Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that the fixed-depth forward pass of a pretrained LLM can be replaced, per input and without finetuning, by a chain of layers that skips, repeats, or reorders existing layers, improving accuracy and cutting depth.

desk verdict The search space is new and the existence results are real, but the method peeks at the gold answer, so the reported accuracy gains are a fitting upper bound, not test-time adaptation. read the letter →

arxiv 2507.07996 v1 pith:7OYALTYK submitted 2025-07-10 cs.LG

Ziyue Li , Yang Li , Tianyi Zhou This is my paper
classification cs.LG
keywords test-timeadaptationlayercompositionchain-of-layersMonteCarloTreeSearchearlyexitloopedtransformerLLMinferencedepth
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 argues that a pretrained large language model should not always run all of its layers in the original order. It claims that each layer can be treated as an independent module that a test-time search may skip, repeat, or rearrange, producing a per-input Chain-of-Layers (CoLa) executed with frozen weights. A Monte Carlo Tree Search finds these paths for each sample, and the experiments report that more than 75% of originally correct samples have a shallower correct path, while more than 60% of originally wrong samples can be corrected. The point of showing this is that static, fixed-depth inference leaves a large amount of accuracy and efficiency on the table.

What carries the argument

The chain-of-layers (CoLa) search space: for a model with layers $L_1,\dots,L_N$, a path $P=[L_i,L_j,\dots]$ is any sequence formed by skipping, repeating (looping), or reordering those frozen layers, with actions that skip or repeat blocks of $k$ layers $r$ times for $k,r\in\{1,2,3,4\}$. The search engine is Monte Carlo Tree Search with a UCB score $\mathrm{UCB}(P)=\frac{Q(P)}{v(P)}+c\sqrt{\frac{\ln V}{v(P)}}-\lambda\frac{\|P\|}{N}$, where $Q(P)$ is cumulative reward (1 for a correct output, optionally penalized by normalized path length), $v(P)$ is visit count, and the length penalty pushes toward compact paths. After a fixed simulation budget, the procedure returns Pareto-optimal paths trading accuracy against depth.

What would settle it

Run the same MCTS search with the reward taken from the model's own output confidence instead of gold labels; if the share of wrong-to-correct corrections collapses to near zero, the claimed error correction depends on oracle access to answers rather than on layer composition alone.

Watch

Extended reading notes

Core claim

The paper's central claim is that a fixed forward pass is rarely the best execution path for any given input. On the paper's terms, the layers of a pretrained LLM are latent modules that can be composed into a custom Chain-of-Layers (CoLa): some layers are skipped, some are repeated like an RNN loop, and the order can change, all without updating any weight. The authors develop an MCTS procedure that searches this space per sample and claims it consistently finds higher-quality and/or shallower paths, correcting more than 60% of originally incorrect predictions and shortening more than 75% of originally correct ones. They further report that the joint skip-plus-recurrence space beats either operation alone and that error-correcting paths are often shorter than paths that merely preserve a correct answer, which they interpret as evidence of 'overthinking' in static models.

Load-bearing premise

The search is scored against the gold answer for each test input, so the reported accuracy and depth gains presuppose that the correct label is available while searching.

Editorial extensions

If this is right

  • For most test inputs the default forward path is not optimal: an MCTS search over layer compositions typically finds a different path that is at least as accurate and usually shorter.
  • Allowing skipping and recurrence together is strictly more expressive than either alone, with the largest accuracy gains appearing on the hardest DART-Math levels.
  • Correcting an originally wrong prediction usually takes fewer layers than preserving an already correct one, implying that many errors come from passing through misleading layers rather than from insufficient depth.
  • Instruction-tuned models behave differently from base models: their layers engage more evenly, and they admit less aggressive depth compression.
  • Across models, early layers are almost never skipped, middle layers are the most frequently skipped, and recurrence tends to concentrate in later layers.

