betaFeedback

WorksMathematics

Bayes-Consistent Losses for Adaptive Inference in Classification

Abstract

Adaptive inference systems balance accuracy and inference cost by routing inputs to models with varying capabilities. Although jointly training models and routers is empirically more sample-efficient than training them separately, principled losses for joint training remain underexplored. We address this gap in a classification setting by proposing cost-sensitive hinge losses for jointly training classifiers with single-stage or cascade routers. Single-stage routers directly select a model per input, whereas cascade routers sequentially decide at each stage whether to predict with the current model or defer to the next stage, which uses a more accurate but more costly model. For both settings, we prove linear calibration bounds establishing Bayes consistency: minimizing the hinge risk recovers the Bayes-optimal risk under the target cost-sensitive zero-one loss. Moreover, since routers can require more labeled data to train than classifiers, we propose a novel post-training prediction-powered inference (PPI) procedure that uses unlabeled data to refine a trained single-stage router. Experiments on synthetic and real datasets demonstrate that the proposed hinge losses are effective for joint training and that PPI can correct routing errors when labeled data is scarce.

Introduction

Adaptive inference systems allocate computation on a per-sample basis to balance accuracy and inference cost. For each sample, they decide which one of several predictive models with different cost-accuracy profiles should be used for prediction. This framework covers a broad range of methods, including learning to defer [38], model cascades [6, 56] (Zhang et al. 2024) and dynamic neural networks (DNNs) [22], including early-exit systems [64]. We study adaptive inference in a classification setting that involves two components: an array of classifiers with varying cost-accuracy profiles, and a router module that decides, for each sample, which classifier makes the prediction.

A central challenge is how to train such systems effectively. Many approaches rely on separate training, where the classifiers are trained independently and routing is determined by a rule chosen separately, such as confidence-score thresholding [47, 26]. This approach is often sample-inefficient: each classifier is trained on the full sample space, even though the trained router may send it samples from only a small subset of the space. This mismatch may lead to suboptimal results [26] and hinder the training process [29]. In contrast, joint training aligns classifier and router objectives, encouraging each classifier to specialize in samples routed to it. Empirically, joint training substantially improves prediction accuracy at a fixed sample size, including with losses designed to align training with test-time routing [23, 67, 44, 29]. These results motivate the design of joint-training losses with consistency guarantees, which are currently scarce.

We address this gap in adaptive inference for classification under two common routing models, as illustrated in Figure 1, considering TT classifiers. Each sample has a base feature XX, available from the start, and additional features Z1,…,ZTZ_1,\ldots,Z_T, where ZtZ_t is revealed once the tt-th classifier is reached. In the single-stage (flat) model, the router directly selects a classifier tt based on XX alone, and the classifier then predicts from (X,Zt)(X,Z_t). This model covers certain configurations of elastic networks [14], learning to defer to multiple experts [58], and width-adaptive DNNs [22]. In the multi-stage (cascade) model, classifiers are invoked sequentially: at each stage tt, the router uses the accumulated features (X,Z1:t)(X,Z_{1:t}) to either exit with the tt-th classifier’s prediction or defer to the next stage, which uses a more powerful classifier. We assume that each classifier tt remains lightweight by predicting from (X,Zt)(X,Z_t) alone. This model covers cascades [6, 56] (Zhang et al. 2024), early-exit systems [64], and, more generally, depth-adaptive DNNs [22].

Illustration of flat and cascade routing models

Figure 1. Illustration of flat (left) and cascade (right) routing models. A flat router selects a single classifier ftf_t based on XX, while a cascade router sequentially invokes classifiers f1,…,fTf_1,\ldots,f_T until the first exit decision is made based on (X,Z1:t)(X,Z_{1:t}), or until the last classifier fTf_T is reached.

