Pith. sign in

REVIEW 3 major objections 5 minor 39 references

A Comparative Study of Specialized LLMs as Dense Retrievers

T0 review · 3 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 Specializing an LLM for math or long reasoning makes it a worse dense retriever, while code and vision-language specialization help or preserve retrieval.

desk verdict A useful empirical map of eight Qwen2.5 checkpoints for dense retrieval, but the causal claim about reasoning and math overreaches what the experiments actually test. read the letter →

arxiv 2507.03958 v2 pith:UCKENEOX submitted 2025-07-05 cs.IR cs.AIcs.CLcs.LG

Hengran Zhang , Keping Bi , Jiafeng Guo This is my paper
classification cs.IRcs.AIcs.CLcs.LG
keywords denseretrievallargelanguagemodelsspecializedLLMszero-shotcodemathematicalreasoningvision-languageMSMARCO
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 asks whether the extra training that turns a general large language model into a specialist changes how well that model can serve as a dense retriever, a system that encodes queries and documents as vectors and matches them by similarity. It compares eight Qwen2.5-7B checkpoints - base, instruction-tuned, code, math, long-reasoning, and vision-language variants - on zero-shot text retrieval (BEIR), zero-shot code retrieval (CoIR), and supervised fine-tuning on MS MARCO. The central finding is that math specialization and long-reasoning training consistently degrade retrieval in all three settings, while code-specialized and vision-language models perform best zero-shot and remain comparable after supervised training. The paper reads this as evidence that stepwise mathematical reasoning conflicts with the global semantic matching that retrieval requires, and that broad structural or cross-modal understanding is the more promising direction for a unified retriever.

What carries the argument

The load-bearing object is a controlled family of eight Qwen2.5-7B checkpoints sharing one architecture, with specialization as the only intended difference: base, instruct, coder, coder-instruct, math, math-instruct, R1-distill, and VL-instruct. All are evaluated as dual-encoder dense retrievers using the hidden state of the EOS token as the embedding, with identical prompts in zero-shot settings and identical contrastive fine-tuning on MS MARCO in the supervised setting. The argument works by comparing systematic gaps in NDCG@10 and MRR and recall metrics across these paired checkpoints, with BM25 serving as the sparse baseline.

What would settle it

Run DeepSeek-R1-Distill-Qwen-7B on the BEIR and CoIR zero-shot sets with chain-of-thought generation enabled and take the final-token hidden state as the embedding; if retrieval is no worse than the base Qwen2.5-7B, then long-reasoning training per se is not what degrades retrieval, and the paper's central attribution is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that among eight Qwen2.5-7B backbones, the math-specialized variants (Qwen2.5-Math and Qwen2.5-Math-Instruct) and the long-reasoning variant (DeepSeek-R1-Distill-Qwen-7B) are consistently the weakest dense retrievers, losing to the base model on zero-shot text retrieval, zero-shot code retrieval, and MS MARCO fine-tuning alike. By contrast, the code-specialized variants and the vision-language model Qwen2.5-VL-7B-Instruct lead zero-shot performance, with the VL model beating BM25 by 48% in average NDCG@10 on CoIR, and these same models stay competitive after supervised training. The paper attributes the consistent math and reasoning deficit to a conflict between reasoning-oriented objectives and the global semantic representation needed for matching queries to documents, and frames code and multimodal specialization as the promising route toward a unified retriever.

Load-bearing premise

The comparisons presume that the eight checkpoints differ only in specialization, but the long-reasoning model is tested without generating any reasoning and the checkpoints also differ in pretraining data size and alignment, so the blamed capability may not be the active factor.

