Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanism

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read COSEE claims that training each early-exit classifier on the samples that will actually exit there, simulated across multiple thresholds, closes the train-test inconsistency that limits signal-based early exiters.

desk verdict COSEE's training recipe is worth a look, but its 'consistently outperforms' claim is not supported by the unmatched speed-up comparisons and borrowed baselines in Table 2. read the letter →

arxiv 2412.13236 v1 pith:PN6D55AH submitted 2024-12-17 cs.LG cs.AI

Jianing He , Qi Zhang , Hongyun Zhang , Xuanjing Huang , Usman Naseem , Duoqian Miao This is my paper · ORCID
classification cs.LGcs.AI
keywords earlyexitingpre-trainedlanguagemodelssampleweightingtraining-testingconsistencyenergyscoreGLUEbenchmarkinferenceefficiencymulti-exitnetworks
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

COSEE is trying to establish that the main bottleneck in signal-based early exiting for pre-trained language models is a mismatch between training and inference: during training every internal classifier treats every sample equally, but at test time easy samples exit early and hard samples travel deeper. The paper's proposed fix is a calibrated sample weighting mechanism that, at each training step, simulates the test-time exiting process under several randomly chosen thresholds, finds where each sample would exit, and up-weights that sample's loss mainly for classifiers near its exit layer. An online signal calibration objective and a normalized energy score make the exiting signal itself more discriminative, so the simulated exits are trustworthy. If the claim is right, a nearly parameter-free change to the training loss improves accuracy at matched speed-up ratios and keeps the ability to adjust the speed-up ratio after training.

What carries the argument

The load-bearing mechanism is the sample weighting mechanism built on simulated exit layers. At each training step the model runs each sample through all layers, reads off an exiting signal, draws $K$ thresholds, and computes $m^*_n$ from the current early-exit rule; the loss weight for classifier $m$ is $w_n^{(m)} = e^{-\beta_t |m - m^*_n|} / \sum_{m'} e^{-\beta_t |m' - m^*_n|}$, with $\beta_t$ warmed up from $\beta_0$ so unstable early-training exits do not distort the weights. Around it, the online signal calibration objective $\max(0, S_{\text{easy}} - S_{\text{hard}} + \epsilon)$ explicitly widens the gap between easy and hard samples' exiting signals, and the normalized energy score $E_{\text{norm}} = (1 + e^{-E})^{-1}$ aligns energy distributions across layers so a single threshold is meaningful everywhere.

What would settle it

Replay a trained COSEE model's training-time simulation: for a fixed threshold, record the $m^*_n$ that SWM would have computed for each development sample from stored per-layer signals, compare it with the layer where that sample actually exits at inference under the same threshold, and compute the disagreement rate; a large disagreement rate would mean the claimed train-test consistency is not achieved.

Watch

Extended reading notes

Core claim

On the paper's own terms, COSEE establishes that consistency between training and testing can be achieved without sacrificing flexible speed-up. The training objective is the mean over $K$ randomly sampled thresholds of a sample-weighted classification loss, where the weight of sample $n$ at classifier $m$ is a soft exponential function of $|m - m^*_n|$ and $m^*_n$ is the layer at which that sample would exit under the current exiting signal at threshold $\tau$. Because the weights are soft rather than one-hot, the model generalizes to thresholds not seen in training. The paper reports that on the GLUE benchmark with BERT-base this yields a better performance-efficiency trade-off than existing signal-based and router-based early exiters, sometimes matching or beating the full BERT-base at roughly $2\times$ speed-up, and that the same framework transfers to ALBERT and to entropy and softmax signals.

Load-bearing premise

The whole weighting scheme rests on the exiting signal being a reliable ranking of sample difficulty from the start of training: the simulated exit layer that sets every sample's weights is read off that signal, and if the signal is miscalibrated early, the weights train classifiers to emphasize the wrong samples.

Editorial extensions

If this is right

  • At matched speed-ups of roughly $2.00\times$, COSEE reports higher accuracy than DeeBERT, PABEE, BERxiT, LeeBERT, GPFEE, GAML-BERT, PALBERT, DisentangledEE, and ConsistentEE on the GLUE tasks, with especially clear gains at high speed-up.
  • Because training averages over $K$ thresholds, the trained multi-exit network can later be operated at any threshold: the acceleration ratio is adjusted at inference time, not fixed by the training router.
  • The OSC objective raises the difficulty-inversion score of the exiting signal, so exit decisions and the weights that depend on them are driven by a more reliable difficulty proxy.
  • Switching to SWM alone speeds convergence because each classifier concentrates on a difficulty slice of the data, while COSEE adds only classifier parameters, under 0.03% of BERT-base.
  • COSEE reports lower premature and delayed exiting rates than conventional training, meaning the resulting exit decisions are more reliable under strict speed-up settings.

