Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Exploring Task Performance with Interpretable Models via Sparse Auto-Encoders

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that sparse-autoencoder feature analysis can catch an LLM misreading a prompt, and that resolving the flagged ambiguity improves math and metaphor performance.

desk verdict New and useful application of SAEs to prompt reformulation, but the missing ablation leaves the causal claim unsupported. read the letter →

arxiv 2507.06427 v1 pith:A3URAZXM submitted 2025-07-08 cs.CL cs.LG

Shun Wang , Tyler Loakman , Youbo Lei , Yi Liu , Bohao Yang , Yuting Zhao , Dong Yang , Chenghua Lin This is my paper
classification cs.CLcs.LG
keywords sparseautoencodersmechanisticinterpretabilitydictionarylearningmonosemanticfeaturespromptreformulationmathematicalreasoningmetaphordetectionlargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a sparse-autoencoder decomposition of an LLM's internal activations can reveal when the model is misreading a prompt, and that acting on that internal signal improves performance on downstream tasks. On mathematical reasoning and metaphor detection, the authors report consistent accuracy gains across four roughly 7-billion-parameter open models after ambiguous symbols or target words are reformulated with targeted annotations. The core proposal is the trigger: prompts are rewritten only when feature analysis flags an ambiguity, not as a blanket input transformation. The authors position the work as a framework for applying interpretability to real-world tasks rather than a new interpretability technique.

What carries the argument

The load-bearing mechanism is a single-hidden-layer sparse autoencoder with tied weights and an L1 sparsity penalty that decomposes the MLP activations of a frozen LLM into 8192 dictionary features, trained on 12.3B tokens from RedPajama and OpenMathInstruct. An auto-interpretability loop samples text where each dictionary feature activates and has a stronger LLM produce a human-readable description of that feature. A rule-based analysis then inspects the top activated features of mathematical symbols and compares them against a curated set of mathematical attributes; if the dominant features fall outside mathematics, the symbol is flagged as ambiguous. For metaphor detection, an LLM examines the top features of the target word to decide whether it is likely to be misinterpreted. When ambiguity is found, GPT-4 writes a supplementary annotation, Gemini 1.0-Pro checks equivalence with the original prompt, and the reformulated prompt is passed back to the original model.

What would settle it

Run the same pipeline but trigger reformulation on a rule unrelated to feature analysis, such as rephrasing every question containing an absolute-value symbol or rephrasing a randomly selected 20% of metaphor sentences, while keeping the GPT-4 annotations and equivalence checking identical; if accuracy matches the reported enhanced numbers, the improvements cannot be credited to the sparse-autoencoder interpretability.

Watch

Extended reading notes

Core claim

The central claim is that monosemantic features recovered by a sparse autoencoder, labelled through auto-interpretability, can identify genuine model-internal ambiguity in mathematical notation and metaphorical language, and that rephrasing the prompt to resolve that ambiguity improves the model's answers. The paper reports average absolute accuracy gains of 12.52 percentage points (a relative gain of 47.78%) on the MATH dataset and 3.76 percentage points (a relative gain of 5.38%) on metaphor detection across Llama 3, Mistral, Gemma, and Phi-3. It also gives case-level evidence: for the symbol "|", the top activated features are labelled as code language, URLs, and alphanumeric characters rather than mathematics, and for the metaphor target "flowed", the top feature is literal liquid movement; after reformulation the model solves both examples correctly. The authors state explicitly that they are not introducing a new interpretability technique, but a modular framework that applies existing dictionary-learning and auto-interpretability methods to downstream tasks.

Load-bearing premise

The load-bearing premise is that the accuracy gains are caused by the sparse-autoencoder feature analysis, but the paper does not include a control where prompts are rephrased with the same annotations without an SAE-based trigger; if generic rephrasing alone produces equal gains, the interpretability pipeline is not doing the causal work.

Editorial extensions

If this is right

  • Across the four tested roughly 7B models, MATH accuracy rises in every domain, with absolute gains averaging 12.52 percentage points and the largest increases in counting/probability, prealgebra, and algebra.
  • Enhanced metaphor detection outperforms the listed task-specific baselines (MelBERT, MrBERT, RoPPT) on MOH-X and beats them on TroFi for the strongest enhanced model.
  • Because reformulation is triggered only when internal feature analysis detects ambiguity, the pipeline applies external LLM intervention selectively rather than rewriting every input.
  • The semantic-equivalence check by a second model is intended to separate disambiguation from arbitrary question alteration, so that reported gains are attributed to resolving ambiguity.

