Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

SkipGPT: Dynamic Layer Pruning Reinvented with Token Awareness and Module Decoupling

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

Pith's one-line read SkipGPT claims that a lightweight per-token router can skip 40% of a large language model's attention and MLP modules while keeping model quality at the dense level.

desk verdict The method itself is solid and worth a serious look, but the abstract's '40% parameter reduction' claim is a mislabeling of per-token compute sparsity, and at 40% sparsity the model is clearly below dense. read the letter →

arxiv 2506.04179 v1 pith:CHGU24OW submitted 2025-06-04 cs.CL

classification cs.CL
keywords dynamiclayerpruningtoken-adaptivecomputationattentionandMLPdecouplingGumbel-Softmaxroutingtwo-stagetrainingLoRAfine-tuningLLMinferenceefficiencymoduleredundancyanalysis
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

SkipGPT tries to establish that layer pruning for large language models should be dynamic in two ways that static pruning ignores: per token, because some tokens need more computation than others, and per module, because self-attention and MLP blocks have different redundancy within the same layer. Its recipe is a tiny router in front of every attention and MLP block that decides, for each token, whether to skip the module or run it, with a global compute budget instead of a fixed per-layer budget. The router is learned first with the pretrained model frozen, and the model is then lightly fine-tuned with LoRA to recover any loss. On LLaMA2-7B, LLaMA2-13B, and LLaMA3.1-8B, the paper reports that this keeps over 90% of dense accuracy at 25% parameter reduction with router tuning alone, and the paper's headline claim is that after LoRA fine-tuning the pruned model matches or exceeds the dense model even at 40% reduction. If true, a large share of an LLM's depth can be skipped at inference time for a negligible routing cost, and the learned router doubles as a map of where the model is genuinely redundant.

What carries the argument

The load-bearing object is the module-level router: one linear projection per attention block and per MLP block that outputs a two-class distribution over skip and execute. Gumbel-Softmax reparameterization with a straight-through estimator makes the discrete choice differentiable, and a global sparsity loss $\mathcal{L}_{\text{sparsity}} = |T - r|$ with $\alpha = 8$ in $\mathcal{L}_{\text{all}} = \mathcal{L}_{\text{lm}} + \alpha \mathcal{L}_{\text{sparsity}}$ enforces the compute budget. The two-stage schedule, router tuning on a frozen model followed by LoRA fine-tuning with a frozen router, is what the paper credits for training stability. The same router doubles as a probe: because stage one leaves the pretrained weights untouched, its skip statistics are read as evidence about which modules and which token positions are intrinsically redundant.

What would settle it

After LoRA fine-tuning, run the frozen router on a fixed evaluation set and compare its decisions with the decisions it made at the end of stage one; if the decisions diverge sharply, or if retraining the router on the LoRA-adapted model yields a clearly better pruned model, the two-stage assumption breaks.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that a per-token, module-level routing policy learned without touching the pretrained weights can identify where computation is wasted, and that this routing survives a subsequent parameter-efficient recovery stage. Concretely, SkipGPT replaces each transformer layer's attention block and MLP block with a gated path: a linear router maps the token hidden state to a two-way categorical distribution, Gumbel-Softmax plus a straight-through estimator turns it into a hard skip/execute decision in the forward pass while keeping gradients flowing in the backward pass, and a sparsity regularizer drives the total fraction of skipped modules toward a user-chosen target. Stage one tunes only the routers; stage two freezes them and LoRA-adapts the model. The authors conclude from their experiments that attention modules carry more redundancy than MLP modules, that later tokens in a long context need more attention but less MLP processing, and that the two-stage separation is what prevents dynamic pruning from collapsing into the poor joint-training solutions they observe in baselines.

Load-bearing premise

The load-bearing premise is that the skip/execute policy learned while the pretrained model is frozen remains near-optimal after LoRA fine-tuning changes the hidden states, since the router is never updated in the second stage.

Editorial extensions

If this is right

  • At 25% parameter reduction, router tuning alone keeps more than 90% of dense accuracy on LLaMA2-7B and LLaMA2-13B and more than 95% on LLaMA3.1-8B, without modifying any pretrained weight.
  • After the optional LoRA stage, the pruned model lands within about a point of the dense model on average accuracy and with only mildly higher perplexity, while static-pruning baselines drop substantially at the same budget.
  • The learned router prunes attention modules more aggressively than MLP modules, indicating that the fixed one-attention-per-MLP layer design carries structural redundancy that future architectures could reduce.
  • The two-stage schedule outperforms joint router-plus-LoRA training, which the paper attributes to the randomly initialized router forcing the pretrained model to adapt to bad early routing decisions.
  • Router tuning is cheap: router parameters are under 0.01% of total model parameters, and the tuning stage runs in about four hours on a single A800 GPU.

