Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Real-Time Hybrid Retrieval in Hyperbolic Space for Retrieval-Augmented Generation on Edge Devices

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A 196K-parameter learned projection into hyperbolic space, combined with BM25, supports competitive zero-shot retrieval on edge devices at a few milliseconds per query.

desk verdict The paper's own Table 1 falsifies its central claim: the hyperbolic reranking is inert and the hybrid reduces to BM25, though the engineering is honest and clearly reported. read the letter →

arxiv 2608.01450 v1 pith:LQVEGDSV submitted 2026-08-02 cs.IR

classification cs.IR
keywords hyperbolicgeometryLorentzmodelretrieval-augmentedgenerationBM25hybridretrievalzero-shotedgecomputingdense
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 tries to establish that retrieval-augmented generation can run entirely on edge devices if documents are embedded in hyperbolic space with a lightweight learned projection, then retrieved by a two-stage BM25-plus-hyperbolic-rerank pipeline. The central claim is that the Lorentz model's exponential volume growth fits the hierarchical organisation of language, and that a projection of frozen word embeddings into that space, trained contrastively with fewer than 200,000 parameters, preserves enough semantic structure to be competitive with much larger retrievers. On five standard retrieval test collections the system reports NDCG@10 scores from 0.150 to 0.654, with per-query latency of about 3 ms on a mid-range phone. The paper's own tables note that every configuration, including pure BM25, produces nearly identical rankings, so the reported quality is carried by the embeddings and lexical matching rather than by the hyperbolic reranker itself. If true, the benefit is practical: RAG over tens of thousands of documents becomes feasible without GPU or server infrastructure.

What carries the argument

The load-bearing object is the learned projection HyTE-H into the Lorentz hyperboloid $H^d_K = \{p=(t,x)\in\mathbb{R}^{1,d} : \langle p,p\rangle_L=-K, t>0\}$, together with the Outward Einstein Midpoint (OEM) aggregation. The time coordinate $t=\sqrt{\|x\|^2+K}$ encodes hierarchical depth; the radial coordinate $r=\operatorname{arcosh}(t/\sqrt{K})$ measures specificity, and the Lorentzian inner product $\langle p,q\rangle_L = -t_p t_q + x_p^\top x_q$ serves directly as the similarity score, since $\operatorname{arcosh}$ is monotonic. BM25 is the other half: it selects the candidate set, and Equation (10) blends the normalised BM25 score with a logistic-sigmoid of the Lorentz similarity throu

What would settle it

Run the pipeline on a corpus where lexical and semantic rankings are known to diverge, such as queries containing rare named entities or numeric values, and test whether $\alpha=0.7$ improves NDCG@10 over $\alpha=1.0$ by more than 0.001. If the Lorentz re-ranker never changes any top-k ranking on any dataset, the claimed precision benefit of hyperbolic reranking is falsified.

Watch

Extended reading notes

Core claim

The discovery this paper reports is that a full hybrid retrieval pipeline can live inside the Lorentz model of hyperbolic geometry and still be cheap enough for a phone. Concretely: frozen pretrained word embeddings are mapped into the hyperboloid $H^{384}_K$ by a two-layer ReLU projection (HyTE-H) trained contrastively on passage-retrieval pairs; document chunks are pooled with the Outward Einstein Midpoint, which weights tokens by their time coordinate so more specific concepts sit farther from the origin; and retrieval is done in two stages, BM25 first, then Lorentzian inner-product reranking, blended by the parameter $\alpha$. On five standard test collections (SciFact, NFCorpus, ArguAna

Load-bearing premise

The system's hybrid score improves over either signal alone only if BM25 and Lorentzian similarity rank documents differently; the paper's own results show that on these five datasets they produce nearly the same top rankings, so the precision gain rests on unseen corpora where the two signals diverge.

Editorial extensions

If this is right

  • RAG over tens of thousands of documents can run fully on-device: index construction is one-time and offline, and queries take about 3 ms on a mid-range phone.
  • A projection with under 200,000 trainable parameters suffices to map frozen word embeddings into a usable hyperbolic retrieval space, removing the need for GPU fine-tuning during deployment.
  • Because hyperbolic scoring runs only on the BM25 candidate set, per-query latency depends on the candidate cutoff, not the corpus size.
  • The radial coordinate learned by the projection encodes specificity for most concept pairs tested, which could later be used to adapt retrieval depth to query generality.
  • On the five evaluated datasets, retrieval quality is bounded by the frozen embeddings; the paper identifies replacing them with a stronger encoder as the next step for accuracy gains.

