Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Decoding Knowledge Attribution in Mixture-of-Experts: A Framework of Basic-Refinement Collaboration and Efficiency Analysis

T0 review · 5 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 The paper claims that MoE models split knowledge work between shared generalists and routed specialists, and that this basic-refinement division yields 31 to 37 percent higher per-layer efficiency than dense models.

desk verdict Plausible qualitative story about MoE experts, but the paper's own tables contradict its headline efficiency and robustness numbers, and the core attribution proxy is unvalidated. read the letter →

arxiv 2505.24593 v2 pith:FM6KO3CV submitted 2025-05-30 cs.CL cs.AIcs.LG

Junzhuo Li , Bo Wang , Xiuze Zhou , Peijie Jiang , Jia Liu , Xuming Hu This is my paper
classification cs.CLcs.AIcs.LG
keywords mixtureofexpertsMoEinterpretabilityknowledgeattributionneuron-levelexpertroutingsharedlayerefficiencyrobustness
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 asks what experts inside a Mixture-of-Experts language model are actually doing: which experts carry which knowledge, how shared and routed experts cooperate, and whether sparse architectures process knowledge more efficiently than dense ones. It extends neuron-level attribution from dense transformers to MoE layers by weighting each expert neuron with its gating probability, and applies the probe across Qwen 1.5-MoE, OLMoE, and Mixtral-8x7B against three dense baselines on twelve knowledge relations. Its central claim is a basic-refinement framework, where shared experts handle general tasks and routed experts refine domain-specific attributes, realized through a mid-activation, late-amplification staging that gives MoE models higher per-layer knowledge gain than their dense counterparts. If correct, the result converts MoE design from trial-and-error into concrete choices: where to place shared experts, how many routed experts are enough, and how much depth a task sensitivity demands. The paper itself restricts these conclusions to 7-billion-parameter, static-routing MoE models.

What carries the argument

The load-bearing instrument is a cross-level attribution algorithm: the neuron-level importance score used for dense FFNs, extended to MoE by weighting an expert neuron's added contribution by its gating probability, I(v_E_j^l) = log p(x_i | g_i_j^l v_E_j^l + u^l) - log p(x_i | u^l). Summed over a layer's neurons, these scores define the per-layer FFN gain, which yields the paper's headline quantities: layer efficiency (total FFN gain divided by layer count), the early, mid, and late stage percentages, and the expert importance rankings used in ablation and blocking experiments. A second instrument is the temporal correlation between attention heads and gated experts, backed by three interventions that argue the attention-to-expert link is causal: head suppression cuts top-4 gating probability by 54 percent and MRR by 29 percent, expert forcing recovers 85 percent of the original MRR, and integrated gradients attribute 28 percent of expert-activation signal to specific heads.

What would settle it

Recompute per-layer FFN gains with activation patching instead of the linear proxy: on the twelve LRE relations, set each expert's output to zero, or to its mean over unrelated tokens, and measure the change in log-probability of the target answer, then recompute layer efficiency and stage shares. If the ordering, Qwen 1.5-MoE 0.307 versus Qwen 1.5-7B 0.203, and the mid-activation peak near 85 percent of depth do not survive, those claims describe the attribution method rather than the models.

Watch

Extended reading notes

Core claim

The central discovery is a division of labour inside MoE language models that the paper calls basic-refinement: shared experts act as generalists for foundational steps such as locating entities, while routed experts specialize in refinement steps such as linking Canada to Ottawa. This collaboration follows a staged, depth-dependent pattern, mid-activation, late-amplification, in which early layers initialize feature extraction, mid-layers screen experts through top-k gating, and late layers amplify gain by combining shared and routed experts. The paper supports the pattern with an extended neuron-level attribution method, reporting that Qwen 1.5-MoE reaches a layer efficiency of 0.307 in 24 layers against 0.203 for Qwen 1.5-7B in 32 layers, a gap stated as 31 percent in the abstract and 37 percent in the introduction, that attention-head activity and expert selection correlate at r=0.68, and that blocking the top-10 experts for geographic tasks drops MRR by 43 percent in deep Qwen 1.5-MoE while shallow OLMoE degrades far more sharply. Three interventions, suppressing head H19-12, forcing expert E20-59, and tracing paths with integrated gradients, are offered as evidence that attention heads steer routing rather than merely co-activating. The paper reads the robustness contrast as evidence that architectural depth plus shared-expert redundancy, rather than expert count alone, decides whether a specialist failure is absorbed.