Reading between the lines

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

  • My inference: because the stage-one router is trained with the model frozen, its per-module skip statistics are a readout of the original pretrained distribution, so the same routing probe could plausibly transfer to other tasks with little or no retraining as a reusable model-diagnosis tool.
  • My inference: if the frozen-router assumption holds, SkipGPT's routing pathway could be composed with other inference accelerators such as quantization, KV-cache eviction, and early exit, since the routing decisions live in a separate lightweight component.
  • My inference: a natural testable extension is to make a small part of the router itself LoRA-adaptable during stage two, which might remove the staleness risk of frozen routing while preserving the stability benefit of two-stage training.
  • My inference: the paper's context-length finding suggests a hardware-aware scheduling policy for long-context serving, spending more attention compute on later tokens and more MLP compute on early tokens, independent of the learned router.
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. SkipGPT proposes a dynamic, token-level, module-level pruning scheme for LLMs: a lightweight linear router before each attention and MLP module makes per-token skip decisions using Gumbel-Softmax with a straight-through estimator, trained under a language-modeling loss plus a sparsity penalty while the base model is frozen; an optional second stage applies LoRA fine-tuning with the router frozen. The method is evaluated on LLaMA2-7B, LLaMA2-13B, and LLaMA3.1-8B against static pruning baselines (ShortGPT, LaCo, Joint Layer Drop, LLM-Pruner, SliceGPT) and dynamic baselines (MoD-D, D-LLM, SkipGPT-Joint), reporting accuracy on seven commonsense benchmarks and perplexity on WikiText-2 and PTB at roughly 25% and 40% target sparsity. The paper claims in the abstract and conclusion that SkipGPT reduces over 40% of model parameters while matching or exceeding the original dense model across benchmarks, and it further claims to discover that attention is more redundant than MLP and that later tokens need more attention but less MLP computation.

Significance. If the claims were supported, the contribution would be practically valuable: per-token, per-module routing with a stable two-stage training recipe, evaluated across multiple Llama models and many baselines, with code released and an informative ablation against joint training. The experimental comparison is broad, and the observation that SkipGPT-RT outperforms existing baselines at matched sparsity is a useful result. However, the headline claims as written are not supported: the reported sparsity is not parameter removal, dense parity at 40% is contradicted by the paper's own tables, and the frozen-router-after-LoRA assumption is untested. These issues are load-bearing for the paper's central claims and require substantial revision.