Reading between the lines

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

  • Because the MCTS reward is the gold answer, the >60% error-correction headline presupposes access to correct labels at search time; a deployment version would need a calibrated confidence proxy, and nothing in the paper guarantees the gains survive without oracle feedback.
  • If per-input search can be made cheap, the same machinery suggests a test-time compute scheduler that allocates depth per query from an estimated difficulty, complementing early-exit and adaptive-computation lines of work.
  • The very large jumps on ARC (for example, LLaMA-3-3B-Base from 27.8% to 95.8% on ARC-Easy) suggest that a few high-impact 'poison' layers in the default path may be responsible for a large share of errors, which could be tested directly by ablating individual layers.
  • A concrete follow-up: train a small amortized predictor that proposes CoLa paths for new inputs without running MCTS, or distill the discovered paths into a single shallower static model.
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 / 6 minor

Summary. The paper proposes Chain-of-Layers (CoLa), a per-input layer composition space in which a frozen pretrained LLM's layers can be skipped, repeated, or reordered, and introduces an MCTS protocol to search this space. On ARC and DART-Math benchmarks, the authors report large accuracy gains over the original forward pass, as well as depth reductions for both originally correct and originally incorrect samples. The central problem is that the search reward is defined as correctness against the gold test answer, and the evaluation protocol explicitly feeds gold-answer comparisons back into the MCTS search. Consequently, the reported accuracy and depth results are oracle upper bounds: they show that there exist per-sample paths that produce the gold answer, not that a label-free test-time adaptation method can find them.

Significance. The compositional search space—jointly allowing skipping, recurrence, and reordering of pretrained layers—is novel and well motivated, and the existence question (does a fixed pretrained LLM contain better per-input paths?) is interesting. The MCTS protocol is clearly specified, with fixed hyperparameters and a reproducible sampling procedure. However, the paper's stated contribution of 'test-time depth adaptation' and its headline findings of error correction and depth reduction are not supported by the evaluation, because the search is rewarded with the gold answer for each test sample. If reframed as an oracle-guided existence proof, the work would be a useful empirical study; as written, the central practical claim is invalid.

major comments (4)
  1. [§3 Search Objective; §4 Evaluation Protocol] The search reward uses the gold test answer. Section 3 states 'A reward of 1 is assigned for a correct output,' and Section 4 states that 'model predictions are compared against gold answers' and that 'these outputs serve as feedback for MCTS to iteratively search optimized layer paths per input.' Algorithm 1's simulation step therefore evaluates each candidate path on the test input using that input's gold label. As a result, Table 1 reports the accuracy of paths selected to maximize the probability of matching the gold answer, i.e., an oracle upper bound, not the accuracy of a deployable test-time adaptation method. The findings that CoLa 'corrects' >60% of originally wrong samples and 'shortens' paths for >75% of originally correct samples follow directly from this reward definition.
  2. [§3, final paragraph; Algorithm 1] No label-free selection mechanism is provided. After the MCTS search, the paper returns Pareto-optimal paths balancing accuracy and length, but at inference time there is no way to know which of the returned paths is correct without the gold label. The paper does not propose a confidence score, a held-out validation criterion, or any surrogate for the reward. Thus, even if the paths exist, the procedure as described cannot be run at test time, and calling it 'test-time depth adaptation' is misleading.
  3. [§5.1, Finding 2 and Finding 3; Figures 2 and 4] The depth-reduction and error-correction claims are circular with respect to the search objective. For W→C samples, the path is selected precisely because it produces the gold answer; for C→C samples, the path is selected among those that already produce the gold answer and then minimized in depth. Therefore the observation that 'reducing depth helps correct errors' and the conclusion that the original model 'overthinks' are confounded by the selection criterion. The data show that an oracle can find short correct paths, not that the original model's depth is generally redundant for its actual predictions.
  4. [Table 1; §4 Datasets] There are no error bars, repeated seeds, or significance tests anywhere in the paper, and all results are based on 500 randomly sampled instances per dataset. Several numbers in Table 1 are identical across very different models (e.g., 95.80 on ARC-E and 98.20 on ARC-C appear for nearly all model variants), which suggests either a saturation effect of the oracle search or an artifact of the sampling protocol; neither possibility is discussed. This lack of uncertainty quantification undermines even the weaker existence-proof reading of the results.
minor comments (6)
  1. [Algorithm 1] The symbol N is used both for the number of layers in the model (Section 3) and as the loop variable in Algorithm 1, which is confusing; please use a different loop variable.
  2. [§3, UCB equation] The UCB formula uses v(P), V, c, and λ without defining their values in the main text; the length penalty ||P||/N should also clarify whether N is the original layer count. The value λ=5.0 appears only in Appendix B.
  3. [§4 Datasets] The description of 'randomly sample 500 instances from each dataset' is underspecified: please state the sampling seed, whether the samples are disjoint across model evaluations, and which official splits of ARC and DART-Math are used.
  4. [§5.2, Finding 4] The entropy values '3.46 v.s. 3.33' and maximum layer concentration '0.035 v.s. 0.040' are reported without stating how they are aggregated across datasets and whether the difference is statistically meaningful.
  5. [References] The reference list mixes arXiv identifiers without a consistent format; please unify the citation style.
  6. [Throughout] The claim 'without training any parameters' is technically true, but the MCTS hyperparameters (number of simulations, λ, c, random-child probability) are manually chosen and fixed; this should be stated explicitly in the main text to avoid implying that no hyperparameter choices are made.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported accuracy and depth gains are the objective of an oracle search rewarded with gold answers, so the headline findings are upper bounds rather than a deployable test-time adaptation result.

