Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Sparse autoencoders make dense retrieval vectors explainable and indexable.

desk verdict Solid, honest SAE-for-IR paper whose CL-SR retrieval framework deserves peer review, but fix the dev-set tuning and add an out-of-distribution CL-SR experiment. read the letter →

arxiv 2506.00041 v2 pith:5J2YITGC submitted 2025-05-28 cs.IR cs.LG

Seongwan Park , Taeklim Kim , Youngjoong Ko This is my paper
classification cs.IRcs.LG
keywords sparseautoencodersdensepassageretrievalinterpretablelatentconceptsconcept-levelinvertedindexneuralvocabularymismatchinterpretability
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 the dense vectors produced by DPR-style retrievers, long treated as opaque, can be decomposed by a sparse autoencoder into a small set of distinct latent concepts, each carrying a human-readable description. It argues that these concepts preserve enough of the original embedding to keep retrieval nearly intact, and that the same concepts can serve as indexing units in an inverted index, giving Concept-Level Sparse Retrieval (CL-SR). CL-SR is reported to match or beat term-based sparse retrievers such as SPLADE while using substantially less storage and computation, and to stay accurate on queries where lexical matching fails. If the claim holds, dense retrieval gains both transparency and efficiency: search systems can explain their scores in concept terms instead of opaque numbers.

What carries the argument

The load-bearing object is the sparse autoencoder, specifically the BatchTopK variant: a single-layer autoencoder whose hidden dimension is 32 times the embedding size and whose activations are sparsified by retaining only the top-k latents per batch. Each hidden unit, called a latent concept, is a learned direction in embedding space; the encoder maps a DPR vector to a sparse activation vector and the decoder reconstructs the vector. The paper generates a natural-language description for each latent by asking an LLM to summarize the passages that activate it most strongly. CL-SR then reuses those latents as indexing units, scoring query-document pairs with a BM25-style equation in which lexical term frequency is replaced by latent activation, length normalization is computed over total latent activation, and IDF weighting suppresses over-frequent abstract latents.

What would settle it

Run the MSMARCO-trained SAE on an out-of-domain benchmark such as BEIR without retraining; if CL-SR no longer beats term-based sparse baselines on vocabulary-mismatch queries, or if the generated latent descriptions no longer identify the passages that activate them, the central claim of transferable concepts would be refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a BatchTopK sparse autoencoder trained on SimLM dense embeddings of 8.8 million MSMARCO passages learns latent concepts that are individually interpretable: a latent intrusion test reaches 0.859 accuracy at k=32, and natural-language concept descriptions let human annotators pick out source passages 94.3% of the time and predict which of two documents the DPR model would rank higher in over 90% of ranking pairs. Reconstructed embeddings retain most retrieval ability, with MRR@10 falling by about 10-20% depending on sparsity and Spearman rank correlations of the ordered lists near 0.93-0.96. Using the same latent concepts as retrieval units, CL-SR reaches MRR@10 of 0.368 on MSMARCO dev with an index of 0.57-1.27 GB, and on queries where BM25 fails to find the gold passage it degrades far less than term-based sparse methods, nearly matching the dense SimLM model. The paper treats the latent concepts as the semantic vocabulary of the retriever, replacing lexical terms rather than supplementing them.

Load-bearing premise

That the latent concepts learned from MSMARCO embeddings are stable, corpus-independent semantic units that also appear in unseen queries and documents; the paper's own limitations section concedes that SAE features are known to be incomplete and dataset-dependent.

Editorial extensions

If this is right

  • DPR rankings become auditable: every query-document score can be traced to shared latent concepts and their descriptions, which lets a search operator diagnose retrieval failures at the concept level.
  • A dense retriever can be served through an inverted index: CL-SR stores only active latent IDs per passage, giving dense-style semantic coverage at a fraction of the FLOPs and storage of lexical neural sparse methods.
  • The decomposition transfers across dense encoders: retraining the same SAE on TAS-B and GTR-T5 embeddings reproduces the same reconstruction trade-offs, so the approach is tied to the embedding space rather than to SimLM.
  • Vocabulary and semantic mismatch can be handled without explicit query or document expansion because paraphrases that share meaning activate the same latent concepts.

