Syntriever: How to Train Your Retriever with Synthetic Data from LLMs

Minsang Kim; Seungjun Baek

REVIEW 4 major objections 6 minor 56 references

Syntriever: How to Train Your Retriever with Synthetic Data from LLMs

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Syntriever trains small dense retrievers on LLM-written passages and LLM preference feedback, reporting nDCG@10 gains of up to 18.6% over prior retrieval models on BeIR benchmarks.

desk verdict Solid training recipe for black-box LLM distillation, but the SOTA claim overreaches: missing direct synthetic-data baseline and miscounted wins. read the letter →

arxiv 2502.03824 v3 pith:YNC2YE3Z submitted 2025-02-06 cs.CL cs.AI

Minsang Kim , Seungjun Baek This is my paper · ORCID

classification cs.CLcs.AI

keywords syntheticdatadenseretrievalLLMdistillationpreferencealignmentPlackett-Lucerankingchain-of-thoughtself-verificationblack-box

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

This paper claims that a small neural search model can be trained to match or beat much larger systems using only text written by a large language model (LLM), without ever seeing the LLM's internal probabilities. It proposes Syntriever, a two-stage recipe: generate synthetic relevant and misleading passages plus chain-of-thought query rewrites, self-check the generations for hallucination, and train the retriever to cluster the relevant ones; then collect pairwise relevance judgments from the LLM on passages the retriever itself found, and fine-tune with a ranking loss that keeps in-batch distractors below the judged pair. If the claims hold, users of black-box LLM APIs can distill retrieval knowledge into small, cheap encoders and improve ranking accuracy on standard benchmarks by up to 18.6% in nDCG@10, a standard ranking-quality metric.

What carries the argument

Two objectives carry the method, plus one data-cleaning mechanism. The distillation objective is a modified soft nearest-neighbor loss: for each query it puts similarities to all positive passages (labeled, synthetic, and chain-of-thought rewrite) in the numerator and similarities to those positives plus synthetic hard negatives and in-batch negatives in the denominator, so multiple relevant passages are clustered and negatives repelled. The alignment objective is partial Plackett-Luce ranking, a marginalization of the Plackett-Luce distribution over rankings in which the LLM-preferred passage ranks first, the dispreferred passage ranks second, and all in-batch samples fill the remaining positions in any order; its negative log-likelihood is the training loss, and the in-batch terms function as contrastive regularization that keeps the aligned model close to the distilled one. The data-cleaning mechanism is self-verification with relabeling: the LLM checks each synthetic positive for hallucination, and passages judged hallucinated are reused as hard negatives instead of being thrown away.

What would settle it

Sample a few hundred query-passage triples from FiQA and NFCorpus, ask the same LLM which passage is more relevant, and compare its choices with the human relevance labels in the BeIR ground truth; if agreement is at chance level, or if replacing the LLM's labels with random labels produces the same nDCG@10 gains, the reported improvements are not evidence that LLM preferences teach true relevance.

Watch

Extended reading notes

Core claim

The central claim is that LLM knowledge can be transferred into dense retrievers purely through synthetic text and preference feedback, with no access to output probabilities. In the distillation stage, for each query the LLM writes a chain-of-thought decomposition, a synthetic positive passage, and a plausible-but-irrelevant hard negative; a self-verification step relabels hallucinated positives as additional hard negatives rather than discarding them, and a modified soft nearest-neighbor loss pulls the query, its labeled passage, the chain-of-thought query, and the synthetic positive together while repelling negatives. In the alignment stage, the retriever retrieves its own top-$K$ passages, the LLM compares passage pairs, and a partial Plackett-Luce ranking objective treats the preferred passage as best, the dispreferred one as second, and all in-batch samples as lower-ranked 'don't care' items, combining preference learning with contrastive regularization. The paper reports nDCG@10 improvements over the second-best baseline of up to 18.6% across seven supervised BeIR datasets and gains on 11 of 15 datasets in zero-shot transfer.

Load-bearing premise

The load-bearing premise is that the LLM's pairwise relevance judgments agree with human judgments of retrieval quality, but the paper never measures that agreement or reports significance tests for the observed gains.

Editorial extensions

If this is right

Any dense retriever can be upgraded by the framework: the paper shows consistent nDCG@$K$ gains when ColBERT, SBERT, or Contriever is used as the base encoder, with the largest gains at nDCG@1.
Since the method needs only generated text and preference labels, it works with modern black-box LLM APIs that withhold output probabilities, a setting where earlier distillation methods fail.
The gains concentrate in hard, specialized-domain datasets, suggesting the framework is most valuable where top-$K$ results are only marginally relevant and fine-grained ranking matters.
The teacher models can be tiered: the paper finds a smaller LLM can supply synthetic passages competitively while the larger LLM matters more for preference alignment, which lowers the cost of the pipeline.