Editorial extensions

If this is right

  • Unified retrieval systems should prefer code- or vision-language-pretrained backbones over math- or reasoning-specialized ones, since the latter lose on zero-shot text retrieval, zero-shot code retrieval, and supervised fine-tuning alike.
  • The math and reasoning deficit survives supervised contrastive training on MS MARCO, so retrieval damage from these specializations is not just a zero-shot prompting artifact.
  • Cross-modal training transfers to code retrieval: the vision-language model's large gains on CoIR suggest visual and code structure share representational benefits useful for matching.
  • Instruction tuning is not uniformly positive for retrieval: it helped code- and math-specialized models but hurt the base model on zero-shot text retrieval.
  • The long-reasoning result is explicitly provisional - the R1-distill model was run without emitting reasoning - so the paper's own causal story for reasoning is an open hypothesis rather than a closed conclusion.

Reading between the lines

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

  • A cleaner test of the long-reasoning claim would run the same checkpoint with chain-of-thought enabled and compare to the no-reasoning setting; the paper's design does not exercise the capability it names.
  • Because the checkpoints differ in pretraining data scale and alignment recipe, the specialization axis is confounded; a matched-token-count pretraining study would be needed to confirm that math content, rather than training scale, causes the degradation.
  • If the conflict is real, removing or reducing math-heavy tokens at a fixed pretraining budget should raise retrieval performance on BEIR and CoIR, which is a testable diagnostic.
  • The vision-language model's code-retrieval success suggests probing whether multimodal backbones transfer to other structured domains such as tables and graphs, which the paper does not evaluate.
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 / 5 minor

Summary. This manuscript presents an empirical comparison of eight Qwen2.5-family 7B checkpoints—base, instruct, code, code-instruct, math, math-instruct, DeepSeek-R1-Distill-Qwen, and VL-instruct—when used as dense retrievers. It evaluates zero-shot text retrieval on six BEIR datasets, zero-shot code retrieval on the CoIR benchmark, and supervised retrieval after fine-tuning on MS MARCO with TREC DL19/DL20 evaluation. The central reported finding is that math-specialized and long-reasoning checkpoints consistently degrade retrieval effectiveness in all three settings, which the authors interpret as evidence of a conflict between mathematical reasoning and semantic matching, while code-specialized and vision-language models perform best, especially in zero-shot code retrieval.

Significance. The value of the paper, if its descriptive results are accepted, is practical and empirical: it documents which publicly available Qwen2.5 checkpoints are suitable as dense-retriever backbones, and it extends the LLM-as-retriever comparison to code and multimodal variants. Strengths include the systematic coverage of eight checkpoints under a shared PromptReps protocol, inclusion of BM25 and supervised baselines, paired t-tests for the supervised results in Table 3, and a public code repository. The main weakness is that the causal headline goes beyond the observational design, particularly for the long-reasoning model, which is never run in its reasoning-emitting mode. The reported rankings may be a useful descriptive benchmark, but the 'conflict between mathematical reasoning and semantic matching' interpretation is not established.