Load-bearing premise

The load-bearing premise is that the gating-weighted importance score of Equation 9 measures how much an expert truly contributes to a prediction; every headline number in the paper, layer efficiency, stage percentages, robustness drops, is computed from these linear scores, and the paper tests the proxy against no counterfactual that isolates an expert's actual causal effect.

Editorial extensions

If this is right

  • Placement rule: shared experts should sit in early layers for cross-task redundancy and routed specialists in later layers for attribute refinement, a strategy the paper presents as validated by Qwen 1.5-MoE's design.
  • Diminishing returns: beyond a small core of shared plus top routed experts, additional routed experts add little, with HIT@10 rising only from 0.82 to 0.85 in Qwen 1.5-MoE, implying an effective expert threshold for knowledge tasks.
  • Task-aware depth: core-sensitive tasks such as geography concentrate knowledge in a few late-layer specialists and need redundancy, while distributed-tolerant tasks such as object attributes can spread work across experts and tolerate shallower designs.
  • Failure tolerance: deep architectures with shared experts buffer expert failure, shallow architectures without shared experts degrade sharply when a critical expert is blocked, and coarse-grained Mixtral is the exception because its dense top-2 routing spreads selection broadly.
  • Scope qualifier: the paper is explicit that the claims cover 7-billion-parameter, static-routing MoE models, leaving larger 100B-plus and fully dynamic routers untested.

Reading between the lines

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

  • A testable extension the paper does not pursue: recompute layer gains with activation patching, zeroing each expert's output, to check whether the reported 0.307-versus-0.203 efficiency ordering is a property of the models or of the linear attribution proxy.
  • The basic-refinement pattern suggests a scaling rule the paper leaves implicit: shared-expert capacity should track the breadth of foundational tasks, while the number of routed experts should track the number of distinct refinement operators a deployment needs.
  • Because the attention-expert linkage is measured on the model's own routing distribution, reordering prompts out-of-distribution would probe whether the r=0.68 correlation reflects genuine semantic steering or a co-activation pattern fixed at pretraining.
  • The paper's tentative reading of Mixtral, weight-copied initialization homogenizing experts, implies that initialization may shape specialization as much as architecture does; training the same top-2 router from random initialization would isolate that factor.
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

5 major / 5 minor

Summary. The paper proposes a cross-level attribution method for Mixture-of-Experts (MoE) models, extending neuron-level importance scoring to routed and shared experts, and uses it to compare three MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) with dense baselines (Qwen 1.5-7B, Llama-7B, Mistral-7B). The authors report that MoE models achieve higher per-layer efficiency via a 'mid-activation, late-amplification' pattern, that shared experts play a 'basic' role while routed experts refine domain-specific knowledge, that attention heads semantically guide expert selection, and that deep MoE architectures are more robust to expert failure than shallow ones. The paper also reports a 76% MRR drop for OLMoE when blocking top-10 experts, contrasted with a 43% drop for Qwen 1.5-MoE.