Reading between the lines

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

  • Extension: The weighting scheme is a continuum between uniform training and router-based one-hot training, controlled by $\beta_t$; a natural extension the paper does not explore is scheduling or adapting $\beta$ per classifier or per task.
  • Extension: If the simulated exit layer is a valid difficulty proxy, the same threshold-simulation loss could be applied to other multi-exit architectures beyond text, for example vision or speech models.
  • Extension: The paper's own premature and delayed exiting rate statistics suggest an immediate diagnostic: those rates could be monitored during training and used to choose $\beta_0$ or to stop training early, since COSEE claims to reduce both.
  • Extension: Because COSEE depends on the signal ranking samples correctly, a direct stress test would be to permute the signal values during training; if the gains persist, the weighting mechanism is not actually carrying the result.
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 introduces COSEE, a training-time modification of signal-based early exiting for pre-trained language models. At each training step, COSEE randomly samples K exiting thresholds, simulates test-time early exiting to obtain a pseudo exit layer m*_n per sample, and weights the cross-entropy loss of each internal classifier by a softmax over distance to that exit layer (Eq. 3). It adds an online signal calibration (OSC) loss (Eqs. 6-8) that pushes the exiting signals of easy and hard samples apart, and it adopts a normalized energy score (Eq. 11) as the default signal. Experiments on six GLUE tasks with BERT-base and ALBERT-base claim that COSEE consistently outperforms prior early-exit methods at similar speed-ups, supplemented by ablations, convergence curves, and failure-case statistics.

Significance. If the empirical claims held, COSEE would be a valuable and simple contribution: SWM is parameter-free, the method adds negligible parameters, it is evaluated across multiple exiting signals and backbones, and code is released. The paper also contains useful ablations (Figures 4-6) and a failure-case analysis (Appendix D). However, the headline claim of consistent state-of-the-art outperformance is not currently supported by the reported evidence because the comparison protocol in Table 2 does not match speed-up ratios or training setups across methods. With a properly matched evaluation, the method may well be competitive; the contribution's value depends on that empirical confirmation.

major comments (4)
  1. [Section 4.4 / Table 2] The claim that COSEE 'consistently outperforms the state-of-the-art methods across all tasks' is not supported because comparisons are made at different speed-up ratios. For example, on RTE COSEE scores 68.7 at 1.96x while ConsistentEE scores 69.0 at 1.85x; on MNLI both score 83.4 but COSEE is at 1.92x while ConsistentEE is at 1.45x; on QNLI COSEE's 90.2 at 2.56x is compared with ConsistentEE's 89.9 at 1.72x. Since accuracy generally declines as speed-up increases, a point that is better on one axis and worse on the other does not establish a superior trade-off. Please report results at matched speed-ups (e.g., tune every method to approximately 2.00x) or present full Pareto curves with error bars.
  2. [Section 4.2 / Table 2] Most baseline numbers are taken from the original papers (DeeBERT, PABEE, BERxiT, etc.) or from GPFEE and DisentangledEE, so training data, hyperparameters, and inference settings are not matched across methods. No multiple seeds or variance estimates are reported. This makes any 'better trade-off' claim fragile. Please rerun the baselines under the same protocol (same backbone checkpoint, same batch-size-1 inference, same seed) or, at minimum, clearly mark borrowed numbers and report variance for every method.
  3. [Section 4.3] The text states that thresholds are adjusted to achieve 'similar speed-up ratio ... approximately 2.00x (±38%)', but the speed-ups in Table 2 range from 1.45x to 2.86x, and COSEE's own ratios range from 1.92x to 2.70x. A ±38% tolerance does not constitute matched speed-up. Please report the threshold used for each method and task, and either enforce a tighter tolerance or analyze the trade-off with interpolation.
  4. [Section 3.3 / Eq. (7)] Both SWM and OSC rely on the exiting signal's ability to rank sample difficulty. The paper acknowledges that early in training signals may not reflect difficulty, and the easy/hard partition is based on the internal classifier's own correctness, which can be noisy. The warm-up in Eq. (4) and the OSC objective are patches, not guarantees. The current evidence that they work is limited: Figure 5 and Appendix D are shown for SST-2 (and QNLI) only, without confidence intervals or comparison to alternative calibration strategies. Please provide quantitative evidence of signal reliability across all tasks, e.g., DIS curves and premature/delayed exiting rates at matched speed-ups, to show that the weighting direction is not sometimes reversed.
minor comments (6)
  1. [Table 2 caption] The caption says 'Best results are marked in bold', but the bold row appears to be COSEE throughout; on RTE and MNLI, COSEE is not the best. Please correct either the bold marking or the caption.
  2. [Abstract and Section 4.4] The phrase 'consistently outperforms the state-of-the-art methods' should be softened to 'is competitive with' or 'often outperforms', given the unmatched comparisons in Table 2.
  3. [Table 3] The ALBERT comparison has the same unmatched speed-up issue and uses borrowed baselines; please add matched-speed-up comparisons or explicitly state the limitation.
  4. [Eq. (11)] Please clarify the monotonic direction of the normalized energy score relative to sample difficulty; currently a reader must infer it from Figure 3.
  5. [Figure 2 caption] The text 'objectiveLossOSC' should be rendered as 'objective Loss_OSC' with proper math formatting.
  6. [Appendix A] Parameter sensitivity is shown only on SST-2 and QNLI; please state whether the chosen defaults generalize to other tasks or provide sensitivity results on at least one additional task.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity: COSEE's self-referential training weighting is a heuristic feedback loop, not a fitted-input prediction; the empirical claims stand independent.