In our LLM experiments, for example, Z1,…,ZTZ_1,\ldots,Z_T correspond to embeddings of the answers produced by pretrained LLMs of increasing size, which serve as inputs to the classifiers (see Appendix § 14 for further justification of this modeling choice). Accessing ZtZ_t incurs a known cost κt\kappa_t, e.g., the cost of running the tt-th LLM. Such systems are typically designed so that κt\kappa_t varies with tt and is non-decreasing in tt. Due to the heterogeneous costs, jointly training the classifiers and router requires a cost-sensitive loss that balances the classification errors, measured by the zero-one loss, against the costs. We propose two such losses, one for each routing model.

Main Contributions

  • We propose hinge-based surrogates of the cost-sensitive zero-one loss for jointly training classifiers and routers in both flat and cascade models. In both cases, we prove linear calibration bounds (Theorems 3.1 and 3.2) that control the excess target risk by the excess hinge risk, thereby establishing Bayes consistency of the surrogate losses. This linear rate is sharp [20]. These results place joint training on a principled theoretical foundation. We validate our theory and demonstrate its practical utility through experiments on synthetic and real datasets.

  • A key technical contribution is the calibration bound for the cascade hinge loss. It is challenging to establish because the exit stage is a stopping time induced by a sequence of adaptive deferral decisions. We address this via a dynamic programming approach [4, 49], applying backward induction to recursively establish stage-wise calibration through a novel risk decomposition that separates excess risks into contributions from suboptimal deferral rules and suboptimal classifiers. A matching lower bound (Proposition C.8) shows that the resulting calibration constant is tight.

  • Since routers can require more labeled data to train than classifiers, we also propose a novel post-training prediction-powered inference (PPI) procedure that leverages unlabeled data to refine a trained flat router. At each input, PPI recovers the oracle routing decision with high probability, whenever the width of its confidence interval on risk improvement is narrower than the gap between the oracle and the second-best routing decisions (Theorem 4.1). This constitutes the first integration of PPI into adaptive inference.

Related Work

Learning to defer allows a model to either predict with a learned classifier or defer to one or more experts with heterogeneous capabilities and costs [33, 27, 62]. Here, only the classifier and the deferral rule are learned; the experts are fixed. Building on foundational works on Bayes consistency [3, 54, 50, 69], a line of works develops Bayes-consistent surrogate losses with a single expert [38, 57] or with multiple experts [58]. For the single-expert setting, later works establish realizable H\mathcal{H}-consistency, a guarantee relative to a restricted hypothesis class [39, 35], where Mao et al. [ [35] further identify surrogates that are simultaneously realizable H\mathcal{H}-consistent, Bayes-consistent, and admit H\mathcal{H}-consistency bounds [2]. Closer to our multi-model setting, Mao et al. [ [36] and Mao et al. [ [34] prove H\mathcal{H}-consistency bounds in the multiple-expert setting. Mao et al. [ [36] consider the standard setting in which the classifier and the deferral rule are learned jointly, while Mao et al. [ [34] consider a two-stage setting where the classifier is first trained then fixed, and the deferral rule is learned in a second stage. Finally, Mao et al. [ [37] establish realizable H\mathcal{H}-consistency, H\mathcal{H}-consistency bounds, and Bayes consistency for learning to defer to multiple experts, in both the standard and two-stage settings.

In contrast to learning to defer, which learns a single classifier, Regol et al. [ [45] and Pugnana et al. [ [42] recently studied a setting where all classifiers involved are learned: they study joint training of two classifiers with a deferral rule (or router), and their analysis is restricted to this setting. Since adaptive inference systems often involve more than two models [6, 56] (Zhang et al. 2024), we address this limitation by developing principled losses for jointly training multiple classifiers with the router under both flat and cascade models. As our work is primarily theoretical, we defer discussion of the rich literature on practical architectures and empirical performance in adaptive inference to Appendix A.