Reading between the lines

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

  • If the near-tied alpha results hold beyond these five datasets, the fair reading is that the reported NDCG@10 is a property of BM25 plus frozen embeddings, and the hyperbolic projection would need a direct Euclidean ablation, same projection dimensions and training loss but cosine similarity in $\mathbb{R}^{384}$, to demonstrate any geometric benefit.
  • The radial hierarchy check, with three of four concept pairs showing the expected specificity ordering and one pair collapsing, suggests the learned specificity axis is noisy; a larger, statistically powered concept-taxonomy test would tell whether radial depth can be trusted as a query-adaptation signal.
  • The exhaustive Lorentz nearest-neighbour search caps practical scale around $10^5$ chunks; extending to million-document corpora would require approximate hyperbolic indexing, and the paper's latency numbers would likely change if that replacement is made.
  • A testable extension: construct a corpus with paired queries where lexical and semantic relevance disagree, such as rare named entities or paraphrase queries, then measure whether any $\alpha \in (0,1)$ beats both $\alpha=0$ and $\alpha=1$. This would convert the paper's $\alpha$ claim from assertion to measurement.
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

4 major / 5 minor

Summary. The paper presents a hybrid document retrieval system for edge-device RAG. It projects frozen BGE-small word embeddings into the Lorentz hyperboloid through a learned two-layer projection (HyTE-H), indexes document chunks, and retrieves via a two-stage pipeline: BM25 candidate generation followed by Lorentzian inner-product reranking, with scores blended by a weight alpha. The system is evaluated on five BEIR datasets, reporting NDCG@10 values, latency, and index size, and includes a radial hierarchy check on four concept pairs. The stated contribution is a BM25-for-recall, Lorentz-for-precision pipeline that runs at interactive latencies on edge devices.

Significance. If the hybrid pipeline were validated, the claimed result—competitive zero-shot retrieval on BEIR with a <200K-parameter projection, no fine-tuned encoder, and no GPU inference—would be of practical interest for on-device RAG. The paper has genuine strengths: the evaluation is externally grounded (projection trained on MS MARCO, tested zero-shot on BEIR), the engineering artifacts (latency, index size, serialization format) are concrete, and source code is referenced. However, the paper's own Table 1 shows that NDCG@10 is invariant to alpha across all datasets, so the Lorentz component contributes no measurable ranking signal beyond BM25. The only direct geometric evidence is partially contradicted by Table 3. Thus the central scientific claim is not established by the reported experiments.