  1. other [Section 3.2, Eq. (2)-(5) and Section 3.3, Eq. (6)-(9)]
    "at each training step, all samples are passed through the entire model to generate predictions and exiting signals at all classifiers. Subsequently, we randomly select K thresholds and simulate the early exiting process based on exiting signals at each threshold to find where the samples exit. This information is used to produce sample-wise loss weights across all classifiers."

    The simulated exiting layer m*_n in Eq. (3) is obtained by applying the same signal-threshold rule that will be used at inference to the current model's own predictions. Weighting each classifier's loss by proximity to m*_n makes the training objective reinforce the model's current exit assignments, so the consistency between training and testing is partly built into the objective rather than independently established. This is a self-referential optimization loop, but it is a training heuristic rather than a fitted parameter renamed as a prediction; the reported accuracy/efficiency gains are empirical and could in principle fail, so the circularity is minor.

full rationale

COSEE's core mechanism assigns per-sample loss weights from the model's own simulated exiting layers, and the OSC objective calibrates exiting signals using easy/hard partitions defined by the model's own classifiers. This creates a feedback loop, but it does not reduce any reported result to an input by construction: no parameter is fitted to a subset of data and then reported as a prediction, no external result is imported solely through self-citation, and no uniqueness theorem is invoked to make the design forced. The comparison in Table 2 does mix speed-up ratios and borrows baseline numbers from their original papers, but that is an experimental-protocol weakness, not circularity. The central empirical claim, that COSEE yields a better accuracy/efficiency trade-off, is externally falsifiable on GLUE and would have been false if the training objective had not improved generalization. Thus the paper is largely self-contained; I assign 2 for the self-referential training design rather than any load-bearing circular derivation.

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

The method introduces no new entities or physical constants. Its central dependence is on four hyperparameters and three domain assumptions about signal validity and weighting geometry.

free parameters (4)
  • beta_0 = per task, e.g., 10.0 for SST-2, 1.0 for QNLI
    Decay factor base in Eq. 4, grid-searched over {0.05, 0.2, 1.0, 10.0}; optimal value differs across tasks per Appendix A.
  • alpha = 0.1 for SST-2, 0.01 for QNLI
    Balance weight between classification and OSC losses in Eq. 9, grid-searched; authors say 0.01 to 0.1 works well.
  • epsilon = 0.3
    Margin in OSC loss, Eq. 7, fixed to 0.3 after sensitivity analysis in Appendix A.
  • K = 5
    Number of random thresholds sampled per training step in Eq. 5, fixed to 5 for computational efficiency.
assumptions (5)
  • domain assumption Exiting signals such as energy, entropy, and softmax score monotonically reflect sample difficulty, so the early-exit rule at a threshold yields the proper exiting layer.
    Used throughout Section 2.2 and 3.5 to define the early-exit rule and the simulated exit layers in Eq. 3.
  • domain assumption Easy and hard samples can be identified by whether the internal classifier predicts correctly, and this partition is used to calibrate signals.
    Section 3.3 defines easy/hard by classifier correctness to construct the OSC loss in Eq. 7.
  • domain assumption Symmetric exponential weighting around the exiting layer is a beneficial inductive bias for training internal classifiers.
    Eq. 3 assigns weights as e^{-beta |m - m*|}; the authors state symmetry is chosen for easy parameter selection but provide no derivation.
  • domain assumption Averaging over K random thresholds at training time captures the full range of acceleration scenarios seen at test time.
    Section 3.2, Eq. 5; the choice of K=5 is empirical and not theoretically justified.
  • standard math Standard cross-entropy loss and sigmoid normalization of energy scores are used as given.
    Eq. 1, 10, 11.

