arxiv: 2604.08885 · v1 · submitted 2026-04-10 · 💻 cs.LG

Recognition: unknown

Uncertainty-Aware Transformers: Conformal Prediction for Language Models

Abhiram Vellore , Niraj K. Jha

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:21 UTC · model grok-4.3

classification 💻 cs.LG

keywords conformal predictiontransformer embeddingsuncertainty quantificationBERTprediction setscalibrationlanguage models

0 comments

The pith

Conformal prediction applied to transformer embeddings produces statistically valid prediction sets and improves accuracy on small language models.

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

The paper establishes that conformal prediction can be applied directly to the internal embeddings of fine-tuned encoder-only transformers such as BERT by building class-conditional nonconformity scores from either the [CLS] token or flattened hidden states. This produces prediction sets that come with coverage guarantees and instance-level explanations while allowing hyperparameter tuning inside the same procedure. A sympathetic reader would care because current transformer models give only point predictions whose uncertainty is hard to interpret, limiting their use in high-stakes or ambiguous-label settings. If the claim holds, early and intermediate layers turn out to yield better-calibrated representations than the final layer, and the resulting sets are both smaller and more accurate than those from prior conformal baselines.

Core claim

CONFIDE constructs class-conditional nonconformity scores from [CLS] embeddings or flattened hidden states of encoder-only transformers and feeds them into the standard conformal prediction pipeline. This yields prediction sets with finite-sample validity guarantees plus per-instance explanations. On BERT-tiny the method raises test accuracy by up to 4.09 percent and improves correct efficiency over NM2 and VanillaNN baselines; early and intermediate layers frequently produce more reliable calibration than the final layer.

What carries the argument

Class-conditional nonconformity scores built from [CLS] token embeddings or flattened hidden states inside the CONFIDE framework for conformal prediction on transformer models.

If this is right

Transformer models can supply users with both a prediction and a small set of alternatives that is guaranteed to contain the true label at a chosen rate.
Early and intermediate layers become preferred sources for uncertainty estimation rather than the final classification head.
Hyperparameter choices can be tuned jointly with the conformal procedure without breaking coverage guarantees.
High-stakes tasks with label ambiguity gain instance-level explanations that softmax probabilities do not supply.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same embedding-based nonconformity construction could be tested on decoder-only or encoder-decoder models by extracting comparable internal states.
Layer selection for calibration might become a standard diagnostic step when deploying transformers in safety-critical pipelines.
Combining CONFIDE sets with human review could reduce error rates in ambiguous domains more effectively than either alone.

Load-bearing premise

The class-conditional nonconformity scores derived from the embeddings satisfy the exchangeability condition needed for conformal prediction to deliver valid coverage.

What would settle it

A new dataset or task where the empirical coverage of the CONFIDE prediction sets falls below the nominal level or where the accuracy gain over softmax baselines disappears entirely.

Figures

Figures reproduced from arXiv: 2604.08885 by Abhiram Vellore, Niraj K. Jha.

**Figure 2.** Figure 2: Classwise and aggregate coverage curves for BERT-tiny on CoLA. The false class severely under [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

**Figure 3.** Figure 3: Classwise and aggregate coverage curves for BERT-tiny on BoolQ (non-classwise). The false class [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

**Figure 4.** Figure 4: Coverage curves on BoolQ using classwise calibration. The classwise method reduces undercoverage [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

**Figure 5.** Figure 5: Coverage curves on CoLA using classwise calibration. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: Coverage curves on CoLA using RoBERTa models. Undercoverage and classwise calibration [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: RoBERTa-large on CB (classwise): neutral class coverage fails entirely for neutral labels despite overall decent accuracy. On CoLA, NM2 similarly exhibits significant calibration issues across both BERT-tiny and RoBERTa-base backbones. As can be seen from [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Classwise coverage curves for CoLA using the NM2 prior method. Both BERT-tiny and RoBERTa [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

**Figure 9.** Figure 9: Classwise coverage curves for BoolQ using BERT-tiny with two prior methods. Both NM2 and [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

