqrt(252) is then not a valid annualized Sharpe without a correction (e.g., Newey-West or block bootstrap). If windows are non-overlapping 14-day periods, the return series is not daily and the formula is misapplied. The paper also does not report transaction costs or turnover. The differences in Table 2 are not accompanied by confidence intervals or significance tests; with a one-year sample and few independent holding periods, several gains (e.g., GritLM-7B SR@3 0.5744 vs. 0.5952) may be within noise. Please clarify the rebalancing
  3. [Section 4.2 / Table 1] Experiment 1 reports retrieval performance but does not state the as-of date for the stock textual profiles or, more importantly, for the past-return features used by the Stage-2 adapter. Since the retrieval embeddings are fusion embeddings that depend on recent returns, the evaluation date matters: if the returns used to compute h'_j during retrieval evaluation come from the same period used to train the temporal adapter, the retrieval comparison is not cleanly out-of-sample. Please specify the point-in-time cutoff for all features used in the retrieval experiment and confirm that it precedes any labels used for training.
minor comments (6)
  1. [Table 1] The 'Ours' columns are blank for the domain-specific and LLM baselines; this is likely intentional, but the table would be clearer with an explicit em dash or a note that THEME was not applied to those models.
  2. [Tables 3 and 4] The ablation tables report only the delta (improvement) over the counterpart setting, not the absolute P@k values for each configuration. Without the baselines, it is impossible to judge the practical size of the effect. Please include absolute numbers for both settings.
  3. [Section 4.3] The note under Table 2 gives the 'average performance of real-world Thematic ETFs' but does not explain how this average was computed (e.g., equal-weighted across all 1,153 ETFs, rebalanced how, over what period). This is needed for a fair benchmark.
  4. [Eq. (4)] The notation [x]_+ is used without definition; please state that it denotes max(x,0).
  5. [Section 3.2] Equation (3) writes the adapter input as A_phi(h_j, r_j), but the architecture for combining a vector embedding with a sequence of L daily returns is not described. Please specify whether r_j is fed through a separate MLP, concatenated, or otherwise encoded.
  6. [References] Several references are incomplete or contain placeholder URLs: [17], [18], and [20] cite 'example.org' and 'To appear.' These should be completed before publication. Also, [16] is cited for GPT-4.1 but the reference is the GPT-4 Technical Report; please align the citation.

Circularity Check

0 steps flagged · score 1.0 of 10

No demonstrated circularity; retrieval evaluation is externally benchmarked and held-out, while the only flagged issue is an unstated temporal split in the portfolio test, which is a missing disclosure rather than a shown reduction.

full rationale

We inspected the derivation chain for definitional or self-citation circularity. Stage 1 (Section 3.1) is a standard LoRA contrastive alignment between theme descriptions and constituent stocks, evaluated on a held-out ETF split (678/97/194, Section 4.1) against external baselines (GPT-4.1, Gemini-2.5, and general-purpose embedding models). The retrieval gains in Table 1 are therefore benchmarked against independent reference points and are not the model predicting its own training labels. The self-citations in Related Works (SimStock [6], Lee et al. [8]) are contextual and not load-bearing; no uniqueness theorem or forced-choice argument is imported from the authors' prior work. The only serious validity concern is temporal: Section 4.1 says the Stage 2 adapter is trained on 'two years of historical U.S. market data' with triplet labels defined by forward returns over H=14 days (Eq. 4), and Experiment 2 (Section 4.3) evaluates 14-day forward returns of top-K stocks over April 23, 2024 – April 29, 2025. The paper never states whether the two-year training window ends before that test window. If training and test windows overlap, the Sharpe and cumulative-return gains in Table 2 would be in-sample fits of the triplet objective rather than out-of-sample predictions. But this is conditional on an unstated fact; the paper's text does not exhibit the overlap, so under the hard rules it is a missing temporal-split disclosure and a correctness/reproducibility risk, not a demonstrated circular reduction. We therefore keep the circularity score low.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on ETF-labeled thematic ground truth, a 14-day forward-return horizon used both for training labels and evaluation, and unstated hyperparameters. The TRS dataset is not released, so these assumptions cannot be independently tested.