Significance. If the quantitative claims were reliable, the paper would make a useful contribution to MoE interpretability, with concrete design principles (shared experts in early layers, routed experts in late layers, depth as redundancy). The expert-blocking ablations are a genuine strength: they are direct interventions and provide independent evidence for some specialization and robustness effects. However, the headline quantitative claims—the 37% efficiency gain, the 76% MRR drop, and the r=0.68 attention–expert correlation—are contradicted by the paper's own tables or by its own text. The efficiency and stage-percentage numbers all derive from an attribution proxy that is neither validated nor compared with a counterfactual. These issues make the central conclusions as stated unreliable and not reproducible.

major comments (5)
  1. [Abstract, §1, Table 1] The central per-layer efficiency claim is internally inconsistent. The abstract and §1 state that Qwen 1.5-MoE achieves 37% higher per-layer efficiency than Qwen 1.5-7B, but the numbers given in §1 (0.294 vs. 0.213) differ from those in Table 1 (0.307 vs. 0.203). The abstract in the full text also reports 31% in one place. These values are not reconcilable, and the efficiency gain computed from Table 1 is actually about 51%, not 37% or 31%. The paper must present a single consistent set of numbers before any efficiency claim can be evaluated.
  2. [Abstract, §5.2, Tables 5/8] The claim that OLMoE suffers a 76% MRR drop when blocking top-10 experts is not supported by the reported data. The largest OLMoE drop in Tables 5 and 8 is for fruit_inside_color (0.76 to 0.41, a 46% relative drop); country_language drops from 0.96 to 0.68 (29%). The value 0.76 appears in the tables as the post-blocking MRR for country_capital_city (from 1.0), suggesting that the authors confused an absolute MRR value with a relative drop. The 43% (Qwen) vs. 76% (OLMoE) robustness contrast in the abstract is therefore false. This is a load-bearing error because the 'architectural depth dictates robustness' conclusion rests on that contrast.
  3. [§2.2, Eq. 9, §3] All efficiency, stage-percentage, and gain numbers are computed from the importance scores defined in Eqs. 7–9, but the proxy is not validated. Equation 9 computes expert-neuron importance by adding the gating-scaled neuron to the intermediate representation and measuring the log-probability change, an additive linearization that ignores the nonlinear expert FFN in Eq. 3. The authors state that this 'accurately attributes importance' without providing any validation against activation patching, zeroing, or other counterfactual interventions. Since the layer-efficiency comparison, the 'mid-activation, late-amplification' stage percentages, and the robustness analysis all depend on these scores, their quantitative conclusions are not established unless the proxy is shown to track causal contribution.
  4. [§4.2, Introduction] The claimed semantic-driven routing correlation is misattributed. The text says 'the correlation coefficient between E18_10 and E20_59 reaches r=0.68 (p<0.001),' but both E18_10 and E20_59 are routed experts, not an attention head and an expert. This does not support the statement that 'attention heads actively steer expert selection.' The introduction reports p<0.0014 while §4.2 reports p<0.001, and the pair of variables being correlated is unclear. This weakens the evidence for the semantic-routing mechanism and needs clarification.
  5. [§3.1, §3.2, §4.2] The proposed processing pattern is not consistently described. The text claims a 'mid-activation, late-amplification' pattern, but §3.1 reports that Qwen 1.5-MoE and OLMoE reach peak FFN gains at 84.8% and 84.6% of network depth, i.e., in the late stage. In §3.2, OLMoE is described as peaking at layer 12 (mid-stage), which for 16 layers is 75% depth, not 84.6%. The stage boundaries also differ between §3.2 (early 1–13, mid 14–19, late 20–24 for Qwen) and §4.2 (early-to-mid 18–20, deeper 21–23). The central 'mid-activation, late-amplification' conclusion is therefore difficult to pin down and needs a single, coherent stage definition.
minor comments (5)
  1. [Table 5] In Table 5, the Top5 experts for Qwen on object_superclass list E21_25 twice ('E 21 25 ,E 21 25'), which appears to be a typo; one of these should be verified.
  2. [§5.1] The text says blocking E21_20 in Qwen causes only a marginal MRR drop, but Table 5 lists E21_48 as the top expert for country_capital_city; please confirm which expert was blocked and report the correct identifier.
  3. [§5.1] The statement that 'blocking the critical expert E13_56 in OLMoE degrades MRR by 30%' is not consistent with Table 5's top-1 exclusions, where the largest drop is about 21% (fruit_inside_color: 0.76 to 0.60). Please provide the specific task and configuration.
  4. [Appendix A] There are typographical errors in Appendix A: 'LLama-7B' should be 'Llama-7B', and the sentence listing dense models has an unbalanced parenthesis.
  5. [Limitations] The Limitations section acknowledges the 7B-parameter and static-routing scope, but it should also explicitly note that the attribution proxy has not been validated against counterfactual interventions; this is a primary limitation of the quantitative claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central efficiency, specialization, and robustness claims rest on new measurements of an attribution metric plus independent ablations and causal interventions, not on predictions that reduce to their own inputs.