major comments (4)
  1. [Section 4.2, Table 1, Eq. (10)] The central claim that the contribution is a 'BM25-for-recall, Lorentz-for-precision pipeline' is contradicted by the paper's own data. NDCG@10 is tied to four decimal places across all configurations on every dataset: alpha=0.3 and alpha=1.0 both give SciFact 0.6537, FiQA 0.2169 vs 0.2168, and identical values on the other three datasets. Since alpha=0.3 assigns 70% weight to the Lorentz score in Eq. (10), any independent ranking signal from s_L would perturb the top-10 list. The text's concession that 'the two scoring signals happen to favour the same top-ranked documents' removes the evidential basis for the claimed BM25/Lorentz synergy. The paper needs an alpha=0.0 ablation and a demonstration that Lorentz reranking changes candidate rankings; without that, the retrieval contribution reduces to BM25.
  2. [Section 3.3 and Section 4.2] No baseline is reported with unprojected BGE embeddings or with a Euclidean projection trained under the same contrastive objective. The reported BEIR scores are consistent with plain BM25 (e.g., SciFact 0.654 vs a typical published BM25 value around 0.665), and Table 1 shows alpha=1.0 matches the hybrid configurations. Therefore the causal role of the hyperbolic geometry is not tested. The paper should include controls: frozen BGE Euclidean cosine similarity, a Euclidean linear projection trained with the same loss, and alpha=0.0. Without these, the claim that hyperbolic space, rather than the frozen embeddings and BM25, drives retrieval quality is unsupported.
  3. [Section 4.3, Table 3] The radial hierarchy verification is only partially successful: the pair mammal→dog shows a -21.9% collapse, directly contradicting the expectation that more specific terms sit farther from the origin. With only four pairs and no error bars or statistical assessment, the claim that the HyTE-H projection 'preserves hierarchical structure' (Abstract) is not supported by the reported geometric evidence. Either a principled explanation for the collapse must be given, or the claim should be qualified.
  4. [Section 4.2, Table 2] The text states 'Retrieval latency grows linearly with corpus size, as exhaustive nearest-neighbour search predicts,' but Table 2 does not show a clean linear relation: ArguAna (8,750 chunks) has 18.7 ms/q, while SciDocs (26,199 chunks) has 16.0 ms/q and NFCorpus (3,664 chunks) has 2.7 ms/q. The engineering claim about scaling on edge devices requires either a clearer explanation of these outliers or a separate controlled scaling experiment.
minor comments (5)
  1. [Section 3.4 vs Section 4.1] The default chunking parameters are inconsistent: Section 3.4 states W=300 words with O=60 overlap, while Section 4.1 says 'default chunking configuration of 500-word windows with 90-word overlap.' Please align.
  2. [Section 4.2, Table 1] No variance or significance testing is reported. Since all NDCG@10 differences are below 0.001, the claim of invariance should be backed by repeated runs or paired tests, or at least by stating that the differences are within evaluation noise.
  3. [Section 4.3, Table 3] The radial hierarchy test uses only four concept pairs. This is a very small sample; please report more pairs or justify the selection.
  4. [Section 2.2] The statement that HypRAG's HyTE-H 'achieves up to 29% gains over Euclidean baselines' is cited to [7], but the relation of this paper's projection module to that claimed gain is not clarified. Please state explicitly whether the same trained weights are used or whether the projection was retrained here.
  5. [Section 3.6, Eq. (12)] The logistic map uses a constant +2 shift with no justification. Since the Lorentz inner product has a fixed range for points on the hyperboloid, please explain the choice of this offset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: BEIR evaluation is externally grounded; alpha-invariance is an evidentiary failure, not a circular derivation.

full rationale

The derivation chain is self-contained and externally grounded. The only learned component, the HyTE-H projection (Eqs. 7-8), is trained contrastively on MS MARCO (Eq. 9) with frozen BGE embeddings, and then evaluated zero-shot on five BEIR datasets; those BEIR numbers are therefore not constructed from the evaluation data. The hybrid scoring rule (Eqs. 10-12) is a defined interpolation, not a derivation that presupposes its conclusion. The radial hierarchy test (Table 3) is an independent check of a geometric property not present in the training loss, and it even reports one failure (mammal->dog), so it is not rigged. The references to HypRAG [7] and HELM [12] are to prior work by other author groups; the sole self-reference is the GitHub link [21], which is not load-bearing. The paper's own Section 4.2 concedes that NDCG@10 is invariant to alpha and that 'the two scoring signals happen to favour the same top-ranked documents'; that is a serious evidence gap for the claimed BM25-for-recall/Lorentz-for-precision synergy, but it is a falsification/correctness problem, not a circular reduction. Similarly, the absence of a Euclidean/unprojected baseline means the hyperbolic hypothesis is untested, but that omission is not a circular step. Therefore no circularity.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The central claim rests on many inherited components and hand-set hyperparameters. The only genuinely fitted parameters in the paper are the projection weights; but the paper's own experimental invariance shows that even these, combined with the Lorentz similarity, do not change outcomes relative to BM25. The ledger is heavy on defaults adopted from prior work and light on independent verification.