Reading between the lines

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

  • A testable extension the paper leaves implicit is using the latent concept vocabulary as a control surface: boosting or suppressing particular concepts could steer retrieval away from sensitive or biased associations without retraining the dense encoder.
  • If the concepts prove stable across corpora, CL-SR could serve as a transparent first-stage retriever for retrieval-augmented generation, where the activated latent IDs double as human-readable provenance for downstream citations.
  • The paper's own limitations section notes that SAE features are known to be incomplete and dataset-dependent, so the widest open question is whether concepts learned on MSMARCO survive on out-of-domain collections; BEIR-style zero-shot evaluation would settle this.
  • CL-SR could also slot into a two-stage pipeline: its tiny index generates a concept-explainable candidate shortlist, and a cross-encoder re-ranks it, preserving explanation for the final ranking.
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. The paper applies sparse autoencoders (SAEs) to the dense embeddings produced by a DPR-style retriever (SimLM) in order to decompose them into a sparse set of interpretable latent concepts. It generates natural-language descriptions for each latent by summarizing its top-activating MSMARCO passages, and then evaluates interpretability through a latent intrusion test, a human passage-identification task, and a human ranking-simulation task. The paper further introduces Concept-Level Sparse Retrieval (CL-SR), which scores query–document pairs using a BM25-style formula applied to latent activations, and reports retrieval effectiveness, efficiency, and robustness results on MSMARCO Dev and TREC DL 2019/2020. The central claims are that SAE latents are semantically meaningful units, that they make dense retrieval more interpretable, and that CL-SR combines the semantic power of dense retrieval with the efficiency and transparency of sparse indexing.

Significance. If the claims hold, the paper offers a novel and potentially useful bridge between dense retrieval and sparse inverted indexes: interpreting dense embeddings through SAE latents and using those latents as indexing units is an appealing idea with clear practical implications for efficiency and transparency. The paper has several strengths: it evaluates reconstruction fidelity with multiple complementary metrics (NMSE, retrieval preservation, ranking correlation), it includes human studies for interpretability, it compares against standard sparse and neural-sparse baselines, and it reports efficiency metrics (FLOPs, index size, latency). The code is released, which supports reproducibility. However, the strongest empirical claims rest on a single in-distribution collection (MSMARCO), and the paper's own Limitations section concedes that SAE features are dataset-dependent and incomplete. Verifying whether the latent concepts and CL-SR transfer to unseen collections is therefore essential for assessing the impact of the work.