Notation

We use xi:=[x]ix_i := [x]_i to denote the ii-th entry of vector xx. We write [n]:={1,2,…,n}[n] := \{1, 2, \ldots, n\}, and for i,j∈Ni,j \in\mathbb{N} with i≤ji \leq j, we use the shorthand ai:j:={ai,ai+1,…,aj}a_{i:j} := \{a_i, a_{i+1}, \ldots, a_j\} for sequences of variables or functions. We denote the nn-dimensional probability simplex by Δn−1\Delta^{n-1}. We use uppercase letters for random variables, lowercase for their realizations (e.g., XX and xx), and calligraphic letters to denote spaces of realizations (e.g., x∈Xx \in\mathcal{X}).

Multi-Model Classification

We study a multi-model classification problem, where the goal is to learn a tuple π=(f1:T,r)\pi= (f_{1:T}, r), consisting of T≥2T \geq2 classifiers f1:Tf_{1:T} and a router rr. Each classifier ftf_t takes the form:

ft:X×Zt→[K],for t∈[T],(1)f_t : \mathcal{X} \times\mathcal{Z}_t \to[K], \quad\text{for } t \in[T], \tag*{(1)}

where X⊂Rp\mathcal{X} \subset\mathbb{R}^{p} denotes the space of the base features (e.g., a multiple-choice question prompt), [K][K] is the label space for KK classes (e.g., the answer choices), and Zt⊂Rpt\mathcal{Z}_t \subset\mathbb{R}^{p_t} denotes the space of the additional feature (e.g., LLM embeddings) available at ftf_t. Accessing Zt\mathcal{Z}_t incurs a known nonnegative cost κt\kappa_t, with κt≥κt−1\kappa_t \ge\kappa_{t-1}, reflecting that later classifiers are more accurate but more costly. The routing rule takes the form

r:H→[T],(2)r : \mathcal{H} \to[T], \tag*{(2)}

which maps the history information H⊆X×∏t=1TZt\mathcal{H} \subseteq\mathcal{X} \times\prod_{t=1}^{T} \mathcal{Z}_t available to the router to a classifier index. The space H\mathcal{H} varies with the router type, as we specify in Section 2.1. We denote by Π\Pi the class of all multi-model classifiers π=(f1:T,r)\pi= (f_{1:T}, r) satisfying (1) and (2).

For a given h∈Hh \in\mathcal{H}, the prediction y^\hat{y} of π=(f1:T,r)\pi= (f_{1:T}, r) is the output of the classifier selected by the router:

y^=fr(h)(x,zr(h)).(3)\hat{y} = f_{r(h)}(x, z_{r(h)}). \tag*{(3)}

Training data. To train π∈Π\pi\in\Pi, we use nn training samples generated i.i.d. according to

X(i),Z1:T(i),Y(i)∼i.i.d.Pfor i∈[n],(4)X^{(i)}, Z_{1:T}^{(i)}, Y^{(i)} \mathrel{\overset{\mathrm{i.i.d.}}{\sim}} P \quad\text{for } i \in[n], \tag*{(4)}

where X(i)∈XX^{(i)} \in\mathcal{X} and Zt(i)∈ZtZ_t^{(i)} \in\mathcal{Z}_t for t∈[T]t \in[T] are features, Y(i)∈[K]Y^{(i)} \in[K] is the label, and PP is the joint distribution over (X(i),Z1:T(i),Y(i))(X^{(i)}, Z_{1:T}^{(i)}, Y^{(i)}). We drop the superscript (i)(i) for a generic sample. In some cascade applications, it is natural to assume the conditional independence

Y⊥Z1:t−1∣(X,Zt),(5)Y \perp Z_{1:t-1} \mid(X, Z_t), \tag*{(5)}

where (X,Zt)(X, Z_t) subsumes all information in Z1:t−1Z_{1:t-1} relevant to predicting YY. We now formalize two routing models and their corresponding multi-model classifiers.