**Figure 10.** Figure 10: BERT-small hyperparameter testing on the CoLA dataset - accuracy [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗

**Figure 11.** Figure 11: BERT-small hyperparameter testing on the MRPC dataset - correct efficiency [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗

**Figure 12.** Figure 12: RoBERTa-large hyperparameter testing on the MRPC dataset - accuracy [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗

**Figure 13.** Figure 13: BERT-tiny hyperparameter testing on the RTE dataset - correct efficiency [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗

**Figure 14.** Figure 14: BERT-tiny hyperparameter testing on the SST2 dataset - accuracy [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗

**Figure 15.** Figure 15: RoBERTa-base hyperparameter testing on the CB dataset - accuracy [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗

**Figure 16.** Figure 16: Success case on BoolQ (BERT-tiny). CONFIDE separates neighbors by class, yielding a confident [PITH_FULL_IMAGE:figures/full_fig_p029_16.png] view at source ↗

**Figure 17.** Figure 17: Classwise and aggregate coverage curves for BERT-tiny on MNLI. All classes are well calibrated [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗

**Figure 18.** Figure 18: Improvement performances for CoLA with classwise calibration. [PITH_FULL_IMAGE:figures/full_fig_p031_18.png] view at source ↗

**Figure 19.** Figure 19: Improvement performances for BoolQ compared to smaller models [PITH_FULL_IMAGE:figures/full_fig_p031_19.png] view at source ↗

**Figure 20.** Figure 20: Coverage curves on CoLA using BERT-small models. Undercoverage and classwise calibration [PITH_FULL_IMAGE:figures/full_fig_p031_20.png] view at source ↗

**Figure 21.** Figure 21: BERT-small CoLA classwise-false [PITH_FULL_IMAGE:figures/full_fig_p032_21.png] view at source ↗

**Figure 23.** Figure 23: RoBERTa-base CoLA classwise-false [PITH_FULL_IMAGE:figures/full_fig_p032_23.png] view at source ↗

**Figure 26.** Figure 26: RoBERTa-large CoLA classwise-true [PITH_FULL_IMAGE:figures/full_fig_p033_26.png] view at source ↗

**Figure 27.** Figure 27: RoBERTa-base MNLI classwise-true [PITH_FULL_IMAGE:figures/full_fig_p033_27.png] view at source ↗

**Figure 29.** Figure 29: BERT-small MRPC classwise-false [PITH_FULL_IMAGE:figures/full_fig_p034_29.png] view at source ↗

**Figure 32.** Figure 32: RoBERTa-large MRPC classwise-true [PITH_FULL_IMAGE:figures/full_fig_p034_32.png] view at source ↗

**Figure 33.** Figure 33: BERT-tiny QNLI classwise-false [PITH_FULL_IMAGE:figures/full_fig_p035_33.png] view at source ↗

**Figure 36.** Figure 36: RoBERTa-base RTE classwise-true [PITH_FULL_IMAGE:figures/full_fig_p035_36.png] view at source ↗

**Figure 38.** Figure 38: BERT-small SST2 classwise-true [PITH_FULL_IMAGE:figures/full_fig_p036_38.png] view at source ↗

**Figure 40.** Figure 40: BERT-small WNLI classwise-true [PITH_FULL_IMAGE:figures/full_fig_p036_40.png] view at source ↗

**Figure 42.** Figure 42: BERT-tiny MultiRC classwise-true [PITH_FULL_IMAGE:figures/full_fig_p037_42.png] view at source ↗

