Pith. sign in

REVIEW 4 major objections 4 minor 28 references

SAFR: Neuron Redistribution for Interpretability

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read SAFR's regularizers make neurons align with word importance.

desk verdict A clean, simple regularization idea with a circular evaluation: the SRS metric largely re-states the training objective, so the interpretability claim needs external validation. read the letter →

arxiv 2501.16374 v2 pith:XEQFRXPU submitted 2025-01-23 cs.LG cs.AI

Ruidi Chang , Chunyuan Deng , Hanjie Chen This is my paper · ORCID
classification cs.LGcs.AI
keywords superpositionmonosemanticitypolysemanticityinterpretabilityregularizationtransformerneuroncapacityVMASK
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a transformer's interpretability can be improved by directly steering how features are spread across neurons during training. It adds two simple terms to the loss: one that makes neurons monosemantic for tokens the model deems important, and one that makes neurons polysemantic for token pairs that attend to each other. On SST-2 and IMDB, the resulting model keeps accuracy near baseline while much larger accuracy drops occur when high-capacity tokens are removed, which the paper takes as evidence that the network's internal capacity aligns with token importance. If the claim holds, it suggests that interpretability can be engineered as a training objective rather than analyzed only after the fact.

What carries the argument

The key machinery is the augmented loss $L = L_{\text{CE}} + \lambda_{\text{Imp}}L_{\text{Importance}} + \lambda_{\text{Inter}}L_{\text{Interaction}}$, where $L_{\text{Importance}}$ penalizes the square root of a token's polysemanticity and $L_{\text{Interaction}}$ penalizes attention-weighted complement of interference. These quantities are defined on hidden representations: interference is the dot product between two token representations, polysemanticity measures how much one representation overlaps all others, and capacity measures the fraction of a representation direction devoted to its own token. Together the two regularizers apply opposing pressures on the same representational space, concentrating capacity on important tokens while allowing correlated tokens to share directions.

What would settle it

Compare the tokens that SAFR assigns the highest capacity against tokens humans mark as important on a small annotated sample of SST-2 or IMDB reviews, or against gradient- and attention-based importance scores; if the capacity-ranked tokens disagree with human rationales or with other importance measures, then SRS is not a valid proxy for interpretability and the paper's central claim would be unsupported. Alternatively, train SAFR with shuffled labels and check whether SRS still rises, which would indicate the metric is capturing training signal rather than meaningful semantic alignment.

Watch

Extended reading notes

Core claim

SAFR is a training-time regularization that redistributes representational capacity across neurons. Important tokens, identified by a variational mask (VMASK), are penalized for having high polysemanticity, forcing their representations to occupy distinct directions; correlated token pairs, identified through attention weights, are penalized for having low interference, pulling their representations into shared dimensions. The result is that in the first feedforward layer, capacity scores become a readable map of token importance: the top 30% of VMASK-important tokens carry about twice the average capacity on SST-2 (0.57 vs 0.30) and more than the average on IMDB (0.20 vs 0.14). The paper reports SRS scores of 17.21 and 28.48 for SST-2 and IMDB, against 4.00 and 3.54 for a baseline transformer, which it interprets as substantially improved interpretability with prediction accuracy essentially unchanged.

Load-bearing premise

The load-bearing premise is that the SRS score actually measures interpretability: if high capacity on VMASK-important tokens only reflects that the model was trained to allocate capacity there, the reported drop in accuracy after deleting those tokens may not mean a human can understand the model's decisions any better.

Editorial extensions

If this is right

  • If SRS is accepted as a proxy for interpretability, SAFR shows that interpretability can be improved by two additive loss terms without requiring special architectures or post-hoc analysis.
  • The reported capacity statistics indicate that VMASK-important tokens receive measurably higher capacity in the expanded feedforward layer, consistent with the intended redistribution.
  • The method yields a visualization: neuron allocation in the first fully connected layer can be read as a token importance map, with larger circles for higher capacity and colored lines for positive or negative interference.
  • The ablation of hyperparameters shows that both regularizer terms are needed; removing either one or setting either weight too high degrades the SRS gain or the accuracy.
  • Because the model uses random embeddings, the regularization is shown to work on purely learned structure from scratch, suggesting the effect does not depend on pretrained representations.