Reading between the lines

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

  • A natural control the paper does not report would rephrase every prompt, or trigger rephrasing on a trivial heuristic, while keeping GPT-4 annotations identical; if that control matches the reported gains, the SAE analysis would not be the cause of the improvements.
  • The framework could transfer to other ambiguity types, such as polysemous technical terms, pronouns, or notation in code and formal logic, as long as the SAE dictionary is trained on activations from the relevant domain.
  • If feature analysis is the true driver, then dictionary interpretability quality (reported above 93% validation precision) should predict downstream gain per model; a cheap check would be correlating feature-dictionary accuracy with the size of the accuracy lift on the same model.
  • The limitations section itself expects task-dependent variation; discourse-level ambiguities, which are not localized to a symbol or target word, would be a sterner test of whether the mechanism generalizes.
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 trains sparse autoencoders (SAEs) on the MLP activations of four 7B-parameter LLMs (Llama 3, Mistral, Gemma, and Phi-3) to obtain monosemantic feature dictionaries. It then uses these features to detect ambiguities in mathematical symbols and metaphorical target words, and automatically reformulates prompts via GPT-4 to resolve the detected ambiguities. The authors report accuracy improvements on the MATH dataset across all four models (Table 4), with paired t-tests reported in Appendix A, and improvements on the MOH-X and TroFi metaphor datasets across the same models (Table 5). The paper also includes an error analysis of math failures and a discussion of why the method is not merely GPT-4-based prompt reformulation.

Significance. If the causal claim were established, this would be a valuable demonstration that SAE-based internal feature analysis can drive measurable downstream-task gains, going beyond descriptive interpretability and showing a concrete practical use for mechanistically interpretable features. The paper has genuine strengths: it evaluates four open models, reports per-domain mathematics accuracy, provides a statistical test for the math improvements, and includes an informative error analysis categorizing ambiguous symbol types. The limitations section also honestly states that no new interpretability technique is proposed. However, the central causal claim is not yet supported, and the metaphor experiments contain an oracle leak; as presented, the evidence does not distinguish SAE-triggered gains from generic prompt reformulation. Because the missing baselines are both necessary and feasible, the paper has substantial room for revision.

major comments (3)
  1. [§3.3, §5.1, Table 4] The central claim that SAE-based ambiguity detection causes the improvements is untested. There is no ablation that separates the trigger signal from the reformulation content: a condition that rephrases all prompts with GPT-4 (or a random subset) while keeping the reformulation content identical would directly test whether the SAE detection module is necessary. The argument in §3.3 that GPT-4 is 'only triggered when the LLMs-autoencoder predicts a high probability of ambiguity' is not supported by any experiment. Without such a baseline, the consistent gains in Table 4 could be due entirely to GPT-4's generic rephrasing, which would undercut the paper's abstract claim of a method that identifies 'model-internal misunderstanding.'
  2. [§3.2, Table 2, §5.3] The metaphor reformulation leaks the intended answer. The enhanced prompt in Table 2 is: 'The champagne flowed at the wedding. “flowed” implies a free and plentiful availability. Is the target word “flowed” a metaphorical or literal expression?' Since the task is a binary metaphor/literal classification, the inserted gloss states the correct interpretation and is effectively an oracle hint. Any model receiving this gloss would likely classify the target as metaphorical, independent of the SAE features. Accuracy gains on MOH-X and TroFi in Table 5 therefore cannot be attributed to the SAE feature analysis. The authors should use a neutral restatement that clarifies the sense without asserting metaphoricity, or an intervention that does not include the answer. In addition, the metaphor results in §5.3 are reported without significance tests or confidence intervals, and the dataset sizes are not given; the 3–4% point gains may be within sampling noise.
  3. [§3.2, Figure 2, Table 2] The math reformulation may also improve performance for reasons unrelated to the SAE-based detection. The example in Table 2/Figure 2 replaces '|4x + 2|' with 'the absolute value of 4x+2' and 'x < 0' with 'x is less than 0'; such normalization can reduce tokenization and parsing difficulty regardless of whether ambiguity was detected. The paper does not report an ablation that applies the same GPT-4 reformulation to all prompts, nor a condition that triggers rephrasing on a trivial heuristic. Without these, the relative contribution of the detection step versus the rephrasing content cannot be quantified, which is load-bearing for the 'ambiguity resolution' claim in the contributions list.
minor comments (5)
  1. [§5.1, Appendix A] The paired t-test compares only 7 domain-level accuracy values per model (n = 7), which is a very small sample; the paper should justify the independence assumption and report effect sizes or confidence intervals in addition to p-values.
  2. [§5.3] Report the number of test examples in MOH-X and TroFi, and add significance tests or bootstrap confidence intervals for the metaphor improvements.
  3. [Appendix B, Eq. (1)] The value of the sparsity coefficient λ in Eq. (1) is not reported; the reported 'sparsity 1% < 5%' is vague and should be stated exactly.
  4. [Appendix B, Eq. (1)] In Eq. (1), the variables s and x̂ are not defined in the text; clarify that s is the encoder output and x̂ is the decoder reconstruction.
  5. [Author list and notation] The author list contains inconsistent spacing ('Y oubo', 'T oubo'), and the terms 'Tro-Fi' and 'TroFi' are used inconsistently; also state which versions of GPT-4 and Gemini were used and the decoding temperatures for reformulation and explanation generation.