Reading between the lines

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

Beyond the paper: the hallucination-as-hard-negative trick could transfer to other synthetic-data pipelines, where false generations are usually discarded but could instead become training signal.
Beyond the paper: the partial Plackett-Luce formulation is a general recipe for any candidate-ranking task in which a judge can confidently order only the top few items; recommendation and reranking could reuse the same in-batch marginalization as regularization.
Beyond the paper: a cheap validity check is to measure agreement between the LLM's pairwise preferences and human relevance judgments before full training; high agreement should predict which datasets benefit most from the alignment stage.
Beyond the paper: the authors' stated limitation is the added cost of LLM inference; a natural extension is to map the cost-quality frontier by using a cheap LLM for synthesis and a strong LLM only for alignment, which their teacher-tiering results suggest could be favorable.

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

Short version: this is a solid, well-written paper on distilling black-box LLM knowledge into small retrievers. The central idea—partial Plackett-Luce ranking with in-batch negatives—is new to me, the derivation is correct, and the gains on BeIR are credible. But the headline 'SOTA' claim is not fully supported: the most relevant baseline, Wang et al. 2023, is cited but not included, and the '8 of 15' count in the text doesn't match the 7 wins in Table 2.

What's genuinely good: the loss in Eq. 7 is a clean max-likelihood objective for 'chosen > rejected > in-batch negatives', and the marginalization in Appendix B is right. The ablation shows that this beats Bradley-Terry alignment, which makes sense because it preserves contrastive separation during alignment. Reusing hallucinated positives as hard negatives is a nice trick. The framework is simple, works with different base encoders and LLMs, and the code is public. That's real work.

Soft spots, in decreasing order:

1. Missing direct baseline. The paper explicitly cites Wang et al. 2023 (GPT-4-generated query-document pairs for embedding training) but never compares against that method or any other synthetic-data-trained retriever. Since Syntriever is fundamentally a synthetic-data distillation method, this is the comparison that matters most. The claim should be qualified as 'among 125M retrievers trained without synthetic data' until that's added.

2. The 'best on 8 of 15' statement is wrong (it's 7). That's a minor but embarrassing internal inconsistency.

3. No error bars or significance tests. I'd want at least a few seeds for the main comparison, especially in zero-shot transfer where the gaps are sometimes small (e.g., Quora).

4. The Table 11 quality check uses GPT-4o to grade passages GPT-4o also generated. That's not evidence of quality relative to human judgment, but it's not load-bearing.

The concern about LLM preference reliability is weak: the final metric is human-annotated nDCG, so the observed improvement already entails empirical alignment with human relevance. The missing baseline and the miscount are the real issues.

Bottom line: this paper is worth a serious referee. The contributions are real and reproducible, and the main weak spots are fixable with an additional baseline and honest reporting. I'd send it out; I'd also ask the authors to rerun with E5-Mistral (or the closest synthetic-data trained retriever) and report variance.

Referee Report

4 major / 6 minor

Summary. The paper proposes Syntriever, a two-stage framework for training dense retrievers using only outputs from black-box LLMs. In Stage 1 (distillation), LLM-generated chain-of-thought sub-queries, synthetic positive passages, and plausible-but-irrelevant negative passages are used to fine-tune a base retriever with a modified soft nearest-neighbor loss; hallucinations detected by self-verification are relabeled as hard negatives. In Stage 2 (alignment), the retriever is further fine-tuned using pairwise LLM preference judgments over its top-K retrieved passages, modeled through a proposed partial Plackett-Luce ranking loss (Eq. 7). The method is evaluated on BeIR benchmarks under supervised fine-tuning and zero-shot transfer settings, and ablations study the contribution of each component, the choice of base encoder, and the choice of teacher LLM. The paper releases code, prompt templates, and hyperparameters.