Reading between the lines

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

  • If SRS is validated against human rationales, SAFR could become a cheap alternative to sparse autoencoders for making small transformers explainable.
  • The same capacity, interference, and polysemanticity metrics could be used as diagnostic tools for larger models, even where the regularization itself is too expensive to apply.
  • A stress test with pretrained embeddings might shift the trade-off between monosemanticity and polysemanticity, potentially making SAFR either more or less effective than the random-embedding setting reported here.
  • The paper's SRS measures alignment with the model's own capacity allocation; whether that equals human interpretability is an untested assumption that would need direct evidence.
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 / 4 minor

Summary. The paper proposes SAFR (Superposition-Aware Feature Regularization), a training-time method that adds two regularizers to the cross-entropy loss of a transformer classifier. The importance regularizer penalizes polysemanticity for tokens deemed important by a VMASK layer, while the interaction regularizer encourages polysemanticity among token pairs with high attention weights. The authors evaluate on SST-2 and IMDB using a single-layer transformer with random embeddings. Their main evidence is the Superposition Regularization Score (SRS), defined as the drop in accuracy after deleting the k% of tokens with highest capacity. They report SRS values of 17.21 and 28.48 for SAFR versus 4.00 and 3.54 for the baseline, concluding that SAFR improves interpretability while maintaining accuracy. They also provide qualitative visualizations of neuron capacity and interference in the FC1 layer.

Significance. If the claimed effect is real, the paper would contribute a simple and inexpensive way to bias neuron allocation in transformers toward monosemantic representations of important tokens, with code provided and a clear, honest limitation section. The central interpretability claim, however, rests entirely on the SRS metric, which has not been validated as an interpretability measure. The training objective directly optimizes the alignment between capacity and VMASK importance, so the SRS gain may largely restate that the training signal was followed. The paper also selects hyperparameters on the test set and reports no error bars, making even the magnitude of the effect uncertain. The central claim is therefore not yet established, although it is defensible with additional experiments or with a more modest framing.

major comments (4)
  1. [§4 (SRS definition) and §5 (Table 1)] The SRS metric is not established as a measure of interpretability. Because L_Importance explicitly penalizes polysemanticity for VMASK-important tokens, the highest-capacity tokens are by construction the tokens the model was trained to treat as important; a larger SRS for SAFR therefore partly restates that the training objective was followed. The paper does not compare capacity rankings with human rationales, gradient-based attribution, attention, or LIME, and the authors' own Limitation section acknowledges the need for "more comprehensive and standardized evaluation metrics." Without such external validation, the central claim that SAFR demonstrates improved interpretability is unsupported.
  2. [§4 (Experimental Setup) and Table 6] The hyperparameters λImp and λInter are tuned on the test set: the reported SAFR results in Table 1 are the best SRS values from the grid in Table 6, with the selected λs stated in the Table 1 caption. This is a form of test-set fitting and biases the reported SRS gains upward. The authors should use a separate validation split for hyperparameter selection and report the corresponding test performance.
  3. [Tables 1, 2, and 6] All accuracy and SRS numbers are single runs without error bars or significance tests. Given the small differences (e.g., 4.00 vs. 8.12 SRS between baseline and λImp=λInter=0 on SST-2) and the use of random embeddings and a single-layer transformer, these differences may be within run-to-run noise. Report means and standard deviations over multiple seeds.
  4. [§3 (Interaction-Based Regularization)] The term I^{Aα}_{i,j} in L_Interaction is described only as "the Interference of the attention weights matrix," but Eq. (1) defines interference for hidden representations, not for attention weight matrices. The loss is therefore not reproducible as written; the authors should define precisely how I^{Aα}_{i,j} is computed from Aα (e.g., a dot product of rows, or a separate interference matrix). Similarly, P^V_i in L_Importance should be pinned down as the Eq. (2) polysemanticity computed at the VMASK output layer.
minor comments (4)
  1. [§3, §5, Appendix C] There are several typos: "Interf erence" in Section 3, "Accross" in Appendix C, and an unexplained "a" in the Figure 3 legend. "BaseLine" should be capitalized consistently.
  2. [Table 3] The table reports average capacity, but it is unclear whether these are per-token averages over the test set and whether the important/less-important split uses VMASK scores computed on the test set itself; clarify the computation and note that the split is in-sample.
  3. [§5 (first paragraph)] The sentence "even without regularization (λImp = 0, λInter = 0), the VMASK layer before the Transformer block enhances SRS scores" is potentially confusing because VMASK is itself trained with the model; consider clarifying that "without regularization" refers to the absence of the SAFR terms.
  4. [References] The reference "A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems." is missing volume/page details and the full author list; the citation to Goh (2016) also lacks a venue.