  1. fitted input called prediction [Section 4, Evaluation Protocol]
    "Each input is prompted as a natural language question, and model predictions are compared against gold answers. These outputs serve as feedback for MCTS to iteratively search optimized layer paths per input."

    The MCTS search is rewarded for correctness against the gold answer on the actual test input, so the 'optimized' path is selected to maximize correctness for that sample. Reporting the accuracy of these selected paths (e.g., the abstract's '>60% of samples with originally incorrect predictions, we can identify CoLa achieving correct predictions') reports the optimized value of the search objective, not the performance of a label-free inference method. This is per-sample fitting to the test labels, not test-time adaptation: a deployable method would need to choose paths without access to the gold answer.

  2. fitted input called prediction [Section 3, Transition and Simulation (UCB objective)]
    "A reward of 1 is assigned for a correct output, optionally minus a penalty proportional to the normalized path length to encourage compact solutions."

    Both headline findings are encoded in the search objective: correctness is the reward term and compactness is the length penalty. The claimed 'shorter CoLa for >75% of samples with correct predictions' and 'correct predictions for >60% of samples with originally incorrect predictions' are therefore the result of optimizing these exact terms. They are properties of the objective being optimized, not independent discoveries about the pretrained model's layer space.

full rationale

The paper's central experimental claim is not self-contained as a test-time adaptation method: MCTS is run per test input with a reward defined as correctness against the gold label (Section 3: 'A reward of 1 is assigned for a correct output...'; Section 4: 'model predictions are compared against gold answers. These outputs serve as feedback for MCTS'). Consequently, the reported accuracy gains and depth reductions are the optimized values of an oracle objective. They are upper bounds on what a label-free method could achieve, so presenting them as evidence of test-time depth adaptation is a fitted-input-called-prediction circularity. The descriptive layer-usage statistics (Finding 4 and 5) are not circular—they analyze the paths found under this oracle—but they do not rescue the central predictive claim. No load-bearing self-citation or imported uniqueness theorem appears; the issue is the gold-answer reward, not citation practice. Score is 6 rather than 8 because the search is not guaranteed to attain the objective (the numbers are empirical upper bounds), and the layer-engagement analyses have independent descriptive content.

Assumptions & free parameters 6 free parameters · 2 assumptions · 0 invented entities

The central claim rests on label availability during search and on the assumption that pretrained layers compose reliably. The per-sample path is the main fitted object; the MCTS hyperparameters are hand-chosen and affect the reported trade-offs.