how reviews work

0 comments
Cite this review

Pith. "Pith review of COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanism." pith.science (2026). https://pith.science/paper/PN6D55AH

@misc{pith2026241213236,
  author       = {Pith},
  title        = {Pith review of: COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanism},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PN6D55AH}},
  note         = {Machine review of arXiv:2412.13236}
}
read the original abstract

Early exiting is an effective paradigm for improving the inference efficiency of pre-trained language models (PLMs) by dynamically adjusting the number of executed layers for each sample. However, in most existing works, easy and hard samples are treated equally by each classifier during training, which neglects the test-time early exiting behavior, leading to inconsistency between training and testing. Although some methods have tackled this issue under a fixed speed-up ratio, the challenge of flexibly adjusting the speed-up ratio while maintaining consistency between training and testing is still under-explored. To bridge the gap, we propose a novel Consistency-Oriented Signal-based Early Exiting (COSEE) framework, which leverages a calibrated sample weighting mechanism to enable each classifier to emphasize the samples that are more likely to exit at that classifier under various acceleration scenarios. Extensive experiments on the GLUE benchmark demonstrate the effectiveness of our COSEE across multiple exiting signals and backbones, yielding a better trade-off between performance and efficiency.

Figures

Figures reproduced from arXiv: 2412.13236 by the authors.

Figure 1
Figure 1. Exiting layer distribution on the QNLI develop [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison between the conventional signal-based early exiting framework and our COSEE. The conventional frame [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Energy distribution across layers 2, 6, and 10 for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 5
Figure 5. Figure 5: DIS heatmap of different models at different layers [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 4
Figure 4. Figure 4: Impact of SWM and OSC on the trade-off between [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Impact of SWM and OSC on training convergence [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Exiting layer distribution on the training and de [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Impact of SWM and OSC on the trade-off between [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Impact of SWM and OSC on the trade-off between [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Sensitivity analysis of β0, α, ϵ, and K on the SST￾2 and QNLI tasks. sp denotes the speed-up ratio. For each task, we start with a set of parameters (β0 = 10.0, α = 0.1, ϵ = 0.3, K = 5 for the SST-2 task, β0 = 1.0, α = 0.01, ϵ = 0.3, K = 5 for the QNLI task) and adjus…
Figure 11
Figure 11. Figure 11: Statistics of failure cases for each training method [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]

Discussion (0). Continue with ORCID 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 of Early Exit Deep Neural Networks in NLP

    cs.LG 2025-01 conditional novelty 3.0 of 10

    A review of early exit deep neural network methods in NLP that has no new experiments but organizes the existing literature.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Akbari, M.; Banitalebi-Dehkordi, A.; and Zhang, Y. 2022. E-lang: Energy-based joint inferencing of super and swift language models. arXiv preprint arXiv:2203.00748

    arXiv 2022

  4. [4]

    Balagansky, N.; and Gavrilov, D. 2022. Palbert: Teaching albert to ponder. Advances in Neural Information Processing Systems, 35: 14002--14012

    work page 2022

  5. [5]

    Devlin, J.; Chang, M.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL-HLT (1) , 4171--4186. Association for Computational Linguistics

    work page 2019

  6. [6]

    Gao, X.; Zhu, W.; Gao, J.; and Yin, C. 2023. F-PABEE: flexible-patience-based early exiting for single-label and multi-label text classification tasks. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5. IEEE

    work page 2023

  7. [7]

    Ji, Y.; Wang, J.; Li, J.; Chen, Q.; Chen, W.; and Zhang, M. 2023. Early exit with disentangled representation and equiangular tight frame. In Findings of the Association for Computational Linguistics: ACL 2023, 14128--14142

    work page 2023

  8. [8]

    Kaya, Y.; Hong, S.; and Dumitras, T. 2019. Shallow-Deep Networks: Understanding and Mitigating Network Overthinking. In ICML , volume 97 of Proceedings of Machine Learning Research, 3301--3310. PMLR

    work page 2019

Show all 29 references
  1. [9]

    Lan, Z.; Chen, M.; Goodman, S.; Gimpel, K.; Sharma, P.; and Soricut, R. 2020. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In ICLR . OpenReview.net

    2020

  2. [10]

    Li, L.; Lin, Y.; Chen, D.; Ren, S.; Li, P.; Zhou, J.; and Sun, X. 2021. CascadeBERT: Accelerating Inference of Pre-trained Language Models via Calibrated Complete Models Cascade. In EMNLP (Findings) , 475--486. Association for Computational Linguistics