full rationale

The paper's load-bearing quantitative claims—per-layer efficiency, the mid-activation/late-amplification pattern, expert specialization, and the robustness contrast—are empirical measurements, not fitted parameters renamed as predictions. The importance scores in Equations 7–9 define a measurement instrument; the layer-efficiency numbers (Table 1) and stage contributions are aggregates of those measurements. A measurement can be invalid or misleading, but it is not circular merely because the paper uses its own metric to summarize model behavior. The expert-blocking experiments in Section 5.2, the attention-head suppression and expert-forcing interventions in Section 4.3, and the temporal correlation r=0.68 in Section 4.2 are all independent outcome measurements: the MRR drops and recoveries are observed facts of the actual model, not consequences of the attribution definition. The attribution metric only selects which experts or heads to intervene on; it does not force the measured outcomes. The only self-citations (Wu et al., 2023; Xu et al., 2023) appear in the Related Work survey of knowledge attribution and are not load-bearing for any of the paper's novel claims; no uniqueness theorem is imported, no ansatz is smuggled in by self-citation, and the base attribution method is explicitly adopted from external prior work (Yu and Ananiadou, 2024). The proposition that the additive importance proxy is unvalidated against counterfactual ablations is a legitimate correctness and validity concern, but it is not a circularity concern under the definitions used here: the paper does not derive a prediction from an input that is defined in terms of that prediction. The inconsistencies among the 31%/37% efficiency figures and the MRR drops are further correctness risks, not evidence of circular reasoning.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or model entities. Its analysis relies on the importance-score proxy, a hand-chosen efficiency metric, and post hoc stage boundaries. The central quantitative claims inherit the validity of these assumptions, none of which are externally benchmarked.

free parameters (2)
  • Stage boundaries (early/mid/late layers) = Qwen 1.5-MoE: 1-13 / 14-19 / 20-24; other models not fully specified
    The three-phase contribution percentages (6.1%, 43.5%, 50.4%) depend on hand-chosen layer partitions; changing the boundaries changes the claimed processing pattern.
  • Top-N neuron thresholds = 100, 200, 300
    The shared-vs-routed neuron ratio (25% routed at top 100 vs 38% at top 300) is threshold-dependent and is used to support the dynamic collaboration claim.
assumptions (3)
  • domain assumption The neuron-level attribution method (Eq. 7-9) measures causal contribution of experts to model predictions.
    No validation that adding gating-weighted neuron vectors to the residual approximates the expert's true contribution; expert FFNs are nonlinear (Eq. 3).
  • ad hoc to paper Layer efficiency = total FFN gain / number of layers is a valid cross-architecture efficiency metric.
    Introduced in Section 3.1 without justification; ignores parameter count, activation budget, and compute per layer, and structurally favors shallow models.
  • domain assumption The selected 12 knowledge relations from LRE are representative and the subset highlighted in Table 5 is chosen without bias.
    Appendix A says only relations with similar performance are analyzed to avoid the influence of model knowledge, a post hoc selection rule that affects the reported specialization and robustness results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoding Knowledge Attribution in Mixture-of-Experts: A Framework of Basic-Refinement Collaboration and Efficiency Analysis." pith.science (2026). https://pith.science/paper/FM6KO3CV

@misc{pith2026250524593,
  author       = {Pith},
  title        = {Pith review of: Decoding Knowledge Attribution in Mixture-of-Experts: A Framework of Basic-Refinement Collaboration and Efficiency Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FM6KO3CV}},
  note         = {Machine review of arXiv:2505.24593}
}
read the original abstract

The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored. Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures. To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mistral-7B). Results show MoE models achieve 37% higher per-layer efficiency via a "mid-activation, late-amplification" pattern: early layers screen experts, while late layers refine knowledge collaboratively. Ablation studies reveal a "basic-refinement" framework--shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes). Semantic-driven routing is evidenced by strong correlations between attention heads and experts (r=0.68), enabling task-aware coordination. Notably, architectural depth dictates robustness: deep Qwen 1.5-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow OLMoE suffers severe degradation (76% drop). Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation. These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.

Figures

Figures reproduced from arXiv: 2505.24593 by the authors.

Figure 1
Figure 1. Schematic of the MoE expert layer: Shared [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of layer importance scores distri [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Cumulative FFN layers contribution curves. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: OLMoE’s performance on knowledge predic [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: HIT@10 scores of four models on 12 different [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Type distribution of neurons in the top 100, [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Robustness comparison heat map. The darker [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Survey on Data Security in Large Language Models

    cs.CR 2025-08 conditional novelty 2.0 of 10

    A survey of data security risks in LLMs that organizes threats, defenses, and evaluation datasets, with notable factual errors in its tables.

Reference graph

Works this paper leans on

43 extracted references · 9 canonical work pages · cited by 1 Pith paper

  1. [1]

    Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. https://arxiv.org/abs/1607.06450 Layer normalization . CoRR, abs/1607.06450

    arXiv 2016

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  3. [3]

    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 Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    2020

  4. [4]

    Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y

    Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. 2024. https://doi.org/10.18653/v1/2024.acl-long.70 D eep S eek M o E : Towards ultimate expert specialization in mixture-of-experts language models . ...

    doi:10.18653/v1/2024.acl-long.70 2024

  5. [5]

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2022 a . https://doi.org/10.18653/v1/2022.acl-long.581 Knowledge neurons in pretrained transformers . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8493--8502, Dublin, Ireland. Association for Computational Linguistics

    doi:10.18653/v1/2022.acl-long.581 2022

  6. [6]

    Damai Dai, Li Dong, Shuming Ma, Bo Zheng, Zhifang Sui, Baobao Chang, and Furu Wei. 2022 b . https://doi.org/10.18653/v1/2022.acl-long.489 S table M o E : Stable routing strategy for mixture of experts . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7085--7095, Dublin, Ireland. Ass...

    doi:10.18653/v1/2022.acl-long.489 2022

  7. [7]

    Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J

    DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huaj...

    arXiv 2024

  8. [8]

    William Fedus, Barret Zoph, and Noam Shazeer. 2021. https://doi.org/10.48550/ARXIV.2101.03961 Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity . arXiv preprint

    arXiv doi:10.48550/arxiv.2101.03961 2021

Show all 43 references
  1. [9]

    Mor Geva, Avi Caciularu, Kevin Wang, and Yoav Goldberg. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.3 Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space . In Proceedings of the 2022 Conference on Empirical Methods in Natural L...

    2022 doi

  2. [10]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.446 Transformer feed-forward layers are key-value memories . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484--5495, ...

    2021 doi

  3. [11]

    Yongxin Guo, Zhenglin Cheng, Xiaoying Tang, Zhaopeng Tu, and Tao Lin. 2025. https://arxiv.org/abs/2405.14297 Dynamic mixture of experts: An auto-tuning approach for efficient transformer models . Preprint, arXiv:2405.14297

    2025 arXiv

  4. [12]

    Dan Hendrycks and Kevin Gimpel. 2023. https://arxiv.org/abs/1606.08415 Gaussian error linear units (gelus) . Preprint, arXiv:1606.08415

    2023 arXiv

  5. [13]

    Evan Hernandez, Arnab Sen Sharma, Tal Haklay, Kevin Meng, Martin Wattenberg, Jacob Andreas, Yonatan Belinkov, and David Bau. 2024. https://openreview.net/forum?id=w7LU2s14kE Linearity of relation decoding in transformer language models . In The Twelfth International Conference...

    2024

  6. [14]

    Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. 2024. https://doi.org/10.18653/v1/2024.acl-long.696 Harder task needs more experts: Dynamic routing in M o E models . In Proceedings of the 62nd...

    2024 doi

  7. [15]

    Jacobs, Michael I

    Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. 1991. https://doi.org/10.1162/neco.1991.3.1.79 Adaptive mixtures of local experts . Neural Computation, 3(1):79--87

    1991 doi

  8. [16]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023 arXiv

  9. [17]

    Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, L \' e lio Renard Lavaud, Lucile Saulnier, Marie...

    2024 doi arXiv

  10. [18]

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2021. https://openreview.net/forum?id=qrwe7XHTmYb GS hard: Scaling giant models with conditional computation and automatic sharding . In Interna...

    2021

  11. [19]

    Ka Man Lo, Zeyu Huang, Zihan Qiu, Zili Wang, and Jie Fu. 2025. https://aclanthology.org/2025.findings-naacl.251/ A closer look into mixture-of-experts in large language models . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 4427--4447, Albuque...

    2025

  12. [20]

    Ang Lv, Yuhan Chen, Kaiyi Zhang, Yulong Wang, Lifeng Liu, Ji-Rong Wen, Jian Xie, and Rui Yan. 2024. https://arxiv.org/abs/2403.19521 Interpreting key mechanisms of factual recall in transformer-based language models . Preprint, arXiv:2403.19521

    2024 arXiv

  13. [21]

    Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. 2013. https://ai.stanford.edu/ amaas/papers/relu_hybrid_icml2013_final.pdf Rectifier nonlinearities improve neural network acoustic models . Proceedings of the 30th International Conference on Machine Learning (ICML-13), pages 338--345

    2013

  14. [22]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/6f1d43d5a82a37e89b0665b33bf3a182-Paper-Conference.pdf Locating and editing factual associations in GPT . Advances in Neural Information Processing Systems, 35

    2022

  15. [23]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023. https://arxiv.org/abs/2210.07229 Mass-editing memory in a transformer . Preprint, arXiv:2210.07229

    2023 arXiv

  16. [24]

    Smith, Pang Wei Koh, Amanpreet Singh, and Hannaneh Hajishirzi

    Niklas Muennighoff, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Jacob Morrison, Sewon Min, Weijia Shi, Pete Walsh, Oyvind Tafjord, Nathan Lambert, Yuling Gu, Shane Arora, Akshita Bhagia, Dustin Schwenk, David Wadden, Alexander Wettig, Binyuan Hui, Tim Dettmers, Douwe Kiela, Ali F...

    2024 arXiv

  17. [25]

    Vinod Nair and Geoffrey E. Hinton. 2010. https://dl.acm.org/doi/10.5555/3104322.3104425 Rectified linear units improve restricted boltzmann machines . In Proceedings of the 27th International Conference on International Conference on Machine Learning, ICML'10, page 807–814, Ma...

    2010

  18. [26]

    Jungwoo Park, Ahn Young Jin, Kee-Eung Kim, and Jaewoo Kang. 2025. https://openreview.net/forum?id=1Ogw1SHY3p Monet: Mixture of monosemantic experts for transformers . In The Thirteenth International Conference on Learning Representations

    2025

  19. [27]

    Stephen Roller, Sainbayar Sukhbaatar, arthur szlam, and Jason Weston. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/92bf5e6240737e0326ea59846a83e076-Paper.pdf Hash layers for large sparse models . In Advances in Neural Information Processing Systems, volume ...

    2021

  20. [28]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. https://openreview.net/forum?id=B1ckMDqlg Outrageously large neural networks: The sparsely-gated mixture-of-experts layer . arXiv preprint arXiv:1701.06538

    2017 arXiv

  21. [29]

    Sheng Shen, Le Hou, Yanqi Zhou, Nan Du, Shayne Longpre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. 2023. https://arxiv.or...

    2023 arXiv

  22. [30]

    Xingwu Sun, Yanfeng Chen, Yiqing Huang, Ruobing Xie, Jiaqi Zhu, Kai Zhang, Shuaipeng Li, Zhen Yang, Jonny Han, Xiaobo Shu, Jiahao Bu, Zhongzhi Chen, Xuemeng Huang, Fengzong Lian, Saiyong Yang, Jianfeng Yan, Yuyuan Zeng, Xiaoqin Ren, Chao Yu, Lulu Wu, Yue Mao, Tao Yang, Suncong...

    2024 arXiv

  23. [31]

    Qwen Team. 2024. https://qwenlm.github.io/blog/qwen-moe/ Qwen1.5-moe: Matching 7b model performance with 1/3 activated parameters"

    2024

  24. [32]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023 a . https://arxiv.org/abs/2302.13971 Lla...

    2023 arXiv

  25. [33]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023 b . https://arxiv.org/abs/2307.09288 Llama 2: Open foundation and fine-tuned chat models . arXiv preprint arX...

    2023 arXiv

  26. [34]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  27. [35]

    Zihan Wang, Deli Chen, Damai Dai, Runxin Xu, Zhuoshu Li, and Yu Wu. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.46 Let the expert stick to his last: Expert-specialized fine-tuning for sparse architectural large language models . In Proceedings of the 2024 Conference on E...

    2024 doi

  28. [36]

    Xinwei Wu, Junzhuo Li, Minghui Xu, Weilong Dong, Shuangzhi Wu, Chao Bian, and Deyi Xiong. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.174 DEPN : Detecting and editing privacy neurons in pretrained language models . In Proceedings of the 2023 Conference on Empirical Metho...

    2023 doi

  29. [37]

    Shaoyang Xu, Junzhuo Li, and Deyi Xiong. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.226 Language representation projection: Can we transfer factual knowledge across languages in multilingual language models? In Proceedings of the 2023 Conference on Empirical Methods in ...

    2023 doi

  30. [38]

    Fuzhao Xue, Zian Zheng, Yao Fu, Jinjie Ni, Zangwei Zheng, Wangchunshu Zhou, and Yang You. 2024. https://dl.acm.org/doi/10.5555/3692070.3694363 Openmoe: an early effort on open mixture-of-experts language models . In Proceedings of the 41st International Conference on Machine L...

    2024

  31. [39]

    Yunzhi Yao, Ningyu Zhang, Zekun Xi, Mengru Wang, Ziwen Xu, Shumin Deng, and Huajun Chen. 2024. https://openreview.net/forum?id=YVXzZNxcag Knowledge circuits in pretrained transformers . In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  32. [40]

    Zeping Yu and Sophia Ananiadou. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.191 Neuron-level knowledge attribution in large language models . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 3267--3280, Miami, Florida, USA....

    2024 doi

  33. [41]

    Zhiyuan Zeng, Qipeng Guo, Zhaoye Fei, Zhangyue Yin, Yunhua Zhou, Linyang Li, Tianxiang Sun, Hang Yan, Dahua Lin, and Xipeng Qiu. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.739 Turn waste into worth: Rectifying top- k router of M o E . In Proceedings of the 2024 Conferen...

    2024 doi

  34. [42]

    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. [43]

    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.