free parameters (6)
  • Per-sample CoLa path (skip/repeat decisions) = selected by MCTS per test sample
    The path is optimized to maximize reward defined as correctness against the gold answer; the reported accuracy is the fit accuracy of this selection. This is the core fitted object of the paper.
  • Number of MCTS simulations = 200
    Appendix B: 200 simulations per input; the search budget directly controls how often a correct path is found, and hence the reported accuracy.
  • Length penalty lambda = 5.0
    Appendix B: UCB length-penalty weight; hand-chosen and affects the depth versus accuracy trade-off and all depth-reduction findings.
  • UCB exploration constant c = not reported
    Appears in the UCB formula in Section 3; hand-set but value not given.
  • Skip/repeat block sizes k, r = {1,2,3,4}
    Section 3 Actions: restricts the search space; an arbitrary restriction that shapes which paths are reachable.
  • Random unexplored child probability = 0.1
    Appendix B: fixed exploration probability in the selection logic.
assumptions (2)
  • domain assumption Transformer layers of a pretrained LLM can be skipped, repeated, and composed in new paths without finetuning and still yield valid predictions.
    Invoked throughout Section 3 and 5; no analysis of distribution shift or normalization instability is provided.
  • domain assumption Ground-truth answers are available to score candidate paths at test time.
    Section 4 Evaluation Protocol and Section 3 reward definition; this is the load-bearing premise that makes the search work, and it is not stated as a limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs." pith.science (2026). https://pith.science/paper/7OYALTYK

@misc{pith2026250707996,
  author       = {Pith},
  title        = {Pith review of: Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7OYALTYK}},
  note         = {Machine review of arXiv:2507.07996}
}
read the original abstract

Can a pretrained neural network adapt its architecture to different inputs without any finetuning? Do we need all layers for simple tasks, and are they adequate for challenging tasks? We found that the layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample. In particular, each layer from the pretrained model can be skipped/pruned or repeated multiple times as recurrent neural networks (RNN), and stacked with others in arbitrary orders, yielding a chain-of-layers (CoLa) per sample. This compositional space greatly expands the scope of existing works on looped/recurrent pretrained modules, layer pruning, or early-exit networks. We develop a Monte Carlo Tree Search (MCTS) protocol to explore and identify the optimal CoLa for each sample from math and commonsense reasoning benchmarks. Compared to a static model of a fixed depth, CoLa allows shortcut paths (fast thinking), recurrence of the same layer(s) (slow thinking), and combining both, offering more flexible, dynamic architectures for different inputs. We conduct an extensive analysis of the MCTS-optimized CoLa, which leads to two key findings: (1) For >75% of samples with correct predictions by the original LLM, we can find shorter CoLa, suggesting a large space for improving inference efficiency; (2) For >60% of samples with originally incorrect predictions, we can identify CoLa achieving correct predictions, suggesting a large space of performance enhancement. Our results highlight the shortcomings of using a fixed architecture of pre-trained LLMs for inference on different samples and pave the way to unlock the generalization power of test-time depth adaptation.

Figures

Figures reproduced from arXiv: 2507.07996 by the authors.

Figure 1
Figure 1. Test-time layer composition search space for CoLa. Starting from the original forward path, each input can dynamically skip or recurrently reuse any layer(s) to construct a custom Chain-of￾Layers (CoLa). This joint space enables both layer pruning and recurrence, supporting fast-slow depth adaptation and dynamic architecture generalization from a pretrained model without any finetuning. In this paper, we extend the … view at source ↗
Figure 2
Figure 2. Depth and non-recurrent depth (# unique layers) of CoLa on four models and seven benchmarks. The average depth of CoLa for inputs whose predictions by the original model and CoLa are both correct (C→C), and whose predictions by the original model are wrong but corrected by CoLa (W→C). Both the depth and non-recurrent depth are effectively reduced in all cases. We observe that, for both C→C (correct before and after)… view at source ↗
Figure 3
Figure 3. Accuracy–depth tradeoff on DART￾4/5 (hardest). Each point represents an architec￾ture search space (original, skip-only, recurrence￾only, CoLa) applied to pretrained or instruction￾tuned LLaMA-3-8B. CoLa consistently achieves the best tradeoff: although its depth lies between skip-only and recurrence-only strategies, it substantially improves accuracy, pushing the accuracy–cost Pareto frontier forward. We further an… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Prediction correctness transitions under CoLa. For each model and dataset, it reports the proportion of four categories of test samples: original path remains optimal, correct→correct (C→C), wrong→correct (W→C), and wrong→wrong (W→W). CoLa substantially improves predic…
Figure 5
Figure 5. Figure 5: Layer selection patterns in CoLa. (a, b) Heatmaps show the frequency of each layer being selected for 3B and 8B models, respectively, on each dataset, with darker shades indicating higher usage. (c, d) Boxplots group layers into early, middle, and late segments, reveal…
Figure 6
Figure 6. Figure 6: Skipping and recurrence rate of each layer on four models. Left: Skip rate—the proportion of CoLa in which layer-i is skipped. Right: Averaged recurrence times of layer-i in CoLa. CoLa models consistently keep early layers but exhibit an elevated skip rate of middle la…
Figure 7
Figure 7. Figure 7: Mean depth and non-recurrent depth of the shortest 5%, 10%, 20%, and 100% of valid execution paths under CoLa. To deepen our understanding beyond average-case trends, we conduct a percentile-based analysis of inference path lengths under CoLa. Instead of aggregating ov…