Circularity Check

1 steps flagged · score 6.0 of 10

Metaphor reformulation embeds the gold label, so that half of the reported gain is forced by construction; math gains remain external but confounded.

  1. fitted input called prediction [Section 3.2 (Metaphor Detection) and Table 2]
    "GPT-4 provides targeted supplementary information to clarify the correct meaning of the target word rather than rewriting the entire sentence. ... The champagne flowed at the wedding. 'flowed' implies a free and plentiful availability. Is the target word 'flowed' a metaphorical or literal expression?"

    In this binary metaphor/literal classification task, the 'supplementary information' generated by GPT-4 asserts the metaphorical interpretation ('flowed implies a free and plentiful availability'), which is exactly the gold label the model is supposed to output. The enhanced prompt therefore contains the answer by construction, and the reported accuracy gain is forced by the injected gloss rather than by the sparse-autoencoder feature analysis. Any model that accepts the gloss will classify the target as metaphorical, independent of the internal features.

full rationale

The paper's central causal claim is that sparse-autoencoder feature analysis identifies genuine model-internal misunderstandings and that prompt reformulations triggered by those features improve downstream accuracy. The mathematical-reasoning half is not circular: the reformulations (e.g., expanding |4x+2| to 'the absolute value of 4x+2') do not embed the numerical answer, and the downstream accuracy is measured on the external MATH benchmark. However, the metaphor half does reduce by construction: the reformulated prompt in Table 2 contains the correct metaphorical gloss, which is the label being predicted, so the model's improved metaphor detection is logically forced by the input annotation rather than by the interpretability pipeline. This is a specific, quotable reduction of the type 'fitted input called prediction,' and it warrants a partial-circularity score of 6. There is no load-bearing self-citation chain: the authors' own prior work (FrameBERT, RoPPT) appears only in related work, and no uniqueness theorem is invoked. The limitations section honestly notes that no new interpretability technique is proposed, but it does not flag the answer-leakage problem in the metaphor reformulation. The auto-interpretability validation (GPT-4 explaining and predicting GPT-4-derived features) is self-referential in spirit, but it is not the central downstream claim, so it is not counted as a separate circular step.

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

The central method relies on several hand-chosen hyperparameters and on an unspecified feature curation set. No fundamentally new entities are introduced; the features are derived quantities. The biggest uncontrolled choice is the curated set of mathematical features used to decide ambiguity, which is load-bearing for the math claim.