read the original abstract

Transformers have had a profound impact on the field of artificial intelligence, especially on large language models and their variants. However, as was the case with neural networks, their black-box nature limits trust and deployment in high-stakes settings. For models to be genuinely useful and trustworthy in critical applications, they must provide more than just predictions: they must supply users with a clear understanding of the reasoning that underpins their decisions. This article presents an uncertainty quantification framework for transformer-based language models. This framework, called CONFIDE (CONformal prediction for FIne-tuned DEep language models), applies conformal prediction to the internal embeddings of encoder-only architectures, like BERT and RoBERTa, while enabling hyperparameter tuning. CONFIDE uses either [CLS] token embeddings or flattened hidden states to construct class-conditional nonconformity scores, enabling statistically valid prediction sets with instance-level explanations. Empirically, CONFIDE improves test accuracy by up to 4.09% on BERT-tiny and achieves greater correct efficiency (i.e., the expected size of the prediction set conditioned on it containing the true label) compared to prior methods, including NM2 and VanillaNN. We show that early and intermediate transformer layers often yield better-calibrated and more semantically meaningful representations for conformal prediction. In resource-constrained models and high-stakes tasks with ambiguous labels, CONFIDE offers robustness and interpretability where softmax-based uncertainty fails. We position CONFIDE as a framework for practical diagnostic and efficiency/robustness improvement over prior conformal baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CONFIDE applies conformal prediction to transformer embeddings for prediction sets but the coverage guarantees look under-supported from the abstract.

read the letter

CONFIDE applies conformal prediction to the internal embeddings of encoder-only transformers like BERT, building class-conditional nonconformity scores from either the [CLS] token or flattened hidden states. It reports up to 4% accuracy gains on BERT-tiny and better correct efficiency than baselines like NM2, while noting that early and intermediate layers often give more useful representations than the final output. The framework also aims to add instance-level explanations alongside the sets. That part is straightforward and potentially useful for people who already run these models and want something beyond softmax probabilities. The empirical numbers on small models are concrete enough to be worth checking. The soft spots sit in the validation. The abstract mentions hyperparameter tuning and layer selection without spelling out that these steps stayed strictly on held-out calibration data separate from test evaluation. If any of those choices touched the evaluation data, the standard conformal coverage guarantee can break. The stress-test concern about score construction introducing dependence is reasonable here; flattening hidden states or using embeddings from a model trained on related data does not automatically preserve the exchangeability needed for marginal coverage. No details on data splits, calibration set size, or explicit checks for validity appear in the summary. This paper is for researchers already working on uncertainty quantification for transformers who want a practical recipe. A reader looking for new theory will not find it, but someone implementing conformal methods on language models might pick up the layer-selection observation. I would send it to peer review so the methods section can be examined for whether the guarantees actually hold or whether extra conditions are needed for transformer internals.

Referee Report

3 major / 2 minor

Summary. The paper introduces CONFIDE, a conformal prediction framework for fine-tuned encoder-only transformers (BERT, RoBERTa). It constructs class-conditional nonconformity scores from [CLS] embeddings or flattened hidden states, permits hyperparameter tuning and layer selection, and claims to deliver statistically valid prediction sets together with instance-level explanations. Empirically, CONFIDE is reported to raise test accuracy by up to 4.09 % on BERT-tiny and to improve correct efficiency relative to NM2 and VanillaNN baselines; early and intermediate layers are said to yield better-calibrated representations.

Significance. If the coverage guarantees are rigorously established and the efficiency gains prove robust across standard splits, the work would supply a practical route to uncertainty quantification and interpretability for transformer models in high-stakes settings where softmax probabilities are insufficient. The emphasis on internal representations rather than output logits is a potentially useful extension of conformal methods to deep language models.

major comments (3)

[Abstract and §3] Abstract and §3 (Methods): The claim of statistically valid prediction sets rests on the standard conformal theorem, yet the manuscript permits hyperparameter tuning and layer selection “for better performance.” No explicit statement confirms that these choices were fixed before the calibration set was examined or that they were performed on a separate validation split, which would violate the fixed-nonconformity-function requirement.
[§4] §4 (Experiments): The abstract reports accuracy and efficiency gains, but the text supplies no description of data splits, whether the calibration set overlaps with training data, or any diagnostic that verifies marginal coverage (1-α) is attained. Without these details the empirical claims cannot be assessed for validity.
[§3.2] §3.2 (Score construction): Class-conditional nonconformity scores are built from [CLS] embeddings or flattened hidden states. The paper invokes the generic CP guarantee but provides neither a proof nor an ablation showing that the chosen score function preserves exchangeability once the underlying transformer has been trained on related data; dependence introduced by the embedding extractor could invalidate the coverage guarantee.