major comments (3)
  1. [§5.2 and Table 5] The robustness claim that CL-SR maintains performance across vocabulary and semantic mismatches is evaluated only on MSMARCO Dev queries with MSMARCO passages, i.e., the same collection and distribution on which the SAE was trained. The Mismatch Set is constructed from MSMARCO Dev queries, and Appendix C transfers the SAE only for reconstruction-based metrics (NMSE, retrieval with reconstructed embeddings) on TREC-COVID and NFCorpus; it never runs CL-SR retrieval on an unseen collection. Since the Limitations section explicitly states that SAE features are dataset-dependent and incomplete, the headline robustness claim is not supported for new corpora. I ask the authors to report CL-SR retrieval results on at least one or two out-of-distribution collections (e.g., BEIR subsets or the already used TREC-COVID/NFCorpus) and to analyze how the activated latent concepts transfer across collections.
  2. [Appendix E and Table 10] The CL-SR hyperparameters (k1, b, k2) are tuned on MSMARCO Dev, and the same split is used for the headline MRR@10 numbers in Table 4 and the comparison in Table 5. This creates a risk of overfitting the reported effectiveness gains, especially because the tuning grid in Table 10 is quite small and no validation/holdout protocol is described. Please specify how the hyperparameters were selected (e.g., a separate validation split, cross-validation, or a fixed grid with a clear criterion) and provide confidence intervals or significance tests for the main retrieval comparisons.
  3. [§4.1–4.2 and Table 3] The human interpretability evaluation is at risk of a self-confirming loop: latent descriptions are generated from the top-30 most activating MSMARCO passages, and the passage-identification task samples 600 passages from the same MSMARCO corpus. If the target passages overlap with the passages used for description generation, high annotator accuracy may reflect the descriptions' memorization of specific source passages rather than stable, generalizable latent semantics. Please report the amount of overlap between the evaluation passages and the description-generation passages, and re-run the passage-identification task on a held-out set of passages that were never used for generating descriptions, ideally including passages from an unseen dataset.
minor comments (5)
  1. [Introduction] There is a typo in the Introduction: "we futher introduce" should read "we further introduce."
  2. [§5, Eq. (5)] Equation (5) writes the query term as fd(q, i), but the following definitions define fq(q, i) and fd(d, i); please correct the notation so that the query and document factors are distinguishable.
  3. [§3.1, Eq. (4)] Equation (4) appears to omit the input vector h inside the BatchTopK operation; it should presumably read BatchTopK(W_enc h + b_enc) to match Eq. (1).
  4. [Table 1] The table caption states that Spearman's correlation is averaged across TREC DL 2019 and 2020 with variance reported and p < 0.05, but the main text does not describe the significance-testing procedure; please add a brief description or a reference to an appendix.
  5. [Table 2] The latent intrusion test is conducted with an LLM (GPT-4.1 mini) rather than human annotators; the paper should clarify whether the reported accuracy is meant as a proxy for human interpretability and, if so, acknowledge that LLM-based evaluation may not perfectly match human judgments.

Circularity Check

1 steps flagged · score 3.0 of 10

Interpretability evaluation is partly self-confirming (descriptions fit to the same MSMARCO corpus used in the passage-identification test), but the CL-SR retrieval claim is an independent, externally benchmarked empirical result.

  1. fitted input called prediction [Section 4.1-4.2 and Appendix B.1-B.2 (latent description generation and passage-identification evaluation)]
    "For each latent that has been activated at least once over the whole MSMARCO corpus, we provide top 30 most activating MSMARCO passage and instruct GPT4.1-mini to generate description for the latent. ... We randomly sampled 600 passages from the MSMARCO corpus and evaluated the embedding interpretability task and report the accuracy in Table 3."

    The natural-language descriptions are fitted to per-latent sets of top-activating MSMARCO passages, and the passage-identification test then uses those same descriptions to identify a target sampled from the same MSMARCO corpus. No held-out split or exclusion of the 600 evaluation passages from the description-generation pool is reported, so any evaluation passage that is itself among a latent's top-30 activators is identified using a description constructed from that very passage. The reported 0.943 accuracy therefore partly reflects the descriptions memorizing the evaluation pool rather than independently demonstrating that the SAE latents are stable, generalizable semantic units.