major comments (3)
  1. [Abstract and §6] The claim that 'mathematical specialization and the long reasoning capability cause consistent degradation in three settings' is not supported by the experiments. DeepSeek-R1-Distill-Qwen-7B is never run in its reasoning-emitting mode; the paper concedes in §5.2 and §6 that the intermediate reasoning process was untested. The observed gaps (e.g., BEIR NDCG@10 of 0.3449 for Base vs 0.2831 for Distill; CoIR 0.1961 vs 0.1767) compare checkpoints under the identical 'Use one word' PromptReps prompt, so they measure a different post-trained checkpoint, not the 'long reasoning capability' in action. Please either test the model with reasoning output enabled or rephrase the conclusion as a descriptive statement about the checkpoint.
  2. [§4.1 and §6] The eight checkpoints are not clean interventions on 'specialization.' They differ in continued pretraining data (Qwen2.5-Math uses over 1T math tokens), post-training recipes (CoT/TIR SFT, DPO, RL distillation), chat templates, and even architecture (Qwen2.5-VL adds a vision encoder and MLP). Any of these confounds could plausibly explain the observed degradation, so the causal attribution to 'mathematical specialization' or 'long reasoning capability' is not justified by the design. The paper should either add controlled comparisons that isolate the target capability or explicitly present the study as observational.
  3. [§4.3, §5.1, §5.2] The zero-shot results are single-point estimates with no variance or significance testing. For example, the BEIR average differences (Base 0.3449 vs Coder-I 0.3904; Math 0.2531 vs Base 0.3449) are reported without error bars or statistical tests, and PromptReps-style embeddings are known to be sensitive to prompt wording. To support the cross-setting degradation claim, the authors should provide multiple runs or paired tests, or at least flag the absence of uncertainty estimates as a limitation.
minor comments (5)
  1. [§4.2] 'CoIR Benchamrk' is a typo; Table 2 also contains 'Hybird' instead of 'Hybrid.'
  2. [References] References [27] and [28] are the same paper ('Repetition improves language model embeddings'); the duplicate should be removed and in-text citations adjusted.
  3. [§5.2] In the final summary paragraph, 'even compressing BM25' should read 'even surpassing BM25.'
  4. [§4.4 and §5.3] There are several typos: 'Deepseed' in §4.4 should be 'DeepSpeed,' 'Deepseel-R1-Distill-Qwen-7B' appears in §5.1 item 4, and 'TRECL DL19' appears in §5.3.
  5. [Table 3] The paired t-test is described only as 'two-sided paired t-test (p<0.05)'; please specify what the pairing unit is (individual queries, topics, or folds) and how many observations were used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the study is an empirical benchmark whose comparisons are made from external checkpoints, not derived from the result being claimed.

full rationale

This paper is an empirical comparative study, not a derivation. It measures zero-shot and supervised retrieval NDCG@10 for eight externally released Qwen2.5-based checkpoints and reports the observed rankings. There is no fitted parameter that is later renamed as a prediction, no equation in which the reported outcome is embedded by construction, and no load-bearing self-citation. The only self-citation, reference [38] (Zhang et al. 2025), appears in related work as a description of prior query-likelihood modeling and is not used to justify the paper's main finding. The central claim that mathematical specialization and long reasoning 'cause' degradation is a causal interpretation of checkpoint comparisons, and the paper itself concedes that the reasoning-emitting mode of DeepSeek-R1-Distill-Qwen-7B was not tested due to computational costs. That concession weakens the causal claim and exposes confounds among architecture, continued pretraining data, chat template, and post-training recipe, but those are threats to internal validity, not circularity: the observed performance gaps are not manufactured by the paper's own model or equations. Since no claim reduces to its own inputs by definition or through a self-citation chain, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No fitted parameters are used to derive a result; the study is an empirical benchmark. Design choices listed are hand-selected and may affect the ranking.

free parameters (3)
  • Zero-shot prompt template = PromptReps-style 'Use one word...' (from Zhuang et al. 2024)
    Chosen from prior work rather than tuned per model; results may shift with different prompts.
  • LoRA rank = 32
    Fixed for all supervised runs, following RepLLaMA-style protocol; affects fine-tuned performance.
  • Training hyperparameters = LR 1e-4, batch size 32, 1 epoch
    Single configuration used across all models; may favor some checkpoints.
assumptions (3)
  • domain assumption Checkpoints are comparable proxies for model specialization
    Differences in retrieval are assumed to stem from specialization type, but checkpoints also differ in architecture, as VL has a vision tower, and in training data scale.
  • domain assumption BEIR/CoIR/MS MARCO are representative of retrieval tasks
    Generalization to other retrieval domains is extrapolated from these benchmarks.
  • domain assumption The inference-time embedding, using the last token hidden state, is a valid retrieval representation
    Standard in prior dense retrieval work, such as RepLLaMA and PromptReps, but different models may use the representation differently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Comparative Study of Specialized LLMs as Dense Retrievers." pith.science (2026). https://pith.science/paper/UCKENEOX