Single-Stage and Cascade Routing Models

The single-stage router directly selects one of the TT classifiers using only the base feature, formalized as:

Definition 2.1 (Single-stage (flat) router). With history information H=X\mathcal{H} = \mathcal{X}, a flat router takes the form

r:X→[T].(6)r : \mathcal{X} \to[T]. \tag*{(6)}

The multi-stage cascade router sequentially decides, at each stage tt, whether to exit with classifier ftf_t or defer to the more powerful classifier ft+1f_{t+1} after collecting additional features, as formalized below:

Definition 2.2 (Cascade router). With history information Ht=X×∏τ=1tZτ\mathcal{H}_t = \mathcal{X} \times\prod_{\tau=1}^{t} \mathcal{Z}_\tau available at stage t∈[T]t \in[T], a cascade router consists of stage-wise deferral rules

dt:Ht→{0,1}for t∈[T−1],(7)d_t : \mathcal{H}_t \to\{0,1\} \quad\text{for } t \in[T-1], \tag*{(7)}

where dt(ht)=0d_t(h_t)=0 indicates exit at stage tt and dt(ht)=1d_t(h_t)=1 indicates deferral to the next stage. We fix dT≡0d_T \equiv0, so the cascade always terminates at the final stage TT. Writing H=HT\mathcal{H}=\mathcal{H}_T for the full history, the stopping time induced by the deferral rules is the cascade routing rule

r(h)∈min⁡{t∈[T]:dt(ht)=0}.(8)r(h) \in\min\{t \in[T] : d_t(h_t)=0\}. \tag*{(8)}

Restricting the class Π\Pi of multi-model classifiers to flat or cascade routers (Definitions 2.1 or 2.2) yields the hypothesis classes Πflat⊂Π\Pi_{\mathrm{flat}} \subset\Pi and Πcasc⊂Π\Pi_{\mathrm{casc}} \subset\Pi, respectively, which are the focus of this work.

The Target Cost-Sensitive Zero-One Loss

To capture the tradeoff between prediction error and cost, we follow prior work on cost-sensitive classification and learning to defer [5, 36] and augment the standard zero-one loss with the costs. This yields the 0-1-c loss, denoted by ℓ:X×∏t=1TZt×[K]×Π×R≥0T→R≥0\ell:\mathcal{X}\times\prod_{t=1}^{T}\mathcal{Z}_t\times[K]\times\Pi\times\mathbb{R}_{\geq0}^{T}\to\mathbb{R}_{\geq0}. For a multi-model classifier π=(f1:T,r)∈Π\pi=(f_{1:T},r)\in\Pi and a cost vector c∈R≥0Tc\in\mathbb{R}_{\geq0}^{T} associated with f1:Tf_{1:T}, we define

ℓ(x,z1:T,y;π,c)=1{fr(h)(x,zr(h))≠y}+cr(h).(9)\ell(x,z_{1:T},y;\pi,c)=\mathbb{1}\{f_{r(h)}(x,z_{r(h)})\ne y\}+c_{r(h)}. \tag*{(9)}

That is, an incorrect prediction incurs a unit loss, while exiting at the tt-th classifier additionally incurs a cost ctc_t. In (9), the semicolon separates the random data arguments from the fixed multi-model classifier π\pi and cost vector cc. The corresponding 0-1-cc population risk is then

R(π)≡R(π;c):=E[ℓ(X,Z1:T,Y;π,c)].(10)\mathcal{R}(\pi)\equiv\mathcal{R}(\pi;c):=\mathbb{E}\left[\ell(X,Z_{1:T},Y;\pi,c)\right]. \tag*{(10)}