free parameters (5)
  • Sparsity coefficient lambda in Eq. 1 = not reported
    Controls the trade-off between reconstruction and sparsity, directly shaping which features are extracted. No value is given in Appendix B.
  • Layer choice (typically 12 or 13) = layer 12/13
    Chosen because 'empirically validated' but no comparison data is shown; affects all extracted features.
  • Data mix ratios in Table 3 = e.g., Common Crawl 8.75B, C4 1.75B, etc.
    Proportions chosen by hand; influence the learned dictionary and downstream behavior.
  • Curated set of mathematical attribute features = not specified
    The rule-based ambiguity detector maps each token's top features against this set; the set is never enumerated, so it is an ad hoc selection that determines which symbols are flagged.
  • Ambiguity detection threshold (top activated features) = not specified
    The rule that a symbol is ambiguous if its dominant features 'do not align' with math attributes requires an implicit threshold; no value is given.
assumptions (4)
  • domain assumption Superposition in LLMs means individual neurons are polysemantic, and sparse autoencoders recover monosemantic features.
    The entire method rests on this interpretability hypothesis (Elhage et al. 2022; Bricken et al. 2023).
  • domain assumption GPT-4-generated feature descriptions correctly capture the meaning of SAE features.
    Auto-interpretability (Bills et al.) is used without independent validation; the paper reports 93% validation but the validation protocol is not detailed.
  • domain assumption GPT-4's rephrasing preserves the mathematical content and difficulty of the original question.
    The equivalence check by Gemini 1.0-Pro is used, but the reliability of that check is not established.
  • standard math The paired t-test over 7 math domains is a valid test of the improvement's significance.
    Standard statistical practice, though the small number of paired observations and non-independence of domains is a limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Task Performance with Interpretable Models via Sparse Auto-Encoders." pith.science (2026). https://pith.science/paper/A3URAZXM

@misc{pith2026250706427,
  author       = {Pith},
  title        = {Pith review of: Exploring Task Performance with Interpretable Models via Sparse Auto-Encoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A3URAZXM}},
  note         = {Machine review of arXiv:2507.06427}
}
read the original abstract

Large Language Models (LLMs) are traditionally viewed as black-box algorithms, therefore reducing trustworthiness and obscuring potential approaches to increasing performance on downstream tasks. In this work, we apply an effective LLM decomposition method using a dictionary-learning approach with sparse autoencoders. This helps extract monosemantic features from polysemantic LLM neurons. Remarkably, our work identifies model-internal misunderstanding, allowing the automatic reformulation of the prompts with additional annotations to improve the interpretation by LLMs. Moreover, this approach demonstrates a significant performance improvement in downstream tasks, such as mathematical reasoning and metaphor detection.

Figures

Figures reproduced from arXiv: 2507.06427 by the authors.

Figure 1
Figure 1. Framework of decomposing LLMs by Dictionary Learning. The framework extracts features from LLMs [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the pipeline for handling ambiguity in mathematical problem-solving. (1) Activated features [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 20 canonical work pages

  1. [1]

    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

  2. [2]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, et al. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . ArXiv, abs/2303.08774

    arXiv 2023

  3. [3]

    Janice Ahn, Rishu Verma, Renze Lou, Di Liu, Rui Zhang, and Wenpeng Yin. 2024. https://aclanthology.org/2024.eacl-srw.17 Large language models for mathematical reasoning: Progresses and challenges . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 225--237, St. ...

    work page 2024

  4. [4]

    Zabir Al-Nazi and Wei Peng. 2024. Large language models in healthcare and medical domain: A review. In Informatics, volume 11, page 57. MDPI

    work page 2024

  5. [5]

    Sanjeev Arora, Yuanzhi Li, Yingyu Liang, Tengyu Ma, and Andrej Risteski. 2018. https://doi.org/10.1162/tacl_a_00034 Linear algebraic structure of word senses, with applications to polysemy . Transactions of the Association for Computational Linguistics, 6:483--495

    doi:10.1162/tacl_a_00034 2018

  6. [6]

    Yonatan Belinkov and James Glass. 2019. https://doi.org/10.1162/tacl_a_00254 Analysis methods in neural language processing: A survey . Transactions of the Association for Computational Linguistics, 7:49--72

    doi:10.1162/tacl_a_00254 2019

  7. [7]

    Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. 2023. https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html Language models can explain neurons in language models . OpenAI Blog

    work page 2023

  8. [8]

    Julia Birke and Anoop Sarkar. 2006. https://aclanthology.org/E06-1042 A clustering approach for nearly unsupervised recognition of nonliteral language . In 11th Conference of the E uropean Chapter of the Association for Computational Linguistics , pages 329--336, Trento, Italy. Association for Computational Linguistics

    work page 2006

Show all 49 references
  1. [9]

    Sebastian Bordt, Mich\` e le Finck, Eric Raidl, and Ulrike von Luxburg. 2022. https://doi.org/10.1145/3531146.3533153 Post-hoc explanations fail to achieve their purpose in adversarial contexts . In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transp...