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. MACRO: Markov Chain Routing of Transformer Layers

    cs.CL 2026-08 conditional novelty 6.0 of 10

    A Markov-chain policy over transformer layers, decoded with top-k Viterbi, learns frozen-model routes that improve test accuracy by 5.0% on average.

Reference graph

Works this paper leans on

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

  1. [1]

    Neural module networks

    Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 39–48, 2016

    2016

  2. [2]

    Early exit optimizations for additive machine learned ranking systems

    B Barla Cambazoglu, Hugo Zaragoza, Olivier Chapelle, Jiang Chen, Ciya Liao, Zhaohui Zheng, and Jon Degenhardt. Early exit optimizations for additive machine learned ranking systems. In Proceedings of the third ACM international conference on Web search and data mining, pages 411–420, 2010

    work page 2010

  3. [3]

    Inner thinking transformer: Leveraging dynamic depth scaling to foster adaptive internal thinking

    Yilong Chen, Junyuan Shang, Zhenyu Zhang, Yanxi Xie, Jiawei Sheng, Tingwen Liu, Shuohuan Wang, Yu Sun, Hua Wu, and Haifeng Wang. Inner thinking transformer: Leveraging dynamic depth scaling to foster adaptive internal thinking. arXiv preprint arXiv:2502.13842, 2025

    arXiv 2025

  4. [4]

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

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018

    arXiv 2018

  5. [5]

    Uni- versal transformers

    Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Uni- versal transformers. arXiv preprint arXiv:1807.03819, 2018

    arXiv 2018

  6. [6]

    Faith and fate: Limits of transformers on compositionality

    Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jiang, Bill Yuchen Lin, Sean Welleck, Peter West, Chandra Bhagavatula, Ronan Le Bras, et al. Faith and fate: Limits of transformers on compositionality. Advances in Neural Information Processing Systems, 36: 70293–70332, 2023

    work page 2023

  7. [7]

    DACT-BERT: Differentiable Adaptive Computation Time for an Efficient BERT Inference

    Cristóbal Eyzaguirre, Felipe del Rio, Vladimir Araujo, and Alvaro Soto. Dact-bert: Differen- tiable adaptive computation time for an efficient bert inference.arXiv preprint arXiv:2109.11745, 2021

    work page Pith review arXiv 2021

  8. [8]

    Reducing transformer depth on demand with structured dropout

    Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019

    arXiv 1909

Show all 26 references
  1. [9]

    Looped transformers for length generalization

    Ying Fan, Yilun Du, Kannan Ramchandran, and Kangwook Lee. Looped transformers for length generalization. arXiv preprint arXiv:2409.15647, 2024

    2024 arXiv

  2. [10]

    Scaling up test-time compute with latent reasoning: A recurrent depth approach

    Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R Bartold- son, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. Scaling up test-time compute with latent reasoning: A recurrent depth approach. arXiv preprint arXiv:2502.05171, 2025

    2025 arXiv

  3. [11]

    Compressing bert: Studying the effects of weight pruning on transfer learning

    Mitchell A Gordon, Kevin Duh, and Nicholas Andrews. Compressing bert: Studying the effects of weight pruning on transfer learning. arXiv preprint arXiv:2002.08307, 2020

    2002 arXiv

  4. [12]