full rationale

The central CL-SR retrieval claim is not circular: it is an empirical pipeline (SAE on SimLM embeddings, latent activations, BM25-style scoring) evaluated against external sparse and dense baselines on MSMARCO Dev and TREC DL 2019/2020. The CL-SR scoring function (Eq. 5) is not identical to the SAE reconstruction objective or to SimLM dot-product similarity, so the retrieval results do not reduce to the training inputs by construction. There are no load-bearing self-citations (the reference list contains no author self-citations) and no uniqueness claim imported from prior work by the same authors. The one credible circularity is in the interpretability study: the descriptions are generated from top-activating MSMARCO passages, and the passage-identification evaluation draws its targets from the same corpus without any reported exclusion, so high accuracy may partly reflect self-memorization rather than general interpretability. The Limitations section honestly concedes that SAE features are 'incomplete' and 'dataset dependent', which limits transferability but is a correctness limitation rather than a circular derivation. Likewise, the robustness claim on the BM25-failure Mismatch Set is evaluated only on in-distribution MSMARCO Dev; that is a missing external-validation issue, not circularity. Overall, the retrieval contribution stands independently, while the headline interpretability result is partially self-confirming.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework rests on a few strong assumptions from the mechanistic interpretability literature (sparse autoencoders recover true features) and on several design choices specific to this paper (BM25-style scoring over latent activations, activation thresholds, and latent indexing limits). The free parameters are mostly hyperparameters; the CL-SR scoring parameters are fitted to the evaluation split. No new physical or abstract entities are introduced beyond the learned latent concepts, which are model-internal features rather than postulated entities.

free parameters (5)
  • k (number of activated latents in SAE) = 32, 48, 64, 128 (experimented)
    Controls the sparsity-reconstruction tradeoff; all downstream results depend on this choice (Section 3.1, Table 1).
  • m (SAE hidden dimension) = 32 * d (d = embedding dimension)
    Hidden layer size chosen as a multiple of the input dimension following common practice (Section 3.1).
  • CL-SR BM25 parameters (k1, b, k2) = Varies with k; see Table 10 (e.g., k=32: 0.6, 1.75, 2.0)
    Tuned on MSMARCO Dev for each configuration; these directly affect the ranking formula (Equation 5).
  • Max latent concepts indexed per passage = 24 (Efficient), 65 (Max)
    Chosen from Figure 3 to balance efficiency and effectiveness (Section 5.1).
  • Activation threshold for latent binarization = Mean of top k-th activation over the dataset
    Determines which latents count as active at inference; a design choice with no theoretical grounding (Section 3.1, Appendix B.1).
assumptions (4)
  • domain assumption Sparse autoencoders recover meaningful interpretable features from neural network activations (superposition hypothesis).
    Invoked in Sections 2.2 and 3.1; the entire interpretability claim rests on this premise from prior LLM interpretability work.
  • domain assumption The DPR embedding space is approximately linearly decomposable into sparse, semantically coherent concepts.
    Assumed when training the SAE (Equations 1-4); no verification that SimLM embeddings satisfy this structure.
  • domain assumption LLM-generated descriptions from top-activating passages accurately summarize each latent concept.
    Used in Section 4.1 and throughout the human interpretability tasks; flawed descriptions would invalidate the interpretation results.
  • ad hoc to paper BM25-style scoring applied to latent activations (Equation 5) is a valid retrieval model.
    Introduced in Section 5 without theoretical justification; the authors tune its parameters but do not derive the functional form.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval." pith.science (2026). https://pith.science/paper/5J2YITGC

@misc{pith2026250600041,
  author       = {Pith},
  title        = {Pith review of: Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5J2YITGC}},
  note         = {Machine review of arXiv:2506.00041}
}
read the original abstract

Despite their strong performance, Dense Passage Retrieval (DPR) models suffer from a lack of interpretability. In this work, we propose a novel interpretability framework that leverages Sparse Autoencoders (SAEs) to decompose previously uninterpretable dense embeddings from DPR models into distinct, interpretable latent concepts. We generate natural language descriptions for each latent concept, enabling human interpretations of both the dense embeddings and the query-document similarity scores of DPR models. We further introduce Concept-Level Sparse Retrieval (CL-SR), a retrieval framework that directly utilizes the extracted latent concepts as indexing units. CL-SR effectively combines the semantic expressiveness of dense embeddings with the transparency and efficiency of sparse representations. We show that CL-SR achieves high index-space and computational efficiency while maintaining robust performance across vocabulary and semantic mismatches.

Figures

Figures reproduced from arXiv: 2506.00041 by the authors.