free parameters (5)
  • temperature tau (Stage 1 contrastive loss) = not reported
    Hyperparameter in Eq. (2) controlling separation of theme-stock pairs; no value or search range given.
  • triplet margin m (Stage 2) = not reported
    Margin in Eq. (4); chosen by hand, not stated.
  • lookback period L = 60 trading days
    Section 4.1 sets L=60 for return inputs; arbitrary modeling choice.
  • forward return horizon H = 14 trading days
    Section 4.1 sets H=14 for both training labels and portfolio evaluation; the same horizon is used to measure the result, creating potential circularity.
  • LoRA rank and adapter dimensions / learning schedule = not reported
    Training details for the LoRA adapter and the 2-layer fusion adapter are not specified; replication would require guessing.
assumptions (4)
  • domain assumption ETF constituent holdings are a valid ground-truth label for thematic relevance.
    Used to define positive pairs in Stage 1 (Section 3.1) and to build TRS (Section 3); if ETF lists are biased or lag thematic shifts, the learned relevance is biased.
  • domain assumption Forward return over H=14 days is an appropriate proxy for stock investment suitability.
    Stage 2 assigns positive and negative labels by forward return (Section 3.2) and the portfolio evaluation uses the same horizon (Section 4.3). No evidence is given that a 14-day horizon corresponds to investor utility.
  • domain assumption General-purpose pretrained embeddings are a useful substrate after LoRA fine-tuning.
    The method freezes the backbone and learns only LoRA adapters; if the base model's text representations lack financial signal, the contrastive step cannot recover it. This is assumed through Eq. (1).
  • standard math Cosine similarity is an appropriate ranking function for theme-stock relevance.
    Used in both losses and inference (Section 3); standard but unstated alternative.
invented entities (2)
  • Thematic Representation Set (TRS)
    purpose: Expanded training corpus of themes with stock memberships from ETFs, sectoral taxonomies, and news-derived themes (Section 3).
    Dataset is not released, so its composition and quality cannot be independently checked; retrieval and portfolio results depend on it.
  • Fusion embedding space after Stage 2 adapter
    purpose: Joint semantic-temporal stock representation h'_j from Eq. (3).
    The adapter's output is the object evaluated in retrieval and portfolio tests, but no external benchmark of the representation exists.

how reviews work

0 comments
Cite this review

Pith. "Pith review of THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics." pith.science (2026). https://pith.science/paper/56TZ25K3

@misc{pith2026250816936,
  author       = {Pith},
  title        = {Pith review of: THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/56TZ25K3}},
  note         = {Machine review of arXiv:2508.16936}
}
read the original abstract

Thematic investing, which aims to construct portfolios aligned with structural trends, remains a challenging endeavor due to overlapping sector boundaries and evolving market dynamics. A promising direction is to build semantic representations of investment themes from textual data. However, despite their power, general-purpose LLM embedding models are not well-suited to capture the nuanced characteristics of financial assets, since the semantic representation of investment assets may differ fundamentally from that of general financial text. To address this, we introduce THEME, a framework that fine-tunes embeddings using hierarchical contrastive learning. THEME aligns themes and their constituent stocks using their hierarchical relationship, and subsequently refines these embeddings by incorporating stock returns. This process yields representations effective for retrieving thematically aligned assets with strong return potential. Empirical results demonstrate that THEME excels in two key areas. For thematic asset retrieval, it significantly outperforms leading large language models. Furthermore, its constructed portfolios demonstrate compelling performance. By jointly modeling thematic relationships from text and market dynamics from returns, THEME generates stock embeddings specifically tailored for a wide range of practical investment applications.

Figures

Figures reproduced from arXiv: 2508.16936 by the authors.

Figure 1
Figure 1. A t-SNE comparison of stock embeddings before [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of how the THEME system is used in practice and how it works. Given a user-provided thematic de [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the two-stage hierarchical contrastive learning framework. By leveraging the explicit relationship between [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 12 canonical work pages

  1. [1]

    Chanyeol Choi, Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, and Jy yong Sohn. 2024. Linq-Embed-Mistral Technical Report. arXiv:2412.03223 [cs.CL] https://arxiv.org/abs/2412.03223

    arXiv 2024

  2. [2]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guil- laume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)

    arXiv 2019

  3. [3]

    Google Gemini Team. 2025. Gemini (May 25 version). https://ai.google.dev/ gemini-api

    work page 2025

  4. [4]

    Shihao Gu, Bryan Kelly, and Dacheng Xiu. 2020. Empirical Asset Pricing via Machine Learning. The Review of Financial Studies 33, 5 (2020), 2223–2273. https://doi.org/10.1093/rfs/hhaa009

    doi:10.1093/rfs/hhaa009 2020

  5. [5]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR 1, 2 (2022), 3

    2022

  6. [6]

    Yoontae Hwang, Junhyeong Lee, Daham Kim, Seunghwan Noh, Joohwan Hong, and Yongjae Lee. 2023. SimStock : Representation Model for Stock Similarities. In 4th ACM International Conference on AI in Finance . ACM, Brooklyn NY USA, 533–540. https://doi.org/10.1145/3604237.3626888

    work page arXiv 2023

  7. [7]

    Seonmi Kim, Seyoung Kim, Yejin Kim, Junpyo Park, Seongjin Kim, Moolkyeol Kim, Chang Hwan Sung, Joohwan Hong, and Yongjae Lee. 2023. LLMs analyzing the analysts: Do BERT and GPT extract more value from financial analyst reports?. In Proceedings of the Fourth ACM International Conference on AI in Finance . 383– 391

    work page 2023

  8. [8]