References

  1. [1]Anastasios N. Angelopoulos, Stephen Bates, Clara Fannjiang, Michael I. Jordan, and Tijana Zrnic. Prediction-Powered Inference. Science, 382(6671):669–674, 2023.arxiv.org/abs/2301.09633
  2. [2]Pranjal Awasthi, Anqi Mao, Mehryar Mohri, and Yutao Zhong. H-Consistency Bounds for Surrogate Loss Minimizers. In Proc. Int. Conf. Mach. Learn. (ICML), 2022.
  3. [3]Peter L Bartlett, Michael I Jordan, and Jon D McAuliffe. Convexity, Classification, and Risk Bounds. J. Am. Stat. Assoc., 101(473):138–156, 2006.DOI
  4. [4]Dimitri P. Bertsekas. Dynamic Programming and Optimal Control. 4th edition, 2017.
  5. [5]Yuzhou Cao, Tianchi Cai, Lei Feng, Lihong Gu, Jinjie Gu, Bo An, Gang Niu, and Masashi Sugiyama. Generalizing Consistent Multi-Class Classification with Rejection to Be Compatible with Arbitrary Losses. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2022.DOI
  6. [6]Lingjiao Chen, Matei Zaharia, and James Zou. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. Trans. Mach. Learn. Res. (TMLR), 2024a.arxiv.org/abs/2305.05176
  7. [7]Shuhao Chen, Weisen Jiang, Baijiong Lin, James Kwok, and Yu Zhang. RouterDC: Query-Based Router by Dual Contrastive Learning for Assembling Large Language Models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2024b.arxiv.org/abs/2409.19886
  8. [8]Yanxi Chen, Xuchen Pan, Yaliang Li, Bolin Ding, and Jingren Zhou. EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism. In Proc. Int. Conf. Mach. Learn. (ICML), 2024c.arxiv.org/abs/2312.04916
  9. [9]C. Chow. On Optimum Recognition Error and Reject Tradeoff. IEEE Trans. Inf. Theory, 16(1): 41–46, 1970.
  10. [10]C. K. Chow. An Optimum Character Recognition System Using Decision Functions. IRE Trans. Electron. Comput., EC-6(4):247–254, 1957.DOI
  11. [11]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiiichiro Nakano, Christopher Hesse, and John Schulman. Training Verifiers to Solve Math Word Problems, 2021. arXiv:2110.14168.arxiv.org/abs/2110.14168
  12. [12]Corinna Cortes, Giulia DeSalvo, and Mehryar Mohri. Learning with Rejection. In Proc. Int. Conf. Algorithmic Learn. Theory, 2016.DOI
  13. [13]Daniel Csillag, Claudio Jose Struchiner, and Guilherme Tegoni Goedert. Prediction-Powered E-Values. In Proc. Int. Conf. Mach. Learn. (ICML), 2025.arxiv.org/abs/2502.04294
  14. [14]Fnu Devvrit, Sneha Kudugunta, Aditya Kusupati, Tim Dettmers, Kaifeng Chen, Inderjit Dhillon, Yulia Tsvetkov, Hannanah Hajishirzi, Sham Kakade, Ali Farhadi, and Prateek Jain. MatFormer: Nested Transformer for Elastic Inference. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2023.arxiv.org/abs/2310.07707
  15. [15]Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and et al. The Llama 3 Herd of Models, 2024. arXiv:2407.21783.arxiv.org/abs/2407.21783
  16. [16]Ran El-Yaniv and Yair Wiener. On the Foundations of Noise-Free Selective Classification. J. Mach. Learn. Res., 11(53):1605–1641, 2010.DOI
  17. [17]Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-Adaptive Transformer. In Proc. Int. Conf. Learn. Representations (ICLR), 2020.
  18. [18]Jessie Finocchiaro, Rafael M. Frongillo, and Bo Waggoner. An Embedding Framework for the Design and Analysis of Consistent Polyhedral Surrogates. J. Mach. Learn. Res., 25(63):1–60, 2024.