@misc{pith2026250703958,
  author       = {Pith},
  title        = {Pith review of: A Comparative Study of Specialized LLMs as Dense Retrievers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UCKENEOX}},
  note         = {Machine review of arXiv:2507.03958}
}
read the original abstract

While large language models (LLMs) are increasingly deployed as dense retrievers, the impact of their domain-specific specialization on retrieval effectiveness remains underexplored. This investigation systematically examines how task-specific adaptations in LLMs influence their retrieval capabilities, an essential step toward developing unified retrievers capable of handling text, code, images, and multimodal content. We conduct extensive experiments with eight Qwen2.5 7B LLMs, including base, instruction-tuned, code/math-specialized, long reasoning, and vision-language models across zero-shot retrieval settings and the supervised setting. For the zero-shot retrieval settings, we consider text retrieval from the BEIR benchmark and code retrieval from the CoIR benchmark. Further, to evaluate supervised performance, all LLMs are fine-tuned on the MS MARCO dataset. We find that mathematical specialization and the long reasoning capability cause consistent degradation in three settings, indicating conflicts between mathematical reasoning and semantic matching. The vision-language model and code-specialized LLMs demonstrate superior zero-shot performance compared to other LLMs, even surpassing BM25 on the code retrieval task, and maintain comparable performance to base LLMs in supervised settings. These findings suggest promising directions for the unified retrieval task leveraging cross-domain and cross-modal fusion.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 12 canonical work pages

  1. [38]

    arXiv preprint arXiv:2504.05216 (2025)

    Zhang, H., Bi, K., Guo, J., Sun, X., Liu, S., Shi, D., Yin, D., Cheng, X.: Unleashing the power of llms in dense retrieval with query likelihood modeling. arXiv preprint arXiv:2504.05216 (2025)

    arXiv 2025

  2. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  3. [2]

    arXiv preprint arXiv:2309.16609 (2023)

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)

    arXiv 2023

  4. [3]

    5-vl technical report

    Bai, S., Chen, K., Liu, X., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  5. [4]

    arXiv preprint arXiv:2404.05961 (2024)

    BehnamGhader, P., Adlakha, V., Mosbach, M., Bahdanau, D., Chapados, N., Reddy, S.: Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961 (2024)

    arXiv 2024

  6. [5]

    In: Advances in Information Retrieval: 38th European Conference on IR Research

    Boteva, V., Gholipour, D., Sokolov, A., Riezler, S.: A full-text learning to rank dataset for medical information retrieval. In: Advances in Information Retrieval: 38th European Conference on IR Research. pp. 716–722. Springer (2016)

    work page 2016

  7. [6]

    arXiv preprint arXiv:2002.03932 (2020)

    Chang, W.C., Yu, F.X., Chang, Y.W., Yang, Y., Kumar, S.: Pre-training tasks for embedding-based large-scale retrieval. arXiv preprint arXiv:2002.03932 (2020)

    arXiv 2020

  8. [7]

    arXiv preprint arXiv:2004.07180 (2020)

    Cohan, A., Feldman, S., Beltagy, I., Downey, D., Weld, D.S.: Specter: Document- level representation learning using citation-informed transformers. arXiv preprint arXiv:2004.07180 (2020)

    arXiv 2020

Show all 39 references
  1. [8]

    arXiv preprint arXiv:2003.07820 (2020)

    Craswell, N., Mitra, B., Yilmaz, E., Campos, D., Voorhees, E.M.: Overview of the trec 2019 deep learning track. arXiv preprint arXiv:2003.07820 (2020)

    2020 arXiv

  2. [9]

    arXiv preprint arXiv:2108.05540 (2021)

    Gao, L., Callan, J.: Unsupervised corpus aware language model pre-training for dense passage retrieval. arXiv preprint arXiv:2108.05540 (2021)