    Hoyoung Lee, Junhyuk Seo, Suhwan Park, Junhyeong Lee, Wonbin Ahn, Chanyeol Choi, Alejandro Lopez-Lira, and Yongjae Lee. 2025. Your AI, Not Your View: The Bias of LLMs in Investment Analysis. arXiv preprint arXiv:2507.20957 (2025)

    arXiv 2025

Show all 28 references
  1. [9]

    Yongjae Lee, Jang Ho Kim, Woo Chang Kim, and Frank J Fabozzi. 2024. An Overview of Machine Learning for Portfolio Optimization. Journal of Portfolio Management 51, 2 (2024)

    2024

  2. [10]

    Yongjae Lee, John RJ Thompson, Jang Ho Kim, Woo Chang Kim, and Francesco A Fabozzi. 2023. An Overview of Machine Learning for Asset Management.Journal of Portfolio Management 49, 9 (2023)

    2023

  3. [11]

    Wei Li, Xin Wang, Yang Liu, and Lei Zhang. 2023. Temporal Knowledge Graphs for Dynamic Financial Event Modeling. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2023), 1234–1244. https: //doi.org/10.1145/3580305.3599456

    2023

  4. [12]

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281 (2023)

    2023 arXiv

  5. [13]

    Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. 2024. SFR-Embedding-Mistral:Enhance Text Retrieval with Transfer Learning. Salesforce AI Research Blog. https://www.salesforce.com/blog/sfr- embedding/

    2024

  6. [14]

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Aman- preet Singh, and Douwe Kiela. 2024. Generative Representational Instruction Tuning. arXiv:2402.09906 [cs.CL]

    2024 arXiv

  7. [15]

    Yuqi Nie, Yaxuan Kong, Xiaowen Dong, John M Mulvey, H Vincent Poor, Qing- song Wen, and Stefan Zohren. 2024. A survey of large language models for financial applications: Progress, prospects and challenges. arXiv preprint arXiv:2406.11903 (2024)

    2024 arXiv

  8. [16]

    OpenAI. 2023. GPT-4 Technical Report. https://openai.com/research/gpt-4

    2023

  9. [17]

    Ananya Rao, John Smith, and Hyejin Lee. 2023. Mapping Companies to Long- Term Themes Using Domain-Specific Language Models. Proceedings of the ACL Workshop on Financial NLP (2023). https://example.org/rao2023 To appear

    2023

  10. [18]

    Ryo Takayanagi, Satoshi Kimura, and Wei Zhang. 2024. SETN: Stock Embedding via Text and Network for Thematic Portfolio Construction. In Proceedings of the Web Conference (WWW). https://example.org/takayanagi2024 To appear

    2024

  11. [19]

    Yixuan Tang and Yi Yang. 2025. FinMTEB: Finance Massive Text Embedding Benchmark. arXiv:2502.10990 [cs.CL] https://arxiv.org/abs/2502.10990

    2025 arXiv

  12. [20]

    Dimitrios Vamvourellis, Neha Gupta, and Tao Li. 2023. Learning Company Similarities from SEC Filings Using Transformer-Based Embeddings. arXiv preprint arXiv:2304.12345 (2023). https://arxiv.org/abs/2304.12345

    2023 arXiv

  13. [21]

    VoyageAI. 2025. VoyageAI (May 25 version). https://api.voyageai.com/v1/ embeddings

    2025

  14. [22]

    Haoran Wang, Qing Liu, Wei Zhang, and Yifan Xu. 2023. KGTransformer: Tempo- ral Knowledge Graph Transformer for Financial Event Prediction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 4567–4579. https://aclanthology.org/20...

    2023

  15. [23]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving Text Embeddings with Large Language Models.arXiv preprint arXiv:2401.00368 (2023)

    2023 arXiv

  16. [24]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual E5 Text Embeddings: A Technical Report. arXiv preprint arXiv:2402.05672 (2024)

    2024 arXiv

  17. [25]

    Stephen Wu, Jacob Austin, Aman Vyas, and et al. 2023. BloombergGPT: A Large Language Model for Finance. arXiv preprint arXiv:2304.03285 (2023). https: //arxiv.org/abs/2304.03285

    2023 arXiv

  18. [26]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]

    2023 arXiv

  19. [27]

    Zhiwei Yang, Yifan Xu, Wei Zhang, Haoran Wang, and Qing Liu. 2023. Fin- LLM: Financial Large Language Models for Algorithmic Trading. arXiv preprint arXiv:2306.09307 (2023). https://arxiv.org/abs/2306.09307

    2023 arXiv

  20. [28]

    Dun Zhang, Jiacheng Li, Ziyang Zeng, and Fulong Wang. 2025. Jasper and Stella: distillation of SOTA embedding models. arXiv:2412.19048 [cs.IR] https: //arxiv.org/abs/2412.19048

    2025 arXiv

Pith tools

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