  19. [19]Marc Finzi, Shikai Qiu, Yiding Jiang, Pavel Izmailov, J. Zico Kolter, and Andrew Gordon Wilson. From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence, 2026. arXiv:2601.03220.arxiv.org/abs/2601.03220
  20. [20]Rafael Frongillo and Bo Waggoner. Surrogate Regret Bounds for Polyhedral Losses. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2021.arxiv.org/abs/2110.14031
  21. [21]Peter Grünwald, Rianne de Heide, and Wouter M Koolen. Safe Testing. In Proc. Inf. Theory Appl. Workshop (ITA), 2020.
  22. [22]Y. Han, G. Huang, S. Song, L. Yang, H. Wang, and Y. Wang. Dynamic Neural Networks: A Survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):7436–7456, 2022a.arxiv.org/abs/2102.04906
  23. [23]Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfen Cao, Wenhui Huang, Chao Deng, and Gao Huang. Learning to Weight Samples for Dynamic Early-Exiting Networks. In Proc. Eur. Conf. Comput. Vis. (ECCV), 2022b.arxiv.org/abs/2209.08310
  24. [24]Kilian Hendrickx, Lorenzo Perini, Dries Van der Plas, Wannes Meert, and Jesse Davis. Machine Learning with a Reject Option: A Survey. Mach. Learn., 113(5):3073–3110, 2024.arxiv.org/abs/2107.11277
  25. [25]Radu Herbei and Marten H. Wegkamp. Classification with Reject Option. Can. J. Stat., 34(4): 709–721, 2006.DOI
  26. [26]Wittawat Jitkrittum, Neha Gupta, Aditya K Menon, Harikrishna Narasimhan, Ankit Rawat, and Sanjiv Kumar. When Does Confidence-Based Cascade Deferral Suffice? In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2023.arxiv.org/abs/2307.02764
  27. [27]Vijay Keswani, Matthew Lease, and Krishnaram Kenthapadi. Towards Unbiased and Accurate Deferral to Multiple Experts. In Proc. AAAI/ACM Conf. AI Ethics Soc. (AIES), 2021.arxiv.org/abs/2102.13004
  28. [28]Valentin Kilian, Stefano Cortinovis, and Francois Caron. Anytime-Valid, Bayes-Assisted, Prediction-Powered Inference. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2025.arxiv.org/abs/2505.18000
  29. [29]Piotr Kubaty, Bartosz Wójcik, Bartłomiej Tomasz Krzepkowski, Monika Michaluk, Tomasz Trzciński, Jary Pomponi, and Kamil Adamczewski. How to Train Your Multi-Exit Model? Analyzing the Impact of Training Strategies. In Proc. Int. Conf. Mach. Learn. (ICML), 2025.arxiv.org/abs/2407.14320
  30. [30]Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast Inference from Transformers via Speculative Decoding. In Proc. Int. Conf. Mach. Learn. (ICML), 2023.arxiv.org/abs/2211.17192
  31. [31]Zichang Liu, Jue Wang, Tri Dao, Tianyi Zhou, Binhang Yuan, Zhao Song, Anshumali Shrivastava, Ce Zhang, Yuandong Tian, Christopher Ré, and Beidi Chen. Deja Vu: Contextual Sparsity for Efficient LLMs at Inference-Time. In Proc. Int. Conf. Mach. Learn. (ICML), 2023.arxiv.org/abs/2310.17157
  32. [33]David Madras, Toniann Pitassi, and Richard Zemel. Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2018.arxiv.org/abs/1711.06664
  33. [34]Anqi Mao, Christopher Mohri, Mehryar Mohri, and Yutao Zhong. Two-Stage Learning to Defer with Multiple Experts. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2023.DOI
  34. [35]Anqi Mao, Mehryar Mohri, and Yutao Zhong. Realizable H-Consistent and Bayes-Consistent Loss Functions for Learning to Defer. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2024a.arxiv.org/abs/2407.13732