Circularity Check

2 steps flagged · score 6.0 of 10

SRS measures the same capacity-importance alignment that SAFR is trained to produce, and the headline SRS values are selected by test-set hyperparameter search; the interpretability claim largely restates the training objective.

  1. self definitional [Section 3 (Importance-Based Regularization), Section 4 (Evaluation), Eqs. (1)-(3), Table 1]
    "To encourage monosemanticity for important tokens, we introduce a regularization term LImportance = 1/T Σ sqrt(P^V_i / E) ... This regularization penalizes important tokens with high polysemanticity. ... By deleting the top k% of tokens based on capacity, SRS calculates the average change in the prediction accuracy ... Higher SRS values indicate that the removed words were highly important, signifying stronger superposition regularization."

    Capacity (Eq. 3) is monotonically decreasing in polysemanticity (Eq. 2), so penalizing P^V_i for VMASK-important tokens is a direct pressure to concentrate capacity on those tokens. SRS then removes the highest-capacity tokens and reports the accuracy drop as 'interpretability.' For SAFR, a large SRS is the expected footprint of the training signal—capacity placed on tokens the model was told are important—rather than an independent demonstration that a human can understand the decisions. No external validation (human rationales, gradients, attention, LIME) is given to show that SRS tracks interpretability.

  2. fitted input called prediction [Appendix D (Table 6) and Table 1 caption]
    "This section presents the results of hyperparameter tuning for λImp and λInter , as summarized in Table 6. ... The optimal parameter settings for SAFR are λImp = 0.1, λInter = 0.1 for SST-2 and λImp = 0.1, λInter = 1for IMDB."

    Table 6 reports SRS for the full λ grid on the same SST-2 and IMDB test sets used in Table 1, and Table 1 quotes the settings with the highest SRS as 'optimal.' The headline values 17.21 and 28.48 are therefore the maximum of a test-set search, i.e., a fitted quantity rather than an independent prediction. This makes the claimed improvement in interpretability partly an artifact of selection on the evaluation metric itself.

full rationale

The central evidence for 'improved interpretability' is the SRS score (Section 4; Table 1). SRS is defined as the accuracy drop after deleting the top-capacity tokens. Capacity (Eq. 3) is derived from the same interference/polysemanticity quantities (Eqs. 1-2) that the importance regularizer minimizes for VMASK-important tokens. Training with L_Importance therefore pushes capacity onto the tokens VMASK labels important; SRS then removes those tokens and reports a large drop. The improvement in SRS is the expected consequence of the training signal, not an independent confirmation that human users can interpret the model. The paper does not compare SRS with human rationales, gradients, attention, or LIME, and its own Limitation section concedes that 'more comprehensive and standardized evaluation metrics' are needed. Compounding this, the 'optimal' λ values in Table 1 are selected from Table 6, which reports SRS on the same test sets, so the headline numbers are the best of a test-set search. The VMASK self-citation (Chen and Ji, 2020) is independently published and does not itself raise the score, but it supplies the importance labels that close the loop between the regularizer and the metric. Because accuracy is roughly preserved and the FC1 capacity effect is not literally identical to the layer where L_Importance is applied, the paper is not a pure tautology; still, the interpretability claim reduces substantially to the objective being optimized, yielding a score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central evaluation rests on two domain assumptions about what the hand-defined quantities (capacity, SRS, attention correlation) mean, and on two fitted hyperparameters selected on the test set. No new physical or theoretical entities are introduced.