free parameters (10)
  • K (curvature) = 1.0
    Chosen as the curvature parameter for the Lorentz model (Eq. 3, 8); no sensitivity analysis.
  • P (OEM exponent) = 1.0
    Default from HypRAG, used for the Outward Einstein Midpoint pooling in Section 2.2.
  • tau (temperature) = 0.05
    Temperature in the contrastive loss (Eq. 9), set without sensitivity analysis.
  • alpha (hybrid weight) = 0.3/0.7/1.0 tested
    Blend weight in Eq. 10, tunable; shown in Table 1 to have no measurable effect on NDCG@10.
  • BM25 k1 = 1.2
    Standard Okapi BM25 term frequency saturation, Section 3.5.
  • BM25 b = 0.75
    Standard length normalization parameter, Section 3.5.
  • Chunk window W = 500 in experiments (default 300 in Section 3.4)
    Sliding window size; Section 4.1 uses 500, Section 3.4 says default 300, inconsistent.
  • Chunk overlap O = 90 in experiments (default 60 in Section 3.4)
    Overlap between chunks; inconsistent between Section 3.4 and Section 4.1.
  • M (BM25 candidate count) = 50
    Number of BM25 candidates passed to Lorentz reranking, Section 3.6.
  • HyTE-H projection weights W1, W2 = learned (196,608 parameters)
    Trained contrastively on MS MARCO, 75,000 steps; these are fitted parameters the system depends on.
assumptions (5)
  • domain assumption Lorentz model is an appropriate geometry for word semantics
    Section 3.2 and introduction assume exponential volume growth matches natural language hierarchy; adopted from prior work (Nickel and Kiela, HypRAG).
  • domain assumption OEM pooling preserves hierarchy
    Section 2.2 leverages HypRAG's OEM operator without independent verification in this paper.
  • domain assumption BGE-small-en-v1.5 word embeddings are a valid frozen token encoder
    Section 3.3 chooses BGE-small without comparison to other embedding sets.
  • domain assumption MS MARCO contrastive training transfers zero-shot to BEIR
    Projection trained on MS MARCO (Section 3.3.1) and evaluated on BEIR; no analysis of transfer failure modes.
  • domain assumption BEIR NDCG@10 is a valid measure of RAG retrieval quality
    Used as the primary metric in Section 4.2 without justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Real-Time Hybrid Retrieval in Hyperbolic Space for Retrieval-Augmented Generation on Edge Devices." pith.science (2026). https://pith.science/paper/LQVEGDSV

@misc{pith2026260801450,
  author       = {Pith},
  title        = {Pith review of: Real-Time Hybrid Retrieval in Hyperbolic Space for Retrieval-Augmented Generation on Edge Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQVEGDSV}},
  note         = {Machine review of arXiv:2608.01450}
}
abstract

This paper presents a hybrid document retrieval system designed for retrieval-augmented generation (RAG) that operates entirely within the Lorentz model of hyperbolic geometry. Unlike conventional dense retrievers confined to Euclidean space, this system projects pretrained word embeddings into hyperbolic space through a learned HyTE-H transformation, whose exponential volume growth suits the hierarchical organization of natural language. Documents are segmented into overlapping chunks, indexed by their Lorentz embeddings, and retrieved through a two-stage pipeline that first applies BM25 lexical scoring, then re-ranks candidates using Lorentzian inner-product similarity. A tunable parameter $\alpha$ blends the BM25 score with the hyperbolic similarity score. The system was evaluated on five datasets from the BEIR benchmark suite, SciFact, NFCorpus, ArguAna, SciDocs, and FiQA, achieving NDCG@10 scores of 0.654, 0.304, 0.342, 0.150, and 0.217 respectively with word embeddings alone, without fine-tuned neural encoders or cross-attention rerankers. The system supports real-time indexing of user-supplied documents and resource-efficient querying over tens of thousands of moderately sized documents, so hyperbolic retrieval can run on edge devices at interactive latencies.

Figures

Figures reproduced from arXiv: 2608.01450 by the authors.