Figure 1
Figure 1. Overview of our method. We first train a SAE to decompose dense embeddings into latent concepts [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Top-5 most activated latent concepts extracted [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance vs FLOPs for CL-SR with vary [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative case study illustrating CL-SR’s ability to handle vocabulary and semantic mismatch. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Frequency distribution of SAE latents across [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: LLM prompt templates for our two tasks: (a) [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Impact of IDF reweighting on an MSMARCO passage—suppressing abstract latents (8033, 3486) and [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Example of human evaluation for dense embeddding interpretability. The left column lists the top-activated [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Example of human evaluation for ranking interpretability. The first column shows the query and its [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 10 canonical work pages

  1. [1]

    Avishek Anand, Lijun Lyu, Maximilian Idahl, Yumeng Wang, Jonas Wallat, and Zijian Zhang. 2022. https://arxiv.org/abs/2211.02405 Explainable information retrieval: A survey . arXiv preprint arXiv:2211.02405

    arXiv 2022

  2. [2]

    Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2018. https://arxiv.org/abs/1611.09268 Ms marco: A human generated machine reading comprehension dataset . arXiv preprint arXiv:1611.09268. Version 3

    arXiv 2018

  3. [3]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Ch...

    2023

  4. [4]

    Bart Bussmann, Patrick Leask, and Neel Nanda. 2024. https://arxiv.org/abs/2412.06410 Batchtopk sparse autoencoders . arXiv preprint arXiv:2412.06410. ArXiv:2412.06410 [cs.LG]

    arXiv 2024

  5. [5]

    Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: how humans interpret topic models. In Proceedings of the 23rd International Conference on Neural Information Processing Systems, NIPS'09, page 288–296, Red Hook, NY, USA. Curran Associates Inc

    work page 2009

  6. [6]

    Bruce Croft, Maarten de Rijke, Jiafeng Guo, and Bhaskar Mitra

    Nick Craswell, W. Bruce Croft, Maarten de Rijke, Jiafeng Guo, and Bhaskar Mitra. 2018. https://doi.org/10.1007/s10791-017-9323-9 Neural information retrieval: Introduction to the special issue . Information Retrieval Journal, 21(2):107--110

    work page doi:10.1007/s10791-017-9323-9 2018

  7. [7]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. https://arxiv.org/abs/2102.07662 Overview of the trec 2020 deep learning track . arXiv preprint arXiv:2102.07662. Version 1

    arXiv 2021

  8. [8]

    Voorhees

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. https://arxiv.org/abs/2003.07820 Overview of the trec 2019 deep learning track . arXiv preprint arXiv:2003.07820. Version 2

    arXiv 2020

Show all 43 references
  1. [9]

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. https://transform...

    2022

  2. [10]

    Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stéphane Clinchant. 2021 a . https://arxiv.org/abs/2109.10086 Splade v2: Sparse lexical and expansion model for information retrieval . Preprint, arXiv:2109.10086. ArXiv:2109.10086 [cs.IR]

    2021 arXiv

  3. [12]

    Leo Gao, Tom Dupre la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2025. https://openreview.net/forum?id=tcsZt9ZNKD Scaling and evaluating sparse autoencoders . In The Thirteenth International Conference on Learning Rep...

    2025

  4. [13]

    Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. https://doi.org/10.48550/arXiv.2104.07186 Coil: Revisit exact lexical match in information retrieval with contextualized inverted list . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Com...

    2021 doi arXiv

  5. [14]

    Sebastian Hofst\" a tter, Sheng-Chieh Lin, Jheng-Hong Yang, Jimmy Lin, and Allan Hanbury. 2021. https://doi.org/10.1145/3404835.3462891 Efficiently teaching an effective dense retriever with balanced topic aware sampling . In Proceedings of the 44th International ACM SIGIR Con...