    2021 arXiv

  3. [10]

    arXiv preprint arXiv:2501.12948 (2025)

    Guo, D., Yang, D., Zhang, H., Song, J., et al.: Deepseek-r1: Incentivizing reason- ing capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

    2025 arXiv

  4. [11]

    ACM Transactions on Information Systems (TOIS) 40(4), 1–42 (2022)

    Guo, J., Cai, Y., Fan, Y., Sun, F., Zhang, R., Cheng, X.: Semantic models for the first-stage retrieval: A comprehensive review. ACM Transactions on Information Systems (TOIS) 40(4), 1–42 (2022)

    2022

  5. [12]

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

    Guo, J., Fan, Y., Ai, Q., Croft, W.B.: A deep relevance matching model for ad- hoc retrieval. In: Proceedings of the 25th ACM International on Conference on Information and Knowledge Management. p. 55–64 (2016)

    2016

  6. [13]

    arXiv preprint arXiv:2105.09938 (2021)

    Hendrycks, D., Basart, S., Kadavath, S., Mazeika, M., et al.: Measuring coding challenge competence with apps. arXiv preprint arXiv:2105.09938 (2021)

    2021 arXiv

  7. [14]

    arXiv preprint arXiv:2105.13239 (2021)

    Huang, J., Tang, D., Shou, L., et al.: Cosqa: 20,000+ web queries for code search and question answering. arXiv preprint arXiv:2105.13239 (2021)

    2021 arXiv

  8. [15]

    5-coder technical report

    Hui, B., Yang, J., Cui, Z., et al.: Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186 (2024)

    2024 arXiv

  9. [16]

    arXiv preprint arXiv:1909.09436 (2019)

    Husain, H., Wu, H.H., Gazit, T., Allamanis, M., Brockschmidt, M.: Codesearch- net challenge: Evaluating the state of semantic code search. arXiv preprint arXiv:1909.09436 (2019)

    2019 arXiv

  10. [17]

    In: EMNLP (1)

    Karpukhin, V., Oguz, B., Min, S., Lewis, P.S., et al.: Dense passage retrieval for open-domain question answering. In: EMNLP (1). pp. 6769–6781 (2020)

    2020

  11. [18]

    In: Proceedings of the 62nd Annual MeetingoftheAssociationforComputationalLinguistics(Volume1:LongPapers)

    Li, C., Liu, Z., Xiao, S., Shao, Y., Lian, D.: Llama2vec: Unsupervised adaptation of large language models for dense retrieval. In: Proceedings of the 62nd Annual MeetingoftheAssociationforComputationalLinguistics(Volume1:LongPapers). pp. 3490–3500 (2024)

    2024

  12. [19]

    arXiv preprint arXiv:2407.02883 (2024) 12 Hengran Zhang et al

    Li, X., Dong, K., Lee, Y.Q., Xia, W., et al.: Coir: A comprehensive benchmark for code information retrieval models. arXiv preprint arXiv:2407.02883 (2024) 12 Hengran Zhang et al

    2024 arXiv

  13. [20]

    arXiv preprint arXiv:2412.19437 (2024)

    Liu, A., Feng, B., Xue, B., et al.: Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  14. [21]

    In: Proceedings of the 14th ACM international conference on web search and data mining

    Ma, X., Guo, J., Zhang, R., Fan, Y., Ji, X., Cheng, X.: Prop: Pre-training with representative words prediction for ad-hoc retrieval. In: Proceedings of the 14th ACM international conference on web search and data mining. pp. 283–291 (2021)

    2021

  15. [22]

    In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Ma, X., Wang, L., Yang, N., Wei, F., Lin, J.: Fine-tuning llama for multi-stage text retrieval. In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 2421–2425 (2024)

    2024

  16. [23]

    In: Companion proceedings of the the web conference 2018

    Maia, M., Handschuh, S., Freitas, A., Davis, B., et al.: Www’18 open challenge: financial opinion mining and question answering. In: Companion proceedings of the the web conference 2018. pp. 1941–1942 (2018)