Figure 1
Figure 1. System architecture. The indexing phase (top) processes documents through chunking, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Hyperbolic manifold geometry in the Poincar´e disk model. Concepts organise into [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The Lorentz hyperboloid H1 K (one spatial dimension). The time coordinate t encodes hierarchical depth. Points farther from the origin (larger t) are more specific concepts. The dashed red curve is a geodesic between two points, as in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The HyTE-H encoding pipeline. 3.4 Sliding Window Chunking Documents are segmented into overlapping chunks using a sliding window with configurable stride and overlap parameters. By default the window is W = 300 words, with O = 60 words of overlap between consecutive ch…
Figure 5
Figure 5. Figure 5: Contrastive HyTE-H training. Query and document tokens pass through identical [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: BM25 scoring process. against the query embedding (encoded through the same HyTE-H pipeline), and the two scores are linearly interpolated: score(d, q) = α·sbm(d, q)+(1−α)·sL(d, q) (10) where the component scores are sbm(d, q) = BM25(d, q) maxd ′ BM25(d ′ , q) , (11) s…
Figure 7
Figure 7. Figure 7: Two-stage retrieval with tunable α. the BM25 inverted index state. The embedding matrix is one contiguous block of floats, one vector per chunk, so scoring walks sequential memory and the file maps directly onto the in￾memory array on load. The texts.z file stores the …
Figure 8
Figure 8. Figure 8: Edge runtime architecture. The Android application layer uses JNI bindings to [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 6 canonical work pages

  1. [1]

    Retrieval- augmented generation for knowledge- intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K¨ uttler, M. Lewis, W. tau Yih, T. Rockt¨ aschel, S. Riedel, and D. Kiela, “Retrieval- augmented generation for knowledge- intensive NLP tasks,” inAdvances in Neural Information Processing Systems 33 (NeurIPS), 2020. [Online]. Available: https://arxiv.org/abs/2005.11401

  2. [2]

    A survey on RAG meeting LLMs: Towards retrieval- augmented large language models,

    W. Fan, Y. Ding, L. Ning, S. Wang, H. Li, D. Yin, T.-S. Chua, and Q. Li, “A survey on RAG meeting LLMs: Towards retrieval- augmented large language models,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2024, pp. 6491–6501. [Online]. Available: https://doi.org/10. 1145/3637528.3671470

  3. [3]

    Hubs in space: Popular nearest neighbors in high-dimensional data,

    M. Radovanovi´ c, A. Nanopoulos, and M. Ivanovi´ c, “Hubs in space: Popular nearest neighbors in high-dimensional data,”Journal of Machine Learning Research (JMLR), vol. 11, pp. 2487–2531,

  4. [4]

    Sentence- BERT: Sentence embeddings using siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence- BERT: Sentence embeddings using siamese BERT-networks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, pp. 3980–3990. [Online]. Available: https://doi.org/10.18653/v1/D19-1410

  5. [5]

    Poincar´ e embeddings for learning hierarchical rep- resentations,

    M. Nickel and D. Kiela, “Poincar´ e embeddings for learning hierarchical rep- resentations,” 2017. [Online]. Available: https://arxiv.org/abs/1705.08039

  6. [6]

    HTS Dipole Magnet with a Mechanical Energy Transfer in the Magnetic Field

    G. Mishne, Z. Wan, Y. Wang, and S. Yang, “The numerical stability of hyperbolic representation learning,” 2023. [Online]. Available: https://arxiv.org/abs/ 2211.00105

  7. [7]

    HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation

    H. Madhu, N. Bui, A. Maatouk, L. Tas- siulas, S. Krishnaswamy, M. Yang, S. Ganguly, K. Srinivasan, and R. Ying, “HypRAG: Hyperbolic dense retrieval for retrieval augmented generation,” inProceedings of the 43rd Interna- tional Conference on Machine Learn- ing (ICML), 2026. [Online]. Available: https://arxiv.org/abs/2602.07739

  8. [8]

    Learning contin- uous hierarchies in the Lorentz model of hyperbolic geometry,

    M. Nickel and D. Kiela, “Learning contin- uous hierarchies in the Lorentz model of hyperbolic geometry,” inProceedings of the 35th International Conference on Machine Learning (ICML), 2018. [Online]. Avail- able: https://arxiv.org/abs/1806.03417

Show all 22 references
  1. [9]

    Hyperbolic neural net- works,

    O.-E. Ganea, G. B´ ecigneul, and T. Hofmann, “Hyperbolic neural net- works,” 2018. [Online]. Available: https://arxiv.org/abs/1805.09112

  2. [10]