    2022

  2. [10]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. 2023. Towards monosemanticity: Decomposing language models with dictionary learning. Transformer Circuits Thread, 2

    2023

  3. [11]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey W...

    2020

  4. [12]

    Eugene Charniak, Don Blaheta, Niyu Ge, Keith Hall, John Hale, and Mark Johnson. 2000. Bllip 1987-89 wsj corpus release 1. Linguistic Data Consortium, Philadelphia, 36

    2000

  5. [13]

    Minjin Choi, Sunkyung Lee, Eunseong Choi, Heesoo Park, Junhyuk Lee, Dongwon Lee, and Jongwuk Lee. 2021. https://doi.org/10.18653/v1/2021.naacl-main.141 M el BERT : Metaphor detection via contextualized late interaction using metaphorical identification theories . In Proceeding...

    2021 doi

  6. [14]

    Alexis Conneau, German Kruszewski, Guillaume Lample, Lo \" c Barrault, and Marco Baroni. 2018. https://doi.org/10.18653/v1/P18-1198 What you can cram into a single \ & ! \# * vector: Probing sentence embeddings for linguistic properties . In Proceedings of the 56th Annual Meet...

    2018 doi

  7. [15]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600

    2023 arXiv

  8. [16]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  9. [17]

    Finale Doshi-Velez and Been Kim. 2017. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608

    2017 arXiv

  10. [18]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, et al. 2024. http://arxiv.org/abs/2407.21783 The llama 3 herd of models

    2024 arXiv

  11. [19]

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. 2022. Toy models of superposition. arXiv preprint arXiv:2209.10652

    2022 arXiv

  12. [20]

    Manaal Faruqui, Yulia Tsvetkov, Dani Yogatama, Chris Dyer, and Noah A. Smith. 2015. https://doi.org/10.3115/v1/P15-1144 Sparse overcomplete word vector representations . In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th Inte...

    2015 doi

  13. [21]

    Pankaj Gupta and Hinrich Sch \"u tze. 2018. https://doi.org/10.18653/v1/W18-5418 LISA : Explaining recurrent neural network judgments via layer-w I se semantic accumulation and example to pattern transformation . In Proceedings of the 2018 EMNLP Workshop B lackbox NLP : Analyz...

    2018 doi

  14. [22]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/be83ab3ecd0db773eb2dc1b0a17836a1-Paper-round2.pdf Measuring mathematical problem solvin...

    2021

  15. [23]

    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

  16. [24]

    Andrej Karpathy, Justin Johnson, and Li Fei-Fei. 2015. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078

    2015 arXiv

  17. [25]

    Honglak Lee, Alexis Battle, Rajat Raina, and Andrew Ng. 2006. https://proceedings.neurips.cc/paper_files/paper/2006/file/2d71b2ae158c7c5912cc0bbde2bb9d95-Paper.pdf Efficient sparse coding algorithms . In Advances in Neural Information Processing Systems, volume 19. MIT Press

    2006

  18. [26]

    Zachary Levonian, Chenglu Li, Wangda Zhu, Anoushka Gade, Owen Henkel, Millie-Ellen Postle, and Wanli Xing. 2023. Retrieval-augmented generation to improve math question-answering: Trade-offs between groundedness and human preference. arXiv preprint arXiv:2310.03184

    2023 arXiv

  19. [27]

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/fi...