Significance. If the empirical claims hold, Syntriever is a useful contribution to black-box LLM distillation for retrieval: it avoids output-probability access, uses only synthetic text, and combines synthesis-based distillation with preference alignment in a principled way. The derivation in Appendix B is correct, and Eq. (7) is indeed the negative log-likelihood of the marginalized Plackett-Luce distribution in Eq. (12), which gives the alignment loss a solid probabilistic grounding. The paper also runs several thoughtful ablations (Table 3, Tables 5-10), tests multiple base encoders (Table 4) and multiple teacher LLMs (Tables 6-7), and releases code and prompts. However, the headline 'state-of-the-art' claim is not established by the presented comparison: the baseline set omits direct synthetic-data-trained retrievers, the reported best-dataset count in the zero-shot table is internally inconsistent, and no error bars or significance tests are provided. These issues currently limit the confidence one can place in the empirical conclusions.

major comments (4)

[Abstract; Section 3.1, Tables 1-2] The state-of-the-art claim is unqualified and the baseline comparison omits the most directly relevant prior work: retrievers trained on LLM-generated synthetic query-document pairs, such as the E5-Mistral-7B model of Wang et al. (2023), which is cited in Related Work but not included in Tables 1 or 2. Because the paper's premise is synthetic-data distillation, a direct baseline of this type is necessary to support the abstract's 'state-of-the-art' statement. At minimum, the claim should be qualified as 'state-of-the-art among the 125M-parameter baselines considered here' or the synthetic-data baselines should be added; otherwise the central claim is unsupported.
[Section 3.2, Table 2] The text states that Syntriever achieves the best performances on 8 of 15 zero-shot datasets, but a careful count of Table 2 yields only 7: MSMARCO (50.1), Trec-covid (75.3), FIQA (39.5), Quora (88.9), SciDocs (19.7), NFCorpus (36.6), and NQ (62.2). This internal inconsistency must be corrected and the entire comparison re-verified, since the number of datasets on which the method is best is a headline quantitative claim.
[Tables 1-3, 5-10] No error bars, standard deviations across runs, or statistical significance tests are reported for any comparison. Several of the claimed improvements are small in absolute terms, for example the 2.5% gain on FiQA and 1.8% gain on SciFact in Table 1, and the ablation gains in Table 3 are reported as single point estimates. Without at least multiple-seed variance or a significance test, it is unclear whether these differences are reliable. The authors should report mean and standard deviation over at least three runs, or explicitly state that results are single-run and treat the margins accordingly.
[Section 3.9, Table 11; Section 2.3] The quality check for synthetic positives is self-referential: GPT-4o evaluates whether passages generated by GPT-4o can derive the ground-truth answer, without any agreement measure against human relevance judgments. Similarly, the alignment stage assumes GPT-4o's pairwise preference judgments in Eq. (6) are reliable teachers, but no analysis of agreement with human labels or of preference consistency is provided. The human-labeled test sets partially mitigate this concern, because improved nDCG on those sets is the ultimate evidence, but a direct preference-quality analysis would materially strengthen the paper's claim that LLM feedback is a valid alignment signal.

minor comments (6)

[Section 3.1] The word 'benckmark' in the zero-shot transfer description is a typo and should read 'benchmark.'
[Section 3.2] The sentence listing improvements over the second-best models omits Fever; please either include the Fever gain or explicitly state why it is not listed.
[Section 2.3] Equation (6) writes (q_i, c+_i, c-_i) = M(PCompare(q_i, c_{i,j}, c_{i,k})), but the prompt template in Appendix C.4 takes only two passages and a question; the notation should clarify that q_i passes through unchanged and that the LLM returns the preferred and dispreferred passages.
[Section 2.3] The subsection header 'From Bradely-Terry to Plackett-Luce model' contains a typo: 'Bradely' should be 'Bradley.'
[Table 7] The caption contains the typo 'Comparion' and should read 'Comparison.'
[Appendix D, Table 12] The table reports batch sizes for Stage 1 and Stage 2 but not the number of training steps or epochs; adding this information would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: training losses are derived from stated objectives and evaluation uses held-out human labels.

full rationale

Training and evaluation are decoupled: the retriever is trained on LLM-generated synthetic passages plus the dataset's own labeled positives and negatives, and evaluated on held-out human-labeled BeIR test collections, so no test label enters any loss (Eqs. 5-7). The stage-2 loss is derived from the Plackett-Luce model by a correct marginalization (Eq. 12 and Appendix B), not assumed equal to the desired metric. The SOTA claim is an empirical comparison, not a derivation, and any concern about omitted baselines or LLM-preference noise is a correctness or robustness issue, not circularity. The only self-referential evidence, Table 11, uses GPT-4o to judge passages GPT-4o generated; this is a weak quality check but it does not enter the training loop and is not used to define the main result, so it does not constitute load-bearing circularity. No self-citation is used as a load-bearing premise.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The framework rests on standard contrastive-learning assumptions plus two LLM-dependent assumptions (self-verification and preference reliability). The only mathematical axiom is the PL marginalization, which is proven in the appendix. No new physical or ontological entities are introduced.