major comments (5)
  1. [§3.2, Table 1/2 caption] The sparsity defined in Section 3.2 is the fraction of module computations skipped per forward pass, and the table captions describe the ratio as 'the proportion of parameters (averaged per token) that do not participate in computations.' This is not parameter reduction: every weight remains resident in memory, additional router parameters are introduced, and no wall-clock latency or memory-footprint measurements are reported. The abstract and Section 7 therefore overstate the result when they say SkipGPT 'reduces over 40% of model parameters.' The paper should reframe the contribution as per-token FLOPs reduction and provide actual efficiency metrics such as latency, throughput, and memory usage.
  2. [Table 2, LLaMA3.1-8B 40.3% row] The central claim that SkipGPT matches or exceeds the dense model at 40% reduction is contradicted by the paper's own data: SkipGPT-RT-L achieves average accuracy 70.29 versus 72.69 for the dense model and average PPL 10.40 versus 8.41, trailing the dense model on every reported benchmark. The statements in the abstract, Section 5.2, and Section 7 that the pruned model 'fully restores its performance, even surpassing the original model despite a 40% parameter reduction' are therefore not supported. At 25% sparsity the method is close to dense on several models but still behind on average PPL for LLaMA3.1-8B (9.40 vs 8.41), so the supported claim is roughly dense-competitive at 25% and far ahead of baselines but clearly below dense at 40%.
  3. [§3.5, Algorithm 1] The two-stage paradigm trains the router on the frozen pretrained model's hidden states, then fine-tunes with LoRA while the router remains frozen. Because LoRA updates change all hidden states, the routing decisions optimized in Stage 1 can become stale in Stage 2; the paper never measures routing-decision agreement before and after LoRA fine-tuning, nor does it compare against a variant that retrains the router on the fine-tuned representations. This untested assumption is load-bearing for all SkipGPT-RT-L performance claims and should be investigated empirically, for example by measuring the fraction of changed skip decisions or by adding a router-retraining ablation.
  4. [§3.4 Eq. (7) versus Algorithm 1 line 13] Equation (7) defines sparsity as r = sum g_t^l[0] / (S x 2L) for an L-layer model with 2L modules, but Algorithm 1 line 13 computes r with denominator S x L. These formulas differ by a factor of two and cannot both be correct. In addition, the module-level sparsity target T in Eq. (8) is not the parameter-level ratio reported in the tables; Section 5.1 mentions that attention and MLP have different parameter counts and that modules are pruned 'proportionally' to keep a consistent average parameter ratio, but the mapping from the optimized module sparsity to the reported parameter reduction is never specified, so the reported ratios are not directly controlled by the loss.
  5. [§6, Figures 5-6] The Section 6 insights about attention versus MLP redundancy and about later tokens needing more attention are properties of the learned router under the paper's objectives, not established properties of the original dense model. Without a proof or an independent measure showing that the router identifies the truly optimal pruning decisions, these claims should be qualified as observations about SkipGPT's routing policy rather than as general facts about LLM architecture.
minor comments (5)
  1. [Figure 5] The x-axis label 'User-Defined Target Sparsity ( )' appears to have a missing variable or unit; please correct the label and clarify whether the axis is percent or a fraction.
  2. [Tables 1-2] No standard deviations, multiple seeds, or significance tests are reported; given that some comparisons to the dense model are within one accuracy point, variance information would materially affect the interpretation.
  3. [§3.5 and §5.1] The text says parameters are 'discarded' or 'removed' during router tuning, but no parameters are discarded; the modules are only skipped per token. Please use consistent terminology such as 'skipped computations' or 'per-token FLOPs reduction.'
  4. [References] The Gumbel-Softmax citation 'Jang et al., 2022' appears to be incorrect; the original Categorical Reparameterization paper was published at ICLR 2017, so please verify and update the reference.
  5. [Algorithm 1] The loop variable and the definition of L in Algorithm 1 are ambiguous: the text says an L-layer model has 2L modules, but the algorithm loops 'for module l = 1 to L' and uses L in the sparsity denominator; please make the module-count notation consistent throughout.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the routing pipeline is self-contained and benchmark-evaluated; the '40% parameter reduction' wording is a terminology overstatement, not a circular derivation.

full rationale

The core derivation chain is not circular. SkipGPT's router is trained on a frozen pretrained model using Gumbel-Softmax with a straight-through estimator, optimizing L_all = L_lm + alpha*|T - r|, where r is the measured fraction of skipped module computations (Eqs. 7-9). The router's decisions are then evaluated on held-out lm-evaluation-harness benchmarks against static and dynamic pruning baselines. The benchmark numbers are independent evidence, not outputs of the training objective, so the main empirical comparison does not reduce to the method's own inputs. The abstract's phrase 'reduces over 40% of model parameters' is misleading, because Eq. 7 counts per-token skipped module computations and no weights are actually removed, but this is a correctness/terminology issue rather than a circular step: the paper does not derive the 40% figure from the benchmark results, and the performance at 40% sparsity is an honest (if weaker than claimed) empirical measurement. The Section 6 router-behavior 'insights' are emergent observations from the authors' own trained routers, so they could reflect artifacts of the routing objective or the probe itself, but they are not used to derive SkipGPT's main results and are not definitionally forced. The untested two-stage premise, that a router trained on frozen representations remains near-optimal after LoRA updates, is a robustness gap rather than a circular argument. Overall, no load-bearing step in the method's derivation is equivalent by construction to its inputs or to a self-citation.

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

The method introduces a per-module linear router, which is a trained architectural component rather than an invented entity. The central claims depend on several empirical assumptions: the cosine-similarity importance metric, the expressiveness of linear routers, and the stability of frozen routing decisions after LoRA adaptation. Training hyperparameters (alpha, learning rate, temperature schedule) are hand-tuned and not subject to sensitivity analysis.