    2021

  6. [15]

    Chao-Wei Huang and Yun-Nung Chen. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1013 Pairdistill: Pairwise relevance distillation for dense retrieval . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18225--18237, Miami, Flo...

    2024 doi

  7. [16]

    Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. 2024. https://openreview.net/forum?id=F76bwRSLeK Sparse autoencoders find highly interpretable features in language models . In The Twelfth International Conference on Learning Representations

    2024

  8. [17]

    Hao Kang, Tevin Wang, and Chenyan Xiong. 2025. https://aclanthology.org/2025.naacl-short.58/ Interpret and control dense retrieval with sparse latent features . In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Li...

    2025

  9. [18]

    Connor Kissane, robertzk, Neel Nanda, and Arthur Conmy. 2024. https://www.lesswrong.com/posts/rtp6n7Z23uJpEH7od/saes-are-highly-dataset-dependent-a-case-study-on-the Saes are highly dataset dependent: a case study on the refusal direction

    2024

  10. [19]

    Carlos Lassance, Herv \'e D \'e jean, Thibault Formal, and St \'e phane Clinchant. 2024. https://arxiv.org/abs/2403.06789 Splade-v3: New baselines for splade . arXiv preprint arXiv:2403.06789

    2024 arXiv

  11. [20]

    Victor Lavrenko and W Bruce Croft. 2001. https://doi.org/10.1145/383952.383970 Relevance based language models . In Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, pages 120--127. ACM

    2001

  12. [21]

    Patrick Leask, Bart Bussmann, Michael T Pearce, Joseph Isaac Bloom, Curt Tigges, Noura Al Moubayed, Lee Sharkey, and Neel Nanda. 2025. https://openreview.net/forum?id=9ca9eHNrdH Sparse autoencoders do not find canonical units of analysis . In The Thirteenth International Confe...

    2025

  13. [22]

    Minghan Li, Sheng-Chieh Lin, Barlas Oguz, Asish Ghoshal, Jimmy Lin, Yashar Mehdad, Wen-tau Yih, and Xilun Chen. 2022. https://arxiv.org/abs/2211.10411 Citadel: Conditional token interaction via dynamic lexical routing for efficient and effective multi-vector retrieval . arXiv ...

    2022 arXiv

  14. [23]

    Jimmy Lin and Xueguang Ma. 2021. https://arxiv.org/abs/2106.14807 A few brief notes on deepimpact, coil, and a conceptual framework for information retrieval techniques . Preprint, arXiv:2106.14807. ArXiv:2106.14807 [cs.IR]

    2021 arXiv

  15. [24]

    Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep, and Rodrigo Nogueira. 2021. https://doi.org/10.1145/3404835.3463238 Pyserini: A python toolkit for reproducible information retrieval research with sparse and dense representations . In Proceedings of the...

    2021

  16. [25]

    Michael Llordes, Debasis Ganguly, Sumit Bhatia, and Chirag Agarwal. 2023. https://doi.org/10.1145/3539618.3591982 Explain like i am bm25: Interpreting a dense model's ranked-list with a sparse approximation . In Proceedings of the 46th International ACM SIGIR Conference on Res...

    2023

  17. [26]

    Ilya Loshchilov and Frank Hutter. 2019. https://openreview.net/forum?id=Bkg6RiCqY7 Decoupled weight decay regularization . In International Conference on Learning Representations

    2019

  18. [27]

    Antonio Mallia, Omar Khattab, Torsten Suel, and Nicola Tonellotto. 2021. https://doi.org/10.1145/3404835.3463030 Learning passage impacts for inverted indexes . In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (...

    2021

  19. [28]

    Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. 2025. https://openreview.net/forum?id=I4e82CIDxv Sparse feature circuits: Discovering and editing interpretable causal graphs in language models . In The Thirteenth International Conferenc...

    2025

  20. [29]

    Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith Hall, Ming-Wei Chang, and Yinfei Yang. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.669 Large dual encoders are generalizable retrievers . In Proceedings of the 2022 Con...