minor comments (2)

[Abstract] Define “correct efficiency” (expected set size conditioned on containing the true label) at first use and state its relation to standard efficiency metrics.
[§4] Add a short table or paragraph listing the exact hyper-parameters tuned and the ranges searched, even if they were chosen on a held-out validation set.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment point by point below, providing clarifications and indicating revisions to the manuscript where appropriate to strengthen the presentation of our methods and results.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (Methods): The claim of statistically valid prediction sets rests on the standard conformal theorem, yet the manuscript permits hyperparameter tuning and layer selection “for better performance.” No explicit statement confirms that these choices were fixed before the calibration set was examined or that they were performed on a separate validation split, which would violate the fixed-nonconformity-function requirement.

Authors: We agree that the nonconformity function must remain fixed prior to inspecting the calibration set. In our experiments, layer selection and all hyperparameters were determined exclusively on a held-out validation split that is completely disjoint from the calibration and test sets. We have revised §3 to explicitly document this two-stage procedure (validation for tuning, followed by fixed nonconformity scores on calibration) and updated §4 with the exact split ratios to confirm compliance with the conformal theorem. revision: yes
Referee: [§4] §4 (Experiments): The abstract reports accuracy and efficiency gains, but the text supplies no description of data splits, whether the calibration set overlaps with training data, or any diagnostic that verifies marginal coverage (1-α) is attained. Without these details the empirical claims cannot be assessed for validity.

Authors: We appreciate the need for greater transparency. The original text in §4 indicated standard held-out splits but lacked explicit percentages and coverage verification. We have expanded §4 with a dedicated paragraph describing the train/calibration/test partitioning (calibration drawn from training data but non-overlapping with test), and added a new table and figure reporting empirical coverage rates across all datasets, models, and α levels, confirming they are statistically consistent with the target 1-α. revision: yes
Referee: [§3.2] §3.2 (Score construction): Class-conditional nonconformity scores are built from [CLS] embeddings or flattened hidden states. The paper invokes the generic CP guarantee but provides neither a proof nor an ablation showing that the chosen score function preserves exchangeability once the underlying transformer has been trained on related data; dependence introduced by the embedding extractor could invalidate the coverage guarantee.

Authors: The transformer is trained on a separate training partition and then frozen; the embedding extractor therefore constitutes a fixed, deterministic function when conformal scores are subsequently computed. Because the calibration and test instances remain exchangeable, the resulting nonconformity scores inherit exchangeability. We have added a clarifying paragraph in §3.2 stating this reasoning and included an ablation (new Table X) that reports coverage for embeddings from early, intermediate, and late layers, showing the guarantee holds in practice. A full proof beyond the standard CP theorem is not supplied, as the fixed-model setting reduces to the classical case. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper applies the standard conformal prediction framework to nonconformity scores derived from transformer embeddings ([CLS] or flattened hidden states). The coverage guarantee is invoked from the general CP theorem rather than derived from the paper's own fitted quantities or self-citations. No equations reduce the prediction sets, efficiency metrics, or layer-selection benefits to a definitional loop or to a parameter fitted on the same data used for the reported test accuracy. Hyperparameter tuning is mentioned as an enabled feature but is not shown to be performed on evaluation data in a way that creates a self-referential construction. The empirical claims are presented as experimental outcomes, not as derivations that collapse back to the inputs by construction. The derivation chain therefore remains independent of the paper's own results.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only abstract available; standard conformal prediction assumptions are invoked implicitly but not detailed. No explicit free parameters or invented entities are named.

axioms (1)

domain assumption Standard exchangeability assumption for conformal prediction to guarantee coverage
Implicit in any conformal prediction application; required for validity of prediction sets.