    Mixture of nested experts: Adaptive processing of visual tokens

    Gagan Jain, Nidhi Hegde, Aditya Kusupati, Arsha Nagrani, Shyamal Buch, Prateek Jain, Anurag Arnab, and Sujoy Paul. Mixture of nested experts: Adaptive processing of visual tokens. arXiv preprint arXiv:2407.19985, 2024

    2024 arXiv

  5. [13]

    Fastbert: a self-distilling bert with adaptive inference time

    Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Haotang Deng, and Qi Ju. Fastbert: a self-distilling bert with adaptive inference time. arXiv preprint arXiv:2004.02178, 2020

    2004 arXiv

  6. [14]

    Faster depth-adaptive transformers

    Yijin Liu, Fandong Meng, Jie Zhou, Yufeng Chen, and Jinan Xu. Faster depth-adaptive transformers. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, pages 13424–13432, 2021

    2021

  7. [15]

    Ebert: Efficient bert inference with dynamic structured pruning

    Zejian Liu, Fanrong Li, Gang Li, and Jian Cheng. Ebert: Efficient bert inference with dynamic structured pruning. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4814–4823, 2021

    2021

  8. [16]

    Anytime dense prediction with confidence adaptivity

    Zhuang Liu, Zhiqiu Xu, Hung-Ju Wang, Trevor Darrell, and Evan Shelhamer. Anytime dense prediction with confidence adaptivity. arXiv preprint arXiv:2104.00749, 2021

    2021 arXiv

  9. [17]

    On limitations of the transformer architecture

    Binghui Peng, Srini Narayanan, and Christos Papadimitriou. On limitations of the transformer architecture. In First Conference on Language Modeling, 2024. 10

    2024

  10. [18]

    Scaling language models: Methods, analysis & insights from training gopher

    Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446, 2021

    2021 arXiv

  11. [19]

    You need multiple exiting: Dynamic early exiting for accelerating unified vision language model

    Shengkun Tang, Yaqing Wang, Zhenglun Kong, Tianchi Zhang, Yao Li, Caiwen Ding, Yanzhi Wang, Yi Liang, and Dongkuan Xu. You need multiple exiting: Dynamic early exiting for accelerating unified vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision ...

    2023

  12. [20]

    Branchynet: Fast inference via early exiting from deep neural networks

    Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd international conference on pattern recognition (ICPR), pages 2464–2469. IEEE, 2016

    2016

  13. [21]

    Dart-math: Difficulty- aware rejection tuning for mathematical problem-solving

    Yuxuan Tong, Xiwen Zhang, Rui Wang, Ruidong Wu, and Junxian He. Dart-math: Difficulty- aware rejection tuning for mathematical problem-solving. 2024. URL https://arxiv.org/ abs/2407.13690

    2024 arXiv

  14. [22]

    Routing experts: Learning to route dynamic experts in multi-modal large language models

    Qiong Wu, Zhaoxi Ke, Yiyi Zhou, Xiaoshuai Sun, and Rongrong Ji. Routing experts: Learning to route dynamic experts in multi-modal large language models. arXiv preprint arXiv:2407.14093, 2024

    2024 arXiv

  15. [23]

    Deebert: Dynamic early exiting for accelerating bert inference

    Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. Deebert: Dynamic early exiting for accelerating bert inference. arXiv preprint arXiv:2004.12993, 2020

    2004 arXiv

  16. [24]

    Lgvit: Dynamic early exiting for accelerating vision transformer

    Guanyu Xu, Jiawei Hao, Li Shen, Han Hu, Yong Luo, Hui Lin, and Jialie Shen. Lgvit: Dynamic early exiting for accelerating vision transformer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9103–9114, 2023

    2023

  17. [25]

    Looped transformers are better at learning learning algorithms

    Liu Yang, Kangwook Lee, Robert Nowak, and Dimitris Papailiopoulos. Looped transformers are better at learning learning algorithms. arXiv preprint arXiv:2311.12424, 2023

    2023 arXiv

  18. [26]

    Bert loses patience: Fast and robust inference with early exit

    Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. Bert loses patience: Fast and robust inference with early exit. Advances in Neural Information Processing Systems, 33:18330–18341, 2020. 11 A Fine-Grained Depth Compression Analysis 20 22 24 26 28 303B D...

    2020

Pith tools

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