free parameters (3)
  • sparsity penalty alpha = 8
    Chosen by experiments rather than derived or swept; no sensitivity analysis is reported (Section 3.4, footnote 2).
  • router learning rate = 2e-3
    Selected by grid search to ensure stability (Section 4.1, footnote 4).
  • Gumbel-Softmax temperature schedule = linear anneal from 5 to 1
    Chosen by hand; no ablation on the schedule is provided (Section 4.1).
assumptions (4)
  • domain assumption Cosine similarity between a module's input and output is a reliable measure of module importance.
    Inherited from Men et al. 2024 and Gromov et al. 2024, and used in Section 2 to motivate horizontal and vertical dynamics. If this metric is not a valid proxy for importance, the motivation is weakened.
  • ad hoc to paper A linear router operating on the token hidden state can express the optimal per-token, per-module skip decision.
    The router is a single linear projection (Section 3.3). The paper provides no analysis of whether this restricted policy class is expressive enough to capture the token-level importance it aims to learn.
  • ad hoc to paper Frozen-model routing decisions remain near-optimal after LoRA fine-tuning modifies the hidden states.
    Stage 2 freezes the router while updating LoRA parameters (Algorithm 1). The paper does not test whether the router and model stay aligned after fine-tuning.
  • standard math Gumbel-Softmax with a straight-through estimator provides usable gradients for the discrete routing decisions.
    Standard technique from Jang et al. 2022 and Bengio et al. 2013, invoked in Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SkipGPT: Dynamic Layer Pruning Reinvented with Token Awareness and Module Decoupling." pith.science (2026). https://pith.science/paper/CHGU24OW

@misc{pith2026250604179,
  author       = {Pith},
  title        = {Pith review of: SkipGPT: Dynamic Layer Pruning Reinvented with Token Awareness and Module Decoupling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CHGU24OW}},
  note         = {Machine review of arXiv:2506.04179}
}
read the original abstract

Large language models (LLMs) achieve remarkable performance across tasks but incur substantial computational costs due to their deep, multi-layered architectures. Layer pruning has emerged as a strategy to alleviate these inefficiencies, but conventional static pruning methods overlook two critical dynamics inherent to LLM inference: (1) horizontal dynamics, where token-level heterogeneity demands context-aware pruning decisions, and (2) vertical dynamics, where the distinct functional roles of MLP and self-attention layers necessitate component-specific pruning policies. We introduce SkipGPT, a dynamic layer pruning framework designed to optimize computational resource allocation through two core innovations: (1) global token-aware routing to prioritize critical tokens, and (2) decoupled pruning policies for MLP and self-attention components. To mitigate training instability, we propose a two-stage optimization paradigm: first, a disentangled training phase that learns routing strategies via soft parameterization to avoid premature pruning decisions, followed by parameter-efficient LoRA fine-tuning to restore performance impacted by layer removal. Extensive experiments demonstrate that SkipGPT reduces over 40% of model parameters while matching or exceeding the performance of the original dense model across benchmarks. By harmonizing dynamic efficiency with preserved expressivity, SkipGPT advances the practical deployment of scalable, resource-aware LLMs. Our code is publicly available at: https://github.com/EIT-NLP/SkipGPT.

Figures

Figures reproduced from arXiv: 2506.04179 by the authors.