pith-pipeline@v0.9.0 · 5573 in / 1172 out tokens · 53624 ms · 2026-05-10T17:21:07.101815+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 27 canonical work pages · 5 internal anchors

[1]

A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification

Anastasios N. Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification.arXiv preprint arXiv:2107.07511,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

URLhttps://doi.org/ 10.5194/gmd-16-1601-2023. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirec- tional transformers for language understanding.arXiv preprint arXiv:1810.04805,

work page internal anchor Pith review Pith/arXiv arXiv doi:10.5194/gmd-16-1601-2023 2023
[3]

Conformal prediction for text infilling and part-of-speech prediction.arXiv preprint arXiv:2111.02592,

Neil Dey, Jing Ding, Jack Ferrell, Carolina Kapper, Maxwell Lovig, Emiliano Planchon, and Jonathan P Williams. Conformal prediction for text infilling and part-of-speech prediction.arXiv preprint arXiv:2111.02592,

work page arXiv
[4]

Dhillon, George Deligiannidis, and Tom Rainforth

Guneet S. Dhillon, George Deligiannidis, and Tom Rainforth. On the expected size of conformal prediction sets, 2024a. URLhttps://arxiv.org/abs/2306.07254. Guneet S. Dhillon, George Deligiannidis, and Tom Rainforth. On the expected size of conformal prediction sets. In Sanjoy Dasgupta, Stephan Mandt, and Yingzhen Li, editors,Proceedings of The 27th Interna...

work page arXiv
[5]

Universal Language Model Fine-tuning for Text Classification

URLhttps://proceedings.mlr.press/v152/giovannotti21a.html. Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification.arXiv preprint arXiv:1801.06146,

work page Pith review arXiv
[6]

Linhui Huang, Sayeri Lala, and Niraj K. Jha. CONFINE: Conformal prediction for interpretable neural networks.arXiv preprint arXiv:2406.00539v2,

work page arXiv
[7]

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

work page Pith review arXiv 1902
[8]

I. T. Jolliffe and J. Cadima. Principal component analysis: A review and recent developments.Philosoph- ical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374(2065): 20150202, Apr

2065
[9]

On the societal impact of open foundation models

Sayash Kapoor, Rishi Bommasani, et al. On the societal impact of open foundation models. arXiv:2404.15958,

work page arXiv
[10]

19 Benjamin Leblanc and Pascal Germain

URLhttps://proceedings.neurips.cc/paper_files/paper/2022/ file/ce9e92e3de2372a4b93353eb7f3dc0bd-Paper-Datasets_and_Benchmarks.pdf. 19 Benjamin Leblanc and Pascal Germain. On the relationship between interpretability and explainability in machine learning.arXiv preprint arXiv:2311.11491,

work page arXiv 2022
[11]

arXiv preprint arXiv:2406.05332 , year=

Junghwan Lee, Chen Xu, and Yao Xie. Transformer conformal prediction for time series.arXiv preprint arXiv:2406.05332,

work page arXiv
[12]

Accessed 2 Nov

URLhttp://www.jstor.org/ stable/43304686. Accessed 2 Nov

work page arXiv
[13]

Journal of the American Statistical Association , volume =

doi: 10.1080/01621459.2012.751873. URLhttps://doi. org/10.1080/01621459.2012.751873. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach.arXiv preprint arXiv:1907.11692,

work page doi:10.1080/01621459.2012.751873 2012
[14]

Deep conformal prediction for robust models

Samir Messoudi, Sylvie Rousseau, and Sébastien Destercke. Deep conformal prediction for robust models. InInformation Processing and Management of Uncertainty in Knowledge-Based Systems (IPMU 2020), volume 1237 ofCommunications in Computer and Information Science, pages 528–540. Springer,

2020
[15]

Deep k-nearest neighbors: Towards confident, inter- pretable and robust deep learning

URLhttps://openai.com/research/ gpt-4o-system-card. Nicolas Papernot and Patrick McDaniel. Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning.arXiv preprint arXiv:1803.04765,

work page arXiv
[16]