free parameters (3)
  • lambdaImp (importance regularization weight) = 0.1 for both SST-2 and IMDB
    Chosen from the ablation table based on test-set SRS, not a held-out validation set.
  • lambdaInter (interaction regularization weight) = 0.1 for SST-2, 1 for IMDB
    Chosen to maximize test-set SRS in Table 6; this selection on the test set is a form of parameter fitting.
  • k (top token percentage for SRS) = 30%
    Evaluation parameter for deleting tokens; the paper does not justify why 30% is the headline number.
assumptions (4)
  • domain assumption Interference, polysemanticity, and capacity as defined in Eqs. 1-3 on token representations are meaningful measures of superposition and interpretability.
    Section 2 extends feature-level definitions from Scherlis et al. and Elhage et al. to per-token representations; no validation that these quantities correspond to human-interpretable structure.
  • domain assumption SRS, the accuracy drop after deleting top-capacity tokens, is a valid proxy for model interpretability.
    Introduced in Section 4; the paper states 'Higher SRS values indicate that the removed words were highly important, signifying stronger superposition regularization', which equates a deletion test with interpretability without external validation.
  • domain assumption Attention weights identify correlated token pairs whose polysemanticity should be encouraged.
    Used in Section 3 LInteraction; attention is known to be a controversial importance/correlation signal (Jain and Wallace, 2019), and the paper itself notes debates about attention-based interpretation.
  • domain assumption VMASK importance scores are an appropriate ground truth for token importance.
    VMASK is adopted from Chen and Ji (2020) and is from the same group as this paper; the regularization and evaluation both lean on this importance signal without comparing to other saliency methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAFR: Neuron Redistribution for Interpretability." pith.science (2026). https://pith.science/paper/XEQFRXPU

@misc{pith2026250116374,
  author       = {Pith},
  title        = {Pith review of: SAFR: Neuron Redistribution for Interpretability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XEQFRXPU}},
  note         = {Machine review of arXiv:2501.16374}
}
read the original abstract

Superposition refers to encoding representations of multiple features within a single neuron, which is common in deep neural networks. This property allows neurons to combine and represent multiple features, enabling the model to capture intricate information and handle complex tasks. Despite promising performance, the model's interpretability has been diminished. This paper presents a novel approach to enhance model interpretability by regularizing feature superposition. We introduce SAFR, which simply applies regularizations to the loss function to promote monosemantic representations for important tokens while encouraging polysemanticity for correlated token pairs, where important tokens and correlated token pairs are identified via VMASK and attention weights respectively. We evaluate SAFR with a transformer model on two classification tasks. Experiments demonstrate the effectiveness of SAFR in improving model interpretability without compromising prediction performance. Besides, SAFR provides explanations by visualizing the neuron allocation within the intermediate layers.

Figures

Figures reproduced from arXiv: 2501.16374 by the authors.

Figure 1
Figure 1. Basic structure of SAFR. i) Promote monosemanticity for important tokens after the embedding layers. ii) Leverage the attention mechanism to enhance polysemanticity among correlated token pairs. degree of polysemanticity, significantly improving interpretability while maintaining prediction per￾formance at comparable levels. 2 Preliminaries Given an input sequence of T tokens S = ⟨s1, . . . , sT ⟩ and a neural netwo… view at source ↗
Figure 3
Figure 3. Sensitivity to k Selection. As tokens are gradually removed, accuracy declines consistently. Average Capacity Across the Vocabulary. Ta￾ble 3 presents an analysis of the average capacity per token status (important or not w.r.t VMASK) on the entire test set. Tokens are categorized based on VMASK scores, with the top 30% identified as important and the rest 70% as less important. These results demonstrate SAFR’s effe… view at source ↗
Figure 2
Figure 2. (a) Important tokens exhibit higher capacity. (b) Circle size represents capacity, with larger circles [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Cross Layers Output: Capacity. VMASK layer uses the importance scores it detects, while the attention [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Cross Layers Output: Interference. The attention layer uses the attention weight matrix. The original [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 11 canonical work pages

  1. [1]

    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

  2. [2]

    Sharkey, L \'e o Grinsztajn, Eric Winsor, Daniel A

    Sid Black, Lee D. Sharkey, L \'e o Grinsztajn, Eric Winsor, Daniel A. Braun, Jacob Merizian, Kip Parker, Carlos Ram'on Guevara, Beren Millidge, Gabriel Alfour, and Connor Leahy. 2022. https://api.semanticscholar.org/CorpusID:253761579 Interpreting neural networks through the polytope lens . ArXiv, abs/2211.12312

    arXiv 2022

  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]

    Hanjie Chen and Yangfeng Ji. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.347 Learning variational word masks to improve the interpretability of neural text classifiers . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4236--4251, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.emnlp-main.347 2020

  5. [5]

    Jiawei Chen, Hongyu Lin, Xianpei Han, Yaojie Lu, Shanshan Jiang, Bin Dong, and Le Sun. 2024. https://aclanthology.org/2024.lrec-main.634 Few-shot named entity recognition via superposition concept discrimination . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), p...

    work page 2024

  6. [6]