free parameters (5)

temperature tau = 0.05
Controls the concentration of normalized embeddings in contrastive losses (Eq. 2 and Eq. 7); chosen by hand.
alignment K = 5
Number of top passages retrieved for preference feedback in Stage 2; default chosen as a trade-off between performance and cost (Sec 3.8).
comparison pairs N = 10
Number of pairwise comparisons per query, K(K-1)/2; the paper shows performance varies with N.
batch size (Stage 1) = 60
Hyperparameter for contrastive training; affects the set of in-batch negatives in Eq. 7.
batch size (Stage 2) = 100
Hyperparameter for alignment; affects in-batch negatives in the partial PL loss.

assumptions (5)

domain assumption LLM self-verification reliably distinguishes relevant synthetic passages from hallucinations.
Invoked in Eq. (5) and Sec 3.5; the whole data-quality pipeline depends on it, but the paper only cites prior work and provides no accuracy measurement.
domain assumption In-batch passages from other queries are negatives for query q_i.
Used in Eq. (2) and Eq. (7); standard contrastive learning assumption, but with large batches false negatives are possible and unquantified.
ad hoc to paper Cosine similarity with temperature can serve as the reward function in the Plackett-Luce model.
The paper sets r(q,y)=s_tau(q,y) in Eq. (12); this design choice is not derived and is central to the alignment loss.
domain assumption GPT-4o's pairwise preference judgments approximate human relevance.
Used in Stage 2 (Eq. 6); no agreement analysis with human labels is provided.
standard math The Plackett-Luce marginalization formula (Eq. 12) is correct.
Derived in Appendix B; the algebra is standard and the derivation checks out.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Syntriever: How to Train Your Retriever with Synthetic Data from LLMs." pith.science (2026). https://pith.science/paper/YNC2YE3Z

@misc{pith2026250203824,
  author       = {Pith},
  title        = {Pith review of: Syntriever: How to Train Your Retriever with Synthetic Data from LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNC2YE3Z}},
  note         = {Machine review of arXiv:2502.03824}
}

abstract

LLMs have boosted progress in many AI applications. Recently, there were attempts to distill the vast knowledge of LLMs into information retrieval systems. Those distillation methods mostly use output probabilities of LLMs which are unavailable in the latest black-box LLMs. We propose Syntriever, a training framework for retrievers using synthetic data from black-box LLMs. Syntriever consists of two stages. Firstly in the distillation stage, we synthesize relevant and plausibly irrelevant passages and augmented queries using chain-of-thoughts for the given queries. LLM is asked to self-verify the synthetic data for possible hallucinations, after which retrievers are trained with a loss designed to cluster the embeddings of relevant passages. Secondly in the alignment stage, we align the retriever with the preferences of LLMs. We propose a preference modeling called partial Plackett-Luce ranking to learn LLM preferences with regularization which prevents the model from deviating excessively from that trained in the distillation stage. Experiments show that Syntriever achieves state-of-the-art performances on benchmark datasets from various domains in nDCG@$K$. The code is available at \href{https://github.com/kmswin1/Syntriever}{https://github.com/kmswin1/Syntriever}.

Figures

Figures reproduced from arXiv: 2502.03824 by the authors.

**Figure 1.** Overview of Syntriever. Stage-1 (Distillation Stage). Given a query, Syntriever uses LLMs to synthesize (i) related sub-queries (prompt Pcot), (ii) relevant passages ( P+) which are self-verified for hallucination ( PRelabel), (iii) plausibly irrelevant passages ( P−). The retriever is trained with the synthetic positive and negative passages. Stage-2 (Alignment Stage). The retriever is aligned with the LLM preferen… view at source ↗

**Figure 2.** Example of LLM synthesis. The correct an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Prompt template design for generating synthetic positive passages. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Prompt template design for generating plausible but irrelevant passages. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

**Figure 5.** Figure 5: Prompt template design of relabeling for synthetic positive passages. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 6.** Figure 6: Prompt template design for comparison of a passage pair. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 10 canonical work pages

[1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...
[4]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...
[5]

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

work page
[6]

Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219

arXiv 2024
[7]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

arXiv 2023
[8]

Hritik Bansal, Arian Hosseini, Rishabh Agarwal, Vinh Q Tran, and Mehran Kazemi. 2024. Smaller, weaker, yet better: Training llm reasoners via compute-optimal sampling. arXiv preprint arXiv:2408.16737

arXiv 2024

Show all 56 references

[9]

Ralph Allan Bradley and Milton E Terry. 1952. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324--345

1952
[10]

Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870--1879

2017
[11]

Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216

2024 arXiv
[12]