Understanding softmax confidence and uncertainty.arXiv preprint arXiv:2106.04972,

Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty.arXiv preprint arXiv:2106.04972,

work page arXiv
[17]

doi: 10.48550/arXiv.2106.04972. J. Ross Quinlan. Induction of decision trees.Machine Learning, 1(1):81–106,

work page doi:10.48550/arxiv.2106.04972
[18]

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683,

work page internal anchor Pith review arXiv 1910
[19]

Mauricio Sadinle, Jing Lei, and Larry Wasserman

URLhttps://arxiv.org/abs/2006.02544. Mauricio Sadinle, Jing Lei, and Larry Wasserman. Least ambiguous set-valued classifiers with bounded error levels.Journal of the American Statistical Association, 114(525):223–234, June

work page arXiv 2006
[20]

M., KUCUKELBIR, A., MCAULIFFE, J

ISSN 1537-274X. doi: 10.1080/01621459.2017.1395341. URLhttp://dx.doi.org/10.1080/01621459.2017.1395341. Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction.Journal of Machine Learning Research, 9(12):371–421,

work page doi:10.1080/01621459.2017.1395341 2017
[21]

Daniel Sikar, Artur d’Avila Garcez, and Tillman Weyde

URLhttp://jmlr.org/papers/v9/shafer08a.html. Daniel Sikar, Artur d’Avila Garcez, and Tillman Weyde. Explorations of the softmax space: Knowing when the neural network doesn’t know...arXiv preprint arXiv:2502.00456,

work page arXiv
[22]

Axiomatic Attribution for Deep Networks

Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks.arXiv preprint arXiv:1703.01365,

work page Pith review arXiv
[23]

Attention Is All You Need

URLhttps://doi.org/10.1186/s12911-024-02444-z. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.arXiv preprint arXiv:1706.03762,

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1186/s12911-024-02444-z
[24]

doi: 10.1007/b106715

doi: 10.1007/b106715. URLhttps://link.springer.com/book/10.1007/ b106715. Vladimir Vovk, Valentina Fedorova, Ilia Nouretdinov, and Alex Gammerman. Criteria of efficiency for conformal prediction.CoRR, abs/1603.04416,

work page doi:10.1007/b106715
[25]

Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R

URLhttp://arxiv.org/abs/1603.04416. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding.arXiv preprint arXiv:1804.07461,

work page arXiv
[26]

Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-purpose language understanding systems.arXiv preprint arXiv:1905.00537,

work page internal anchor Pith review arXiv 1905
[27]

arXiv preprint arXiv:1908.04626 , year=

Sarah Wiegreffe and Yuval Pinter. Attention is not not explanation.arXiv preprint arXiv:1908.04626,

work page arXiv 1908
[28]

Over 9,000 training examples, but highly noisy and linguis- tically ambiguous

•BoolQ (Boolean Questions):Binary QA task requiring yes/no answers to naturally occurring queries paired with Wikipedia passages. Over 9,000 training examples, but highly noisy and linguis- tically ambiguous. •CB (CommitmentBank):Three-class NLI task with only∼250 training examples. Requires fine- grained reasoning over implicatures in short discourse con...

2025
[29]

same-class

Trends are monotonic on SST-2 and QNLI; they are nearly flat on MRPC, where the small training set limits variation in nearest-neighbor search time across layers. We also compare per- example latency across model sizes on MRPC (Table 11); as expected, RoBERTa-base has higher average latency than BERT-tiny. Dataset∆calibration time / block (s)∆test time / ...

work page arXiv
[30]

entailment

We observe a strong match between coverage and correct efficiency across mostϵvalues in the QNLI task usingBERT-tiny, especially whenclasswisecalibrated. This suggeststhat evenwith a limited modelcapacity, the learned representations in QNLI provide sufficient semantic separation between the “entailment” and “not-entailment” classes. QNLI tends to involve...

work page arXiv