Figure 1
Figure 1. An overview of SkipGPT. Unlike conventional static structured pruning, SkipGPT dynamically prunes layers by con￾sidering both horizontal and vertical dynamics. In horizontal dynamics, different tokens receive varying computational alloca￾tions. In vertical dynamics, the MLP and attention modules are decoupled to account for their distinct roles within each layer. 1. Introduction Large language models (LLMs) are buil… view at source ↗
Figure 2
Figure 2. Token-Wise Cosine Similarities Across Modules in LLaMA-2-7B, which consists of 32 layers, corresponding to 64 modules in total. Due to space constraints, we showcase only the results for the initial and final modules. Higher values indicate greater redundancy. The Necessity of Vertical Dynamics Existing pruning methods, whether dynamic or static, typically treat an entire transformer layer as the smallest pruning un… view at source ↗
Figure 3
Figure 3. Illustration of the forward and backward passes during the router tuning stage. In the forward pass, the router makes hard decisions to either execute (1) or skip (0). In the backward pass, the gradients are propagated back using soft probabilities. soft probabilities facilitate gradient propagation, allowing the router weights to be updated effectively. With this routing design, in principle, we can train the route… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Training loss curves of LoRA-finetuned static and dynamic pruning baselines, and the two-stage training of SkipGPT. 30 40 50 60 70 User-Defined Target Sparsity ( ) 0 20 40 60 80 Attention and MLP Sparsity (%) 33.19 45.19 54.41 66.81 74.35 26.33 34.30 46.03 52.57 64.53 …
Figure 5
Figure 5. Figure 5: Average sparsity of the attention and MLP modules in five pruned models of LLaMA-2-13B, generated through router tuning under different target sparsity levels T . RT consistently prunes more attention than MLP modules, across both low and high pruning rates. This obser…
Figure 7
Figure 7. Figure 7: Perplexity (PPL) of Joint Layer Drop and SkipGPT-RT under different sparsity levels. 0.0 0.2 0.4 0.6 0.8 1.0 Attn 30 MLP 30 Attn 31 MLP 31 Attn 32 MLP 32 0.99 0.97 0.98 0.99 1.00 0.99 0.99 0.99 0.98 0.99 0.99 0.99 0.99 0.99 1.00 0.86 0.88 0.90 0.93 0.97 0.91 0.87 0.92 …
Figure 8
Figure 8. Figure 8: Token-Wise Cosine Similarities Across Modules in LLaMA-2-7B. E. Additional Case Studies on Motivation In this section, we present additional case studies that illustrate Token-Wise Cosine Similarities Across Modules in LLaMA2-7B, as shown in Figures 8 and 9. For LLaMA2…
Figure 9
Figure 9. Figure 9: Token-Wise Cosine Similarities Across Modules in LLaMA-2-7B. 0.0 0.2 0.4 0.6 0.8 1.0 Attn 38 MLP 38 Attn 39 MLP 39 Attn 40 MLP 40 0.95 0.98 0.99 0.96 0.99 0.98 0.98 0.97 0.98 0.96 0.96 0.98 0.91 0.92 0.91 0.98 0.93 0.94 0.94 0.93 0.95 0.89 0.92 0.93 0.98 0.98 0.96 0.96…
Figure 10
Figure 10. Figure 10: Token-Wise Cosine Similarities Across Modules in LLaMA-2-13B. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Token-Wise Cosine Similarities Across Modules in LLaMA-2-13B. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. SkipOPU: An FPGA-based Overlay Processor for Large Language Models with Dynamically Allocated Computation

    cs.AR 2026-03 conditional novelty 6.0 of 10

    An FPGA accelerator for SkipGPT-style pruned LLMs that fuses router/normalization/attention computation to hide nonlinear latency, packs two FP16 products in each DSP, and serves reused KV entries from on-chip URAM to...

  2. ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention

    cs.CV 2026-02 conditional novelty 6.0 of 10

    Frozen visual tokens with sparse cross-attention at selected layers preserve 98% accuracy while reducing vision-side FLOPs to 4% in LLaVA-1.5 models.

Reference graph

Works this paper leans on

67 extracted references · 11 canonical work pages · cited by 2 Pith papers

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Fluctuation-based adaptive structured pruning for large language models

    An, Y., Zhao, X., Yu, T., Tang, M., and Wang, J. Fluctuation-based adaptive structured pruning for large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 10865--10873, 2024

  3. [3]

    L., Nascimento, M

    Ashkboos, S., Croci, M. L., Nascimento, M. G. d., Hoefler, T., and Hensman, J. Slicegpt: Compress large language models by deleting rows and columns. arXiv preprint arXiv:2401.15024, 2024

  4. [4]

    A., Bourne, J

    Bartol, Thomas M, J., Bromer, C., Kinney, J., Chirillo, M. A., Bourne, J. N., Harris, K. M., and Sejnowski, T. J. Nanoconnectomic upper bound on the variability of synaptic plasticity. eLife, 4: 0 e10778, nov 2015. ISSN 2050-084X. doi:10.7554/eLife.10778. URL https://doi.org/10.7554/eLife.10778

  5. [5]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Bengio, Y., L \'e onard, N., and Courville, A. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013

  6. [6]

    Piqa: Reasoning about physical commonsense in natural language

    Bisk, Y., Zellers, R., Gao, J., Choi, Y., et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pp.\ 7432--7439, 2020

  7. [7]

    A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M

    Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J. Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., G...

  8. [8]

    Brown, T. B. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020

Show all 67 references
  1. [9]

    A survey on mixture of experts, 2024

    Cai, W., Jiang, J., Wang, F., Tang, J., Kim, S., and Huang, J. A survey on mixture of experts, 2024. URL https://arxiv.org/abs/2407.06204

  2. [11]

    Compressing large language models by streamlining the unimportant layer

    Chen, X., Hu, Y., and Zhang, J. Compressing large language models by streamlining the unimportant layer. arXiv preprint arXiv:2403.19135, 2024 b

  3. [12]