    2022 doi

  21. [30]

    Rodrigo Nogueira and Jimmy Lin. 2019. From doc2query to doctttttquery. arXiv preprint arXiv:1904.08375

    2019 arXiv

  22. [31]

    Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. 2020. https://doi.org/10.23915/distill.00024.001 Zoom in: An introduction to circuits . Distill. Https://distill.pub/2020/circuits/zoom-in

    2020 doi

  23. [32]

    Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Kramár, and Neel Nanda. 2024. https://arxiv.org/abs/2407.14435 Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders . Preprint, arXiv:2407.14435

    2024 arXiv

  24. [33]

    Robert\_AIZI. 2024. https://www.lesswrong.com/posts/BduCMgmjJnCtc7jKc/research-report-sparse-autoencoders-find-only-9-180-board Research report: Sparse autoencoders find only 9/180 board state features in othellogpt

    2024

  25. [34]

    Michaud, Stephen Casper, Max Tegmark, William Saunders, David Bau, Eric Todd, Atticus Geiger, Mor Geva, Jesse Hoogland, Daniel Murfet, and Tom McGrath

    Lee Sharkey, Bilal Chughtai, Joshua Batson, Jack Lindsey, Jeff Wu, Lucius Bushnaq, Nicholas Goldowsky-Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, Stella Biderman, Adria Garriga-Alonso, Arthur Conmy, Neel Nanda, Jessica Rumbelow, Martin Wattenberg, Nandi Schoots, ...

    2025 arXiv

  26. [35]

    Daniel Freeman, Theodore R

    Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner, Callum McDougall, Monte MacDiarmid, C. Daniel Freeman, Theodore R. Sumers, Edward Rees, Joshua ...

    2024

  27. [36]

    o lske, Alexander Bondarenko, Maik Fr \

    Michael V \"o lske, Alexander Bondarenko, Maik Fr \"o be, Benno Stein, Jaspreet Singh, Matthias Hagen, and Avishek Anand. 2021. https://doi.org/10.1145/3471158.3472256 Towards axiomatic explanations for neural ranking models . In Proceedings of the 2021 ACM SIGIR International...

    2021

  28. [37]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2023 a . https://doi.org/10.18653/v1/2023.acl-long.125 Simlm: Pre-training with representation bottleneck for dense passage retrieval . In Proceedings of the 61st Annua...

    2023 doi

  29. [38]

    Liang Wang, Nan Yang, and Furu Wei. 2023 b . https://doi.org/10.18653/v1/2023.emnlp-main.585 Query2doc: Query expansion with large language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9414--9423, Singapore. Associat...

    2023 doi

  30. [39]

    Ho, Chao Zhang, and Carl Yang

    Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Yanqiao Zhu, May Dongmei Wang, Joyce C. Ho, Chao Zhang, and Carl Yang. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1241 Bmretriever: Tuning large language models as better biomedical text retrievers . In Proceedings of the 2024 C...

    2024 doi

  31. [40]

    Christine Ye, Charles O'Neill, John F Wu, and Kartheik G. Iyer. 2024. https://openreview.net/forum?id=oacksuh7Tu Steering semantic search with interpretable features from sparse autoencoders . In MINT: Foundation Model Interventions

    2024

  32. [41]

    Zhang, and Shaoping Ma

    Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, M. Zhang, and Shaoping Ma. 2021. https://api.semanticscholar.org/CorpusID:244715082 Interpreting dense retrieval as mixture of topics . ArXiv, abs/2111.13957

    2021 arXiv

  33. [42]

    Wayne Xin Zhao, Jing Liu, Ruiyang Ren, and Ji‐Rong Wen. 2024. https://doi.org/10.1145/3637870 Dense text retrieval based on pretrained language models: A survey . ACM Transactions on Information Systems, 42(4):1--60

    2024 doi

  34. [43]

    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...

  35. [44]

    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 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.