    2018

  17. [24]

    Nguyen, T., Rosenberg, M., Song, X., Gao, J., Tiwary, S., Majumder, R., Deng, L.: Ms marco: A human-generated machine reading comprehension dataset (2016)

    2016

  18. [25]

    arXiv preprint arXiv:2010.08191 (2020)

    Qu, Y., Ding, Y., Liu, J., Liu, K., et al.: Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2010.08191 (2020)

    2020 arXiv

  19. [26]

    Foundations and Trends® in Information Retrieval 3(4), 333–389 (2009)

    Robertson, S., Zaragoza, H., et al.: The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval 3(4), 333–389 (2009)

    2009

  20. [28]

    arXiv preprint arXiv:2402.15449 (2024)

    Springer, J.M., Kotha, S., Fried, D., Neubig, G., Raghunathan, A.: Repetition improves language model embeddings. arXiv preprint arXiv:2402.15449 (2024)

    2024 arXiv

  21. [29]

    arXiv preprint arXiv:2104.08663 (2021)

    Thakur, N., Reimers, N., Rücklé, A., Srivastava, A., Gurevych, I.: Beir: A heteroge- nous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663 (2021)

    2021 arXiv

  22. [30]

    Advances in neural information processing systems30 (2017)

    Vaswani, A., Shazeer, N., Parmar, N., et al.: Attention is all you need. Advances in neural information processing systems30 (2017)

    2017

  23. [31]

    In: Proceedings of the 56th Annual Meeting of the Associ- ation for Computational Linguistics (Volume 1: Long Papers)

    Wachsmuth, H., Syed, S., Stein, B.: Retrieval of the best counterargument without prior topic knowledge. In: Proceedings of the 56th Annual Meeting of the Associ- ation for Computational Linguistics (Volume 1: Long Papers). pp. 241–251 (2018)

    2018

  24. [32]

    ArXiv pp

    Wang, L.L., Lo, K., Chandrasekhar, Y., Reas, R., et al.: Cord-19: The covid-19 open research dataset. ArXiv pp. arXiv–2004 (2020)

    2020

  25. [33]

    arXiv preprint arXiv:2205.12035 (2022)

    Xiao, S., Liu, Z., Shao, Y., Cao, Z.: Retromae: Pre-training retrieval-oriented lan- guage models via masked auto-encoder. arXiv preprint arXiv:2205.12035 (2022)

    2022 arXiv

  26. [34]

    arXiv preprint arXiv:2310.04951 (2023)

    Yan, W., Tian, Y., Li, Y., et al.: Codetransocean: A comprehensive multilingual benchmark for code translation. arXiv preprint arXiv:2310.04951 (2023)

    2023 arXiv

  27. [35]

    Yang, A., Yang, B., Hui, B., et al.: Qwen2 technical report (2024)

    2024

  28. [36]

    Yang, A., Yang, B., Zhang, B., et al.: Qwen2.5 technical report (2025)

    2025

  29. [37]

    5-math technical report: Toward mathematical expert model via self-improvement

    Yang, A., Zhang, B., Hui, B., et al.: Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122 (2024)

    2024 arXiv

  30. [39]

    arXiv preprint arXiv:2402.14658 (2024)

    Zheng, T., Zhang, G., Shen, T., et al.: Opencodeinterpreter: Integrating code gen- eration with execution and refinement. arXiv preprint arXiv:2402.14658 (2024)

    2024 arXiv

  31. [40]

    arXiv preprint arXiv:2404.18424 (2024)

    Zhuang, S., Ma, X., Koopman, B., Lin, J., Zuccon, G.: Promptreps: Prompting large language models to generate dense and sparse representations for zero-shot document retrieval. arXiv preprint arXiv:2404.18424 (2024)

    2024 arXiv

Pith tools

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