    Zhongtian Chen, Edmund Lau, Jake Mendel, Susan Wei, and Daniel Murfet. 2023. https://api.semanticscholar.org/CorpusID:263831043 Dynamical versus bayesian phase transitions in a toy model of superposition . ArXiv, abs/2310.06301

    arXiv 2023

  7. [7]

    Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben, Wojciech Samek, and Sebastian Lapuschkin. 2024. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 8212--8217

    work page 2024

  8. [8]

    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

    arXiv 2022

Show all 28 references
  1. [9]

    Gabriel Goh. 2016. https://gabgoh.github.io/ThoughtVectors/ Decoding the thought vector

    2016

  2. [10]

    Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsimas. 2023. https://openreview.net/forum?id=JYs1R9IMJr Finding neurons in a haystack: Case studies with sparse probing . Transactions on Machine Learning Research

    2023

  3. [11]

    Kaarel H \"a nni, Jake Mendel, Dmitry Vaintrob, and Lawrence Chan. 2024. https://openreview.net/forum?id=OcVJP8kClR Mathematical models of computation in superposition . In ICML 2024 Workshop on Mechanistic Interpretability

    2024

  4. [12]

    Tom Henighan, Shan Carter, Tristan Hume, Nelson Elhage, Robert Lasenby, Stanislav Fort, Nicholas Schiefer, and Christopher Olah. 2023. Superposition, memorization, and double descent. Transformer Circuits Thread, 6:24

    2023

  5. [13]

    Chenhui Hu, Pengfei Cao, Yubo Chen, Kang Liu, and Jun Zhao. 2024. https://api.semanticscholar.org/CorpusID:271865810 Knowledge in superposition: Unveiling the failures of lifelong knowledge editing for large language models . ArXiv, abs/2408.07413

    2024 arXiv

  6. [14]

    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

  7. [15]

    Sarthak Jain and Byron C Wallace. 2019. Attention is not explanation. arXiv preprint arXiv:1902.10186

    2019 arXiv

  8. [16]

    Pavan Katta. 2024. https://api.semanticscholar.org/CorpusID:270870435 On implications of scaling laws on feature superposition . ArXiv, abs/2407.01459

    2024 arXiv

  9. [17]

    Michael Lan, Philip Torr, Austin Meek, Ashkan Khakzar, David Krueger, and Fazl Barez. 2024. https://arxiv.org/abs/2410.06981 Sparse autoencoders reveal universal feature spaces across large language models . Preprint, arXiv:2410.06981

    2024 arXiv

  10. [18]

    Victor Lecomte, Kushal Thaman, Rylan Schaeffer, Naomi Bashkansky, Trevor Chow, and Sanmi Koyejo. 2024. https://openreview.net/forum?id=AHfE6WeJLQ What causes polysemanticity? an alternative origin story of mixed selectivity from incidental causes . In ICLR 2024 Workshop on Rep...

    2024

  11. [19]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. https://aclanthology.org/P11-1015 Learning word vectors for sentiment analysis . In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Hu...

    2011

  12. [20]

    Simon C Marshall and Jan H Kirchner. 2024. Understanding polysemanticity in neural networks through coding theory. arXiv preprint arXiv:2401.17975

    2024 arXiv

  13. [21]

    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

  14. [22]

    Adam Scherlis, Kshitij Sachan, Adam S Jermyn, Joe Benton, and Buck Shlegeris. 2022. Polysemanticity and capacity in neural networks. arXiv preprint arXiv:2210.01892

    2022 arXiv

  15. [23]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. https://aclanthology.org/D13-1170 Recursive deep models for semantic compositionality over a sentiment treebank . In Proceedings of the 2013 Conference on Emp...

    2013

  16. [24]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems

    2017

  17. [25]

    Jiachuan Wang, Shimin Di, Lei Chen, and Charles Wang Wai Ng. 2023. https://api.semanticscholar.org/CorpusID:266362804 Learning from emergence: A study on proactively inhibiting the monosemantic neurons of artificial neural networks . In Knowledge Discovery and Data Mining

    2023

  18. [26]

    Jiachuan Wang, Shimin Di, Lei Chen, and Charles Wang Wai Ng. 2024. https://arxiv.org/abs/2312.11560 Learning from emergence: A study on proactively inhibiting the monosemantic neurons of artificial neural networks . Preprint, arXiv:2312.11560

    2024 arXiv

  19. [27]

    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...

  20. [28]

    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 10, 2026 · model on record in the stance chip above.