    Hyperbolic graph con- volutional neural networks,

    I. Chami, Z. Ying, C. R´ e, and J. Leskovec, “Hyperbolic graph con- volutional neural networks,” inAdvances in Neural Information Processing Systems 32 (NeurIPS), 2019. [Online]. Available: https://arxiv.org/abs/1910.12933

  3. [11]

    Low- distortion and GPU-compatible tree em- beddings in hyperbolic space,

    M. Spengler and P. Mettes, “Low- distortion and GPU-compatible tree em- beddings in hyperbolic space,” 2025. [Online]. Available: https://arxiv.org/abs/ 2502.17130

  4. [13]

    Unsupervised dense informa- tion retrieval with contrastive learning,

    G. Izacard, M. Caron, L. Hosseini, S. Riedel, P. Bojanowski, A. Joulin, and E. Grave, “Unsupervised dense informa- tion retrieval with contrastive learning,” Transactions on Machine Learning Re- search (TMLR), 2022. [Online]. Available: https://arxiv.org/abs/2112.09118

  5. [14]

    The probabilistic relevance framework: BM25 and beyond,

    S. Robertson and H. Zaragoza, “The probabilistic relevance framework: BM25 and beyond,”Foundations and Trends® in Information Retrieval, vol. 4, no. 1–2, pp. 1–174, 2009. [Online]. Available: https://doi.org/10.1561/1500000019

  6. [15]

    SPLADE: Sparse lexical and ex- pansion model for first stage ranking,

    T. Formal, B. Piwowarski, and S. Clin- chant, “SPLADE: Sparse lexical and ex- pansion model for first stage ranking,” in Proceedings of the 44th International ACM SIGIR Conference on Research and Devel- opment in Information Retrieval (SIGIR), 2021, pp. 2288–2292. [Online]. Av...

  7. [16]

    ColBERT: Efficient and effective passage search via contextualized late interaction over BERT,

    O. Khattab and M. Zaharia, “ColBERT: Efficient and effective passage search via contextualized late interaction over BERT,” inProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2020, pp. 39–48. [Online]. Ava...

  8. [17]

    C-pack: Packaged resources to advance general Chinese embedding,

    S. Xiao, Z. Liu, P. Zhang, N. Muennighoff, D. Lian, and J.-Y. Nie, “C-pack: Packaged resources to advance general Chinese embedding,” inProceedings of the 47th In- ternational ACM SIGIR Conference on Re- search and Development in Information Re- trieval (SIGIR), 2024. [Online]...

  9. [18]

    MS MARCO: A human generated MAchine Reading COm- prehension dataset,

    P. Bajaj, D. Campos, N. Craswell, L. Deng, J. Gao, X. Liu, R. Majumder, A. McNamara, B. Mitra, T. Nguyen, M. Rosenberg, X. Song, A. Stoica, S. Ti- wary, and T. Wang, “MS MARCO: A human generated MAchine Reading COm- prehension dataset,” in30th Conference on Neural Information ...

  10. [19]

    Decou- pled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decou- pled weight decay regularization,” in Proceedings of the 7th International Conference on Learning Representa- tions (ICLR), 2019. [Online]. Available: https://arxiv.org/abs/1711.05101

  11. [20]

    BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models,

    N. Thakur, N. Reimers, A. R¨ uckl´ e, A. Srivastava, and I. Gurevych, “BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models,” inAdvances in Neural Information Process- ing Systems 34 (NeurIPS), Datasets and Benchmarks Track, 2021. [Online]. ...

  12. [21]

    Hybrid hyper- bolic retrieval system for on- device RAG: Source code,

    A. Chakrabarti, “Hybrid hyper- bolic retrieval system for on- device RAG: Source code,” GitHub repository, 2026. [Online]. Avail- able: https://github.com/TimeATronics/ Zetla/tree/main/src/zetla/rag 12

  13. [2010]

    Available: https://jmlr

    [Online]. Available: https://jmlr. org/papers/v11/radovanovic10a.html 10 Table 3: Radial hierarchy verification of HyTE-H embeddings. Positive percentage change indicates that the specific term is encoded further from the origin than its general counterpart. General Specific R...

  14. [2025]

    Available: https://arxiv

    [Online]. Available: https://arxiv. org/abs/2505.24722

Pith tools

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