Yihe Deng, Weitong Zhang, Zixiang Chen, and Quanquan Gu. 2023. Rephrase and respond: Let large language models ask better questions for themselves. arXiv preprint arXiv:2311.04205

2023 arXiv
[13]

Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

2018 arXiv
[14]

Ronen Eldan and Yuanzhi Li. 2023. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759

2023 arXiv
[15]

Nicholas Frosst, Nicolas Papernot, and Geoffrey Hinton. 2019. Analyzing and improving representations with the soft nearest neighbor loss. In International conference on machine learning, pages 2012--2020. PMLR

2019
[16]

Luyu Gao and Jamie Callan. 2022. Unsupervised corpus aware language model pre-training for dense passage retrieval. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2843--2853

2022
[17]

Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio C \'e sar Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. 2023. Textbooks are all you need. arXiv preprint arXiv:2306.11644

2023 arXiv
[18]

Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. 2024. Deepseek-coder: When the large language model meets programming--the rise of code intelligence. arXiv preprint arXiv:2401.14196

2024 arXiv
[19]

Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning, pages 3929--3938. PMLR

2020
[20]

Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729--9738

2020
[21]

Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118

2021 arXiv
[22]

Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

2023 arXiv
[23]

Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)...

2020
[24]

Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

2014 arXiv
[25]

Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

2022
[26]

Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for C...

2019
[27]

Angeliki Lazaridou, Elena Gribovskaya, Wojciech Stokowiec, and Nikolai Grigorev. 2022. Internet-augmented language models through few-shot prompting for open-domain question answering. arXiv preprint arXiv:2203.05115

2022 arXiv
[28]

Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6086--6096

2019
[29]

u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

2020
[30]

Yuanzhi Li, S \'e bastien Bubeck, Ronen Eldan, Allie Del Giorno, Suriya Gunasekar, and Yin Tat Lee. 2023. Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463

2023 arXiv
[31]

Zhuowan Li, Cheng Li, Mingyang Zhang, Qiaozhu Mei, and Michael Bendersky. 2024. Retrieval augmented generation or long-context llms? a comprehensive study and hybrid approach. arXiv preprint arXiv:2407.16833

2024 arXiv
[32]

R Duncan Luce. 1959. Individual choice behavior, volume 4. Wiley New York

1959
[33]

Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2024. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36

2024
[34]

Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. Ms marco: A human-generated machine reading comprehension dataset

2016
[35]

Zach Nussbaum, John X Morris, Brandon Duderstadt, and Andriy Mulyar. 2024. Nomic embed: training a reproducible long context text embedder. arXiv preprint arXiv:2402.01613

2024 arXiv
[36]

Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748

2018 arXiv
[37]

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 3...

2022
[38]

Robin L Plackett. 1975. The analysis of permutations. Journal of the Royal Statistical Society Series C: Applied Statistics, 24(2):193--202

1975
[39]

Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. Rocketqa: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North Amer...

2021
[40]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36

2024
[41]

Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJC...

2019
[42]

Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, Qiaoqiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. Rocketqav2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language P...

2021
[43]

Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

2009
[44]

Joshua David Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. 2021. Contrastive learning with hard negative samples. In International Conference on Learning Representations

2021
[45]

Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. 2023. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950

2023 arXiv
[46]

Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Richard James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2024. Replug: Retrieval-augmented black-box language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computa...

2024
[47]

Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

2023 arXiv
[48]

Nandan Thakur, Nils Reimers, Andreas R \"u ckl \'e , Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663

2021 arXiv
[49]

Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533

2022 arXiv
[50]

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 arXiv
[51]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

2022
[52]

Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text degeneration with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020

2020
[53]

Yixuan Weng, Minjun Zhu, Fei Xia, Bin Li, Shizhu He, Shengping Liu, Bin Sun, Kang Liu, and Jun Zhao. 2023. Large language models are better reasoners with self-verification. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2550--2575

2023
[54]

Bennett, Junaid Ahmed, and Arnold Overwijk

Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. https://openreview.net/forum?id=zeFrfgyZln Approximate nearest neighbor negative contrastive learning for dense text retrieval . In International Conference o...

2021
[55]

Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. 2024. Distilling system 2 into system 1. arXiv preprint arXiv:2407.06023

2024 arXiv
[56]

Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593

2019 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

[4] [4]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

[5] [5]

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

work page

[6] [6]

Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219

arXiv 2024

[7] [7]

Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

arXiv 2023

[8] [8]

Hritik Bansal, Arian Hosseini, Rishabh Agarwal, Vinh Q Tran, and Mehran Kazemi. 2024. Smaller, weaker, yet better: Training llm reasoners via compute-optimal sampling. arXiv preprint arXiv:2408.16737

arXiv 2024