    Unveiling the key factors for distilling chain-of-thought reasoning, 2025

    Chen, X., Sun, Z., Guo, W., Zhang, M., Chen, Y., Sun, Y., Su, H., Pan, Y., Klakow, D., Li, W., and Shen, X. Unveiling the key factors for distilling chain-of-thought reasoning, 2025. URL https://arxiv.org/abs/2502.18001

  4. [13]

    Ee-llm: Large-scale training and inference of early-exit large language models with 3d parallelism, 2024 c

    Chen, Y., Pan, X., Li, Y., Ding, B., and Zhou, J. Ee-llm: Large-scale training and inference of early-exit large language models with 3d parallelism, 2024 c . URL https://arxiv.org/abs/2312.04916

  5. [14]

    Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., Schuh, P., Shi, K., Tsvyashchenko, S., Maynez, J., Rao, A., Barnes, P., Tay, Y., Shazeer, N., Prabhakaran, V., Reif, E., Du, N., Hutchinson, B., Pope,...

  6. [15]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Clark, C., Lee, K., Chang, M.-W., Kwiatkowski, T., Collins, M., and Toutanova, K. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019

  7. [16]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018

  8. [17]

    Redpajama: an open dataset for training large language models

    Computer, T. Redpajama: an open dataset for training large language models. 2023. URL https://github.com/togethercomputer/RedPajama-Data

  9. [18]

    Skipdecode: Autoregressive skip decoding with batching and caching for efficient llm inference

    Del Corro, L., Del Giorno, A., Agarwal, S., Yu, B., Awadallah, A., and Mukherjee, S. Skipdecode: Autoregressive skip decoding with batching and caching for efficient llm inference. arXiv preprint arXiv:2307.02628, 2023

  10. [19]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

  11. [20]

    Not all layers of llms are necessary during inference, 2024

    Fan, S., Jiang, X., Li, X., Meng, X., Han, P., Shang, S., Sun, A., Wang, Y., and Wang, Z. Not all layers of llms are necessary during inference, 2024. URL https://arxiv.org/abs/2403.02181

  12. [21]

    and Alistarh, D

    Frantar, E. and Alistarh, D. Sparsegpt: Massive language models can be accurately pruned in one-shot. In International Conference on Machine Learning, pp.\ 10323--10337. PMLR, 2023

  13. [22]

    A framework for few-shot language model evaluation, 07 2024

    Gao, L., Tow, J., Abbasi, B., Biderman, S., Black, S., DiPofi, A., Foster, C., Golding, L., Hsu, J., Le Noac'h, A., Li, H., McDonell, K., Muennighoff, N., Ociepa, C., Phang, J., Reynolds, L., Schoelkopf, H., Skowron, A., Sutawika, L., Tang, E., Thite, A., Wang, B., Wang, K., a...

  14. [23]

    Transformer feed-forward layers are key-value memories, 2021

    Geva, M., Schuster, R., Berant, J., and Levy, O. Transformer feed-forward layers are key-value memories, 2021. URL https://arxiv.org/abs/2012.14913

  15. [24]

    Gromov, A., Tirumala, K., Shapourian, H., Glorioso, P., and Roberts, D. A. The unreasonable ineffectiveness of the deeper layers. arXiv preprint arXiv:2403.17887, 2024

  16. [25]

    Gumbel, E. J. Statistical theory of extreme values and some practical applications: a series of lectures, volume 33. US Government Printing Office, 1954

  17. [26]

    What matters in transformers? not all attention is needed, 2024

    He, S., Sun, G., Shen, Z., and Li, A. What matters in transformers? not all attention is needed, 2024. URL https://arxiv.org/abs/2406.15786

  18. [27]

    Adaskip: Adaptive sublayer skipping for accelerating long-context llm inference, 2025

    He , Z., Yizhen Yao , Pengfei Zuo , Bin Gao , Qinya Li , Zhenzhe Zheng , and Fan Wu . Adaskip: Adaptive sublayer skipping for accelerating long-context llm inference, 2025. URL https://arxiv.org/abs/2501.02336

  19. [28]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

  20. [29]

    Categorical reparameterization with gumbel-softmax

    Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2022

  21. [30]

    Jiang, Y., Wang, H., Xie, L., Zhao, H., zhang chao, Qian, H., and Lui, J. C. D-llm: A token adaptive computing resource allocation strategy for large language models. In Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS 2024), 2024. URL https...