    2022

  20. [28]

    Haitao Li, Qingyao Ai, Jia Chen, Qian Dong, Zhijing Wu, Yiqun Liu, Chong Chen, and Qi Tian. 2024. http://arxiv.org/abs/2403.18365 Blade: Enhancing black-box large language models with small domain-specific models

    2024 arXiv

  21. [29]

    Yucheng Li, Shun Wang, Chenghua Lin, Frank Guerin, and Lo \" c Barrault. 2023. Framebert: Conceptual metaphor detection with frame embedding learning. arXiv preprint arXiv:2302.04834

    2023 arXiv

  22. [30]

    Zichao Lin, Shuyan Guan, Wending Zhang, Huiyan Zhang, Yugang Li, and Huaping Zhang. 2024. Towards trustworthy llms: a review on debiasing and dehallucinating in large language models. Artificial Intelligence Review, 57(9):1--50

    2024

  23. [31]

    Zachary C. Lipton. 2018. https://doi.org/10.1145/3233231 The mythos of model interpretability . Commun. ACM, 61(10):36–43

    2018 doi

  24. [32]

    Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, et al. 2024. http://arxiv.org/abs/2403.08295 Gemma: Open models based on gemini research and technology

    2024 arXiv

  25. [33]

    George A. Miller. 1994. https://aclanthology.org/H94-1111 W ord N et: A lexical database for E nglish . In H uman L anguage T echnology: Proceedings of a Workshop held at P lainsboro, N ew J ersey, M arch 8-11, 1994

    1994

  26. [34]

    Saif Mohammad, Ekaterina Shutova, and Peter Turney. 2016. https://doi.org/10.18653/v1/S16-2003 Metaphor as a medium for emotion: An empirical study . In Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics, pages 23--33, Berlin, Germany. Association...

    2016 doi

  27. [35]

    W James Murdoch, Peter J Liu, and Bin Yu. 2018. Beyond word importance: Contextual decomposition to extract interactions from lstms. arXiv preprint arXiv:1801.05453

    2018 arXiv

  28. [36]

    Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. 2020. Zoom in: An introduction to circuits. Distill, 5(3):e00024--001

    2020

  29. [37]

    B A Olshausen and D J Field. 1997. Sparse coding with an overcomplete basis set: a strategy employed by v1? Vision Res, 37(23):3311--3325

    1997

  30. [38]

    Yasaman Razeghi, Robert L Logan IV, Matt Gardner, and Sameer Singh. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.59 Impact of pretraining term frequencies on few-shot numerical reasoning . In Findings of the Association for Computational Linguistics: EMNLP 2022, pages...

    2022 doi

  31. [39]

    Cynthia Rudin. 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence, 1(5):206--215

    2019

  32. [40]

    Dong Shu, Haoran Zhao, Xukun Liu, David Demeter, Mengnan Du, and Yongfeng Zhang. 2024. Lawllm: Law large language model for the us legal system. arXiv preprint arXiv:2407.21065

    2024 arXiv

  33. [41]

    KV Srivatsa and Ekaterina Kochmar. 2024. What makes math word problems challenging for llms? arXiv preprint arXiv:2403.11369

    2024 arXiv

  34. [42]

    Yuan Tian, Nan Xu, and Wenji Mao. 2024. A theory guided scaffolding instruction framework for llm-enabled metaphor reasoning. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volu...

    2024

  35. [43]

    Shubham Toshniwal, Ivan Moshkov, Sean Narenthiran, Daria Gitman, Fei Jia, and Igor Gitman. 2024. Openmathinstruct-1: A 1.8 million math instruction tuning dataset. arXiv preprint arXiv:2402.10176

    2024 arXiv

  36. [44]

    Shun Wang, Yucheng Li, Chenghua Lin, Loic Barrault, and Frank Guerin. 2023. https://doi.org/10.18653/v1/2023.eacl-main.102 Metaphor detection with effective context denoising . In Proceedings of the 17th Conference of the European Chapter of the Association for Computational L...

    2023 doi

  37. [45]

    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

  38. [46]

    Zheyang Xiong, Ziyang Cai, John Cooper, Albert Ge, Vasilis Papageorgiou, Zack Sifakis, Angeliki Giannou, Ziqian Lin, Liu Yang, Saurabh Agarwal, Grigorios Chrysos, Samet Oymak, Kangwook Lee, and Dimitris Papailiopoulos. 2024. https://doi.org/10.48550/arXiv.2410.05603 Everything...

    2024 doi arXiv

  39. [47]

    Zeyu Yun, Yubei Chen, Bruno Olshausen, and Yann LeCun. 2021. https://doi.org/10.18653/v1/2021.deelio-1.1 Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors . In Proceedings of Deep Learning Inside Out (D...

    2021 doi

  40. [48]

    URL: " 'urlintro :=

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

  41. [49]

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