  35. [36]Anqi Mao, Mehryar Mohri, and Yutao Zhong. Principled Approaches for Learning to Defer with Multiple Experts. In Proc. Artif. Intell. and Image Analysis, 2024b.arxiv.org/abs/2310.14774
  36. [37]Anqi Mao, Mehryar Mohri, and Yutao Zhong. Mastering Multiple-Expert Routing: Realizable H-Consistency and Strong Guarantees for Learning to Defer. In Proc. Int. Conf. Mach. Learn. (ICML), 2025.arxiv.org/abs/2506.20650
  37. [38]Hussein Mozannar and David Sontag. Consistent Estimators for Learning to Defer to an Expert. In Proc. Int. Conf. Mach. Learn. (ICML), 2020.arxiv.org/abs/2006.01862
  38. [39]Hussein Mozannar, Hunter Lang, Dennis Wei, Prasanna Sattigeri, Subhro Das, and David Sontag. Who Should Predict? Exact Algorithms for Learning to Defer to Humans. In Proc. Int. Conf. Artif. Intell. Stat. (AISTATS), 2023.arxiv.org/abs/2301.06197
  39. [40]Niels Mü​ndler, Jingxuan He, Slobodan Jenko, and Martin Vechev. Self-Contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation. In Proc. Int. Conf. Learn. Representations (ICLR), 2024.
  40. [41]Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M Waleed Kadous, and Ion Stoica. RouteLLM: Learning to Route LLMs from Preference Data. In Proc. Int. Conf. Learn. Representations (ICLR), 2025.arxiv.org/abs/2406.18665
  41. [42]Andrea Pugnana, Giovanni De Toni, Cesare Barbera, Roberto Pellungrini, Bruno Lepri, and Andrea Passerini. To Ask or Not to Ask: Learning to Require Human Feedback, 2025. arXiv:2510.08314.arxiv.org/abs/2510.08314
  42. [43]Aaditya Ramdas and Ruodu Wang. Hypothesis Testing with E-Values. Found. Trends Stat., 1(1-2): 1–390, 2025.DOI
  43. [44]Florence Regol, Joud Chataoui, and Mark Coates. Jointly-Learned Exit and Inference for a Dynamic Neural Network. In Proc. Int. Conf. Learn. Representations (ICLR), 2024.arxiv.org/abs/2310.09163
  44. [45]Florence Regol, Joseph Cotnareanu, Theodore Glavas, and Mark Coates. Is the Acquisition Worth the Cost? Surrogate Losses for Consistent Two-Stage Classifiers. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2025.DOI
  45. [46]Pavel Rumiantsev, Soumyasundar Pal, Yingxue Zhang, and Mark Coates. FEval-TTC: Fair Evaluation Protocol for Test-Time Compute. In NeurIPS Workshop on Evaluating the Evolving LLM Lifecycle, 2025.arxiv.org/abs/2511.01203
  46. [47]Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. Confident Adaptive Language Modeling. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2022.arxiv.org/abs/2207.07061
  47. [48]Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. In Proc. Int. Conf. Learn. Representations (ICLR), 2017.arxiv.org/abs/1701.06538
  48. [49]Albert N Shiryaev. Optimal Stopping Rules. Springer, 2008.
  49. [50]Ingo Steinwart. How to Compare Different Loss Functions and Their Risks. Constr. Approx., 26(2): 225–287, Aug 2007.DOI
  50. [51]Manxi Sun, Wei Liu, Jian Luan, Pengzhi Gao, and Bin Wang. Mixture of Diverse Size Experts. In Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), Industry Track, 2024.arxiv.org/abs/2409.12210
  51. [52]Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge. In Proc. Conf. North Amer. Chapter Assoc. Comput. Linguist. (NAACL), 2019.arxiv.org/abs/1811.00937
  52. [53]Bernadetta Tarigan and Sara A. van de Geer. A Moment Bound for Multi-Hinge Classifiers. J. Mach. Learn. Res., 9(71):2171–2185, 2008.