  22. [31]

    Shortened llama: A simple depth pruning for large language models

    Kim, B.-K., Kim, G., Kim, T.-H., Castells, T., Choi, S., Shin, J., and Song, H.-K. Shortened llama: A simple depth pruning for large language models. arXiv preprint arXiv:2402.02834, 11, 2024

  23. [32]

    Attention is not only a weight: Analyzing transformers with vector norms

    Kobayashi, G., Kuribayashi, T., Yokoi, S., and Inui, K. Attention is not only a weight: Analyzing transformers with vector norms. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.\ 7057--7075, 2020

  24. [33]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

  25. [34]

    Gshard: Scaling giant models with conditional computation and automatic sharding, 2020

    Lepikhin, D., Lee, H., Xu, Y., Chen, D., Firat, O., Huang, Y., Krikun, M., Shazeer, N., and Chen, Z. Gshard: Scaling giant models with conditional computation and automatic sharding, 2020. URL https://arxiv.org/abs/2006.16668

  26. [35]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization, 2019. URL https://arxiv.org/abs/1711.05101

  27. [36]

    Llm-pruner: On the structural pruning of large language models

    Ma, X., Fang, G., and Wang, X. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36: 0 21702--21720, 2023

  28. [37]

    J., Tarlow, D., and Minka, T

    Maddison, C. J., Tarlow, D., and Minka, T. A* sampling, 2015. URL https://arxiv.org/abs/1411.0030

  29. [38]

    Marcus, M., Santorini, B., and Marcinkiewicz, M. A. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19 0 (2): 0 313--330, 1993

  30. [39]

    Shortgpt: Layers in large language models are more redundant than you expect

    Men, X., Xu, M., Zhang, Q., Wang, B., Lin, H., Lu, Y., Han, X., and Chen, W. Shortgpt: Layers in large language models are more redundant than you expect. arXiv preprint arXiv:2403.03853, 2024

  31. [40]

    Locating and editing factual associations in gpt, 2023

    Meng, K., Bau, D., Andonian, A., and Belinkov, Y. Locating and editing factual associations in gpt, 2023. URL https://arxiv.org/abs/2202.05262

  32. [41]

    Pointer sentinel mixture models

    Merity, S., Xiong, C., Bradbury, J., and Socher, R. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016

  33. [42]

    Language models implement simple word2vec-style vector arithmetic, 2024

    Merullo, J., Eickhoff, C., and Pavlick, E. Language models implement simple word2vec-style vector arithmetic, 2024. URL https://arxiv.org/abs/2305.16130

  34. [43]

    Can a suit of armor conduct electricity? a new dataset for open book question answering

    Mihaylov, T., Clark, P., Khot, T., and Sabharwal, A. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018

  35. [44]

    In-context learning and induction heads

    Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Johnston, S., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., Amodei, D., Brown, T., Cla...

  36. [45]

    OpenAI, Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., Avila, R., Babuschkin, I., Balaji, S., Balcom, V., Baltescu, P., Bao, H., Bavarian, M., Belgum, J., Bello, I., Berdine, J., Bernadett-Shapi...

  37. [46]

    C., and Santoro, A

    Raposo, D., Ritter, S., Richards, B., Lillicrap, T., Humphreys, P. C., and Santoro, A. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. arXiv preprint arXiv:2404.02258, 2024

  38. [47]

    L., Bhagavatula, C., and Choi, Y

    Sakaguchi, K., Bras, R. L., Bhagavatula, C., and Choi, Y. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 0 (9): 0 99--106, 2021

  39. [48]

    From words to watts: Benchmarking the energy costs of large language model inference, 2023

    Samsi, S., Zhao, D., McDonald, J., Li, B., Michaleas, A., Jones, M., Bergeron, W., Kepner, J., Tiwari, D., and Gadepally, V. From words to watts: Benchmarking the energy costs of large language model inference, 2023. URL https://arxiv.org/abs/2310.03003

  40. [49]

    Confident adaptive language modeling

    Schuster, T., Fisch, A., Gupta, J., Dehghani, M., Bahri, D., Tran, V., Tay, Y., and Metzler, D. Confident adaptive language modeling. Advances in Neural Information Processing Systems, 35: 0 17456--17472, 2022

  41. [50]

    A., Dong, X., Heinrich, G., Breuel, T., Kautz, J., Krueger, D., and Molchanov, P

    Siddiqui, S. A., Dong, X., Heinrich, G., Breuel, T., Kautz, J., Krueger, D., and Molchanov, P. A deeper look at depth pruning of llms. arXiv preprint arXiv:2407.16286, 2024

  42. [51]

    Sleb: Streamlining llms through redundancy verification and elimination of transformer blocks

    Song, J., Oh, K., Kim, T., Kim, H., Kim, Y., and Kim, J.-J. Sleb: Streamlining llms through redundancy verification and elimination of transformer blocks. arXiv preprint arXiv:2402.09025, 2024

  43. [52]

    Llama: Open and efficient foundation language models, 2023 a

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., and Lample, G. Llama: Open and efficient foundation language models, 2023 a . URL https://arxiv.org/abs/2302.13971

  44. [53]

    Llama 2: Open foundation and fine-tuned chat models

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023 b

  45. [54]

    Accelerating llama inference by enabling intermediate layer decoding via instruction tuning with lite, 2023

    Varshney, N., Chatterjee, A., Parmar, M., and Baral, C. Accelerating llama inference by enabling intermediate layer decoding via instruction tuning with lite, 2023. URL https://arxiv.org/abs/2310.18581

  46. [55]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention is all you need. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Informatio...

  47. [56]

    Efficient large language models: A survey, 2024

    Wan, Z., Wang, X., Liu, C., Alam, S., Zheng, Y., Liu, J., Qu, Z., Yan, S., Zhu, Y., Zhang, Q., Chowdhury, M., and Zhang, M. Efficient large language models: A survey, 2024. URL https://arxiv.org/abs/2312.03863

  48. [57]

    Wang, X., Yu, F., Dou, Z.-Y., Darrell, T., and Gonzalez, J. E. Skipnet: Learning dynamic routing in convolutional networks, 2018. URL https://arxiv.org/abs/1711.09485

  49. [58]

    Chain-of-thought prompting elicits reasoning in large language models, 2023

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., and Zhou, D. Chain-of-thought prompting elicits reasoning in large language models, 2023. URL https://arxiv.org/abs/2201.11903

  50. [59]

    Sheared llama: Accelerating language model pre-training via structured pruning

    Xia, M., Gao, T., Zeng, Z., and Chen, D. Sheared llama: Accelerating language model pre-training via structured pruning. arXiv preprint arXiv:2310.06694, 2023

  51. [60]

    Llmcdsr: Enhancing cross-domain sequential recommendation with large language models

    Xin, H., Sun, Y., Wang, C., and Xiong, H. Llmcdsr: Enhancing cross-domain sequential recommendation with large language models. ACM Transactions on Information Systems, 2025

  52. [61]

    Laco: Large language model pruning via layer collapse

    Yang, Y., Cao, Z., and Zhao, H. Laco: Large language model pruning via layer collapse. arXiv preprint arXiv:2402.11187, 2024

  53. [62]

    Jump to conclusions: Short-cutting transformers with linear transformations

    Yom Din, A., Karidi, T., Choshen, L., and Geva, M. Jump to conclusions: Short-cutting transformers with linear transformations. In Calzolari, N., Kan, M.-Y., Hoste, V., Lenci, A., Sakti, S., and Xue, N. (eds.), Proceedings of the 2024 Joint International Conference on Computat...

  54. [63]

    Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019

    Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., and Choi, Y. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019

  55. [64]

    Learning to skip for language modeling, 2023

    Zeng, D., Du, N., Wang, T., Xu, Y., Lei, T., Chen, Z., and Cui, C. Learning to skip for language modeling, 2023. URL https://arxiv.org/abs/2311.15436

  56. [65]

    Finercut: Finer-grained interpretable layer pruning for large language models, 2024

    Zhang, Y., Li, Y., Wang, X., Shen, Q., Plank, B., Bischl, B., Rezaei, M., and Kawaguchi, K. Finercut: Finer-grained interpretable layer pruning for large language models, 2024. URL https://arxiv.org/abs/2405.18218

  57. [66]

    Unveiling in-context learning: A coordinate system to understand its working mechanism

    Zhao, A., Ye, F., Fu, J., and Shen, X. Unveiling in-context learning: A coordinate system to understand its working mechanism. arXiv preprint arXiv:2407.17011, 2024

  58. [67]

    Llama-moe: Building mixture-of-experts from llama with continual pre-training, 2024

    Zhu, T., Qu, X., Dong, D., Ruan, J., Tong, J., He, C., and Cheng, Y. Llama-moe: Building mixture-of-experts from llama with continual pre-training, 2024. URL https://arxiv.org/abs/2406.16554

  59. [68]

    Aligning books and movies: Towards story-like visual explanations by watching movies and reading books

    Zhu, Y., Kiros, R., Zemel, R., Salakhutdinov, R., Urtasun, R., Torralba, A., and Fidler, S. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV), December 2015

Pith tools

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