    2021

  3. [11]

    Liao, K.; Zhang, Y.; Ren, X.; Su, Q.; Sun, X.; and He, B. 2021. A Global Past-Future Early Exit Method for Accelerating Inference of Pre-trained Language Models. In NAACL-HLT , 2013--2023. Association for Computational Linguistics

    2021

  4. [12]

    Liu, W.; Zhou, P.; Wang, Z.; Zhao, Z.; Deng, H.; and Ju, Q. 2020. FastBERT: a Self-distilling BERT with Adaptive Inference Time. In ACL , 6035--6044. Association for Computational Linguistics

    2020

  5. [13]

    Liu, Y.; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy, O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR, abs/1907.11692

    2019 arXiv

  6. [14]

    Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In ICLR (Poster) . OpenReview.net

    2019

  7. [15]

    Mangrulkar, S.; MS, A.; and Sembium, V. 2022. BE3R: BERT based Early-Exit Using Expert Routing. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 3504--3512

    2022

  8. [16]

    Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; Sutskever, I.; et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8): 9

    2019

  9. [17]

    Schwartz, R.; Stanovsky, G.; Swayamdipta, S.; Dodge, J.; and Smith, N. A. 2020. The Right Tool for the Job: Matching Model and Instance Complexities. In ACL , 6640--6651. Association for Computational Linguistics

    2020

  10. [18]

    Sun, T.; Liu, X.; Zhu, W.; Geng, Z.; Wu, L.; He, Y.; Ni, Y.; Xie, G.; Huang, X.; and Qiu, X. 2022. A Simple Hash-Based Early Exiting Approach For Language Understanding and Generation. In ACL (Findings) , 2409--2421. Association for Computational Linguistics

    2022

  11. [19]

    Wang, A.; Singh, A.; Michael, J.; Hill, F.; Levy, O.; and Bowman, S. R. 2019. GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. In ICLR (Poster) . OpenReview.net

    2019

  12. [20]

    L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A

    Wolf, T.; Debut, L.; Sanh, V.; Chaumond, J.; Delangue, C.; Moi, A.; Cistac, P.; Rault, T.; Louf, R.; Funtowicz, M.; Davison, J.; Shleifer, S.; von Platen, P.; Ma, C.; Jernite, Y.; Plu, J.; Xu, C.; Scao, T. L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A. M. 2020. Transforme...

    2020

  13. [21]

    Xin, J.; Tang, R.; Lee, J.; Yu, Y.; and Lin, J. 2020. DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference. In ACL , 2246--2251. Association for Computational Linguistics

    2020

  14. [22]

    Xin, J.; Tang, R.; Yu, Y.; and Lin, J. 2021. BERxiT: Early Exiting for BERT with Better Fine-Tuning and Extension to Regression. In EACL , 91--104. Association for Computational Linguistics

    2021

  15. [23]

    Zeng, Z.; Hong, Y.; Dai, H.; Zhuang, H.; and Chen, C. 2024. ConsistentEE: A Consistent and Hardness-Guided Early Exiting Method for Accelerating Language Models Inference. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 19506--19514

    2024

  16. [24]

    Zhang, J.; Tan, M.; Dai, P.; and Zhu, W. 2023. Leco: Improving early exiting via learned exits and comparison-based exiting mechanism. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), 298--309

    2023

  17. [25]

    Zhang, Z.; Zhu, W.; Zhang, J.; Wang, P.; Jin, R.; and Chung, T. 2022. PCEE-BERT: Accelerating BERT Inference via Patient and Confident Early Exiting. In NAACL-HLT (Findings) , 327--338. Association for Computational Linguistics

    2022

  18. [26]

    J.; Xu, K.; and Wei, F

    Zhou, W.; Xu, C.; Ge, T.; McAuley, J. J.; Xu, K.; and Wei, F. 2020. BERT Loses Patience: Fast and Robust Inference with Early Exit. In NeurIPS

    2020

  19. [27]

    Zhu, W. 2021. LeeBERT: Learned early exit for BERT with cross-level optimization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 2968--2980

    2021

  20. [28]

    Zhu, W.; Wang, P.; Ni, Y.; Xie, G.; and Wang, X. 2023. BADGE: speeding up BERT inference after deployment via block-wise bypasses and divergence-based early exiting. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry ...

    2023

  21. [29]

    Zhu, W.; Wang, X.; Ni, Y.; and Xie, G. 2021. GAML-BERT: improving BERT early exiting by gradient aligned mutual learning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 3033--3044

    2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.