  53. [54]Ambuj Tewari and Peter L. Bartlett. On the Consistency of Multiclass Classification Methods. J. Mach. Learn. Res., 8(36):1007–1025, 2007.
  54. [56]Antonios Valkanas, Soumyasundar Pal, Pavel Rumiantsev, Yingxue Zhang, and Mark Coates. C3PO: Optimized Large Language Model Cascades with Probabilistic Cost Constraints for Reasoning. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2025b.arxiv.org/abs/2511.07396
  55. [57]Rajeev Verma and Eric T. Nalisnick. Calibrated Learning to Defer with One-vs-All Classifiers. In Proc. Int. Conf. Mach. Learn. (ICML), 2022.arxiv.org/abs/2202.03673
  56. [58]Rajeev Verma, Daniel Barrejón, and Eric Nalisnick. Learning to Defer to Multiple Experts: Consistent Surrogate Losses, Confidence Calibration, and Conformal Ensembles. In Proc. Int. Conf. Artif. Intell. Stat. (AISTATS), 2023.arxiv.org/abs/2210.16955
  57. [59]Vladimir Vovk and Ruodu Wang. E-Values: Calibration, Combination and Applications. Ann. Stat., 49(3):1736–1754, 2021.DOI
  58. [60]An Wang, Xingwu Sun, Ruobing Xie, Shuaipeng Li, Jiaqi Zhu, Zhen Yang, Pinxue Zhao, Weidong Han, Zhanhui Kang, Di Wang, Naoaki Okazaki, and Cheng-zhong Xu. HMoE: Heterogeneous Mixture of Experts for Language Modeling. In Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), 2025.arxiv.org/abs/2408.10681
  59. [61]Yair Wiener and Ran El-Yaniv. Agnostic Selective Classification. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2011.
  60. [62]Bryan Wilder, Eric Horvitz, and Ece Kamar. Learning to Complement Humans. In Proc. Int. Joint Conf. Artif. Intell. (IJCAI), 2021.arxiv.org/abs/2005.00582
  61. [63]Heming Xia, Zhe Yang, Qingxiu Dong, Peiyi Wang, Yongqi Li, Tao Ge, Tianyu Liu, Wenjie Li, and Zhifang Sui. Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding. In Proc. Findings Assoc. Comput. Linguist. (ACL), 2024.arxiv.org/abs/2401.07851
  62. [64]Ji Xin, Raphael Tang, Yaoliang Yu, and Jimmy Lin. BERxiT: Early Exiting for BERT with Better Fine-Tuning and Extension to Regression. In Proc. Conf. Eur. Chapter Assoc. Comput. Linguist. (EACL), 2021.DOI
  63. [65]Yilun Xu, Shengjia Zhao, Jiaming Song, Russell Stewart, and Stefano Ermon. A Theory of Usable Information Under Computational Constraints. In Proc. Int. Conf. Learn. Representations (ICLR), 2020.arxiv.org/abs/2002.10689
  64. [66]An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Junyang Lin, Jingren Zhou, et al. Qwen2.5 technical report. Technical report, Alibaba Group, 2025.arxiv.org/abs/2412.15115
  65. [67]Haichao Yu, Haoxiang Li, Gang Hua, Gao Huang, and Humphrey Shi. Boosted Dynamic Neural Networks. In Proc. AAAI Conf. Artif. Intell. (AAAI), 2023.arxiv.org/abs/2211.16726
  66. [68]Ziqian Zeng, Yihuai Hong, Hongliang Dai, Huiping Zhuang, and Cen Chen. ConsistentEE: A Consistent and Hardness-Guided Early Exiting Method for Accelerating Language Model Inference. In Proc. AAAI Conf. Artif. Intell. (AAAI), 2024.arxiv.org/abs/2312.11882
  67. [69]Tong Zhang. Statistical behavior and consistency of classification methods based on convex risk minimization. Ann. Stat., 32(1):56–85, 2004.DOI

Paper details

vlnc:W_84bc5h8u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 14

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents