Pith. sign in

REVIEW 5 major objections 4 minor 32 references

CoVeR calibrates token clusters to give autoregressive decoding a 1−α coverage guarantee that does not decay with sequence length.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

CoVeR is a cluster-aware conformal decoding method that claims full-sequence coverage for LLM outputs without the (1-alpha)^L decay of prior conformal beam search.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A promising conformal decoding scheme for long-tail sequences, but the advertised PAC guarantee does not follow from the theorem, and there are no experiments. the 5 major comments →

arxiv 2509.04733 v1 pith:UOG3K2OU submitted 2025-09-05 cs.LG cs.AI

CoVeR: Conformal Calibration for Versatile and Reliable Autoregressive Next-Token Prediction

Yuzhu Chen , Yingjie Wang , Shunyu Liu , Yongcheng Jing , Dacheng Tao This is my paper
classification cs.LG cs.AI
keywords conformal predictionautoregressive decodingnext-token predictioncoverage guaranteePAC boundlong-tail trajectoriesbeam searchrisk control
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper proposes CoVeR, a conformal-prediction-based decoding strategy for autoregressive models. It aims to produce a candidate set of full reasoning sequences that covers the desired trajectory with probability at least 1−α, while keeping the search space compact and retaining long-tail solutions. The central advance is a finite-sample PAC-style bound that avoids the exponential coverage decay of earlier stepwise conformal beam search, so the guarantee does not weaken as sequences grow longer. This matters because existing decoding methods either lack provable coverage or sacrifice long-tail coverage for efficiency.

Core claim

CoVeR decomposes the full-sequence non-coverage event into a disjoint union of cluster-step events, where each event is a token at a given step falling in a learned cluster and failing to pass that step's conformal threshold. It learns per-step token clusters from quantile embeddings of conformity scores and fits a separate quantile threshold for each cluster-step pair under a dual objective: maximize thresholds to shrink the candidate set while enforcing full-path coverage of at least 1−α on a calibration set. The paper proves that, with high probability over the calibration data, the full-path failure probability is at most α plus variance-sensitive finite-sample terms that shrink as calib

What carries the argument

The cluster-step decomposition is the load-bearing object: the full non-coverage event is partitioned into disjoint events E_{l,m} where the first uncovered failure occurs at step l for a token assigned to cluster m. Per-cluster quantile levels β_{m,l} then let the algorithm assign heterogeneous thresholds (tight for easy clusters, loose for long-tail clusters). The proof reframes each threshold decision as a binary classifier and applies the empirical Bernstein inequality, giving a PAC bound expressed as α plus sums over clusters of variance- and sample-size-dependent terms.

Load-bearing premise

The proof treats the learned token clusters and the fitted per-cluster thresholds as fixed before calibration, and assumes the clustering 'successfully partitions' the class-step space; the paper gives no bound on the error introduced by solving the threshold optimization on the same calibration data used to estimate coverage.

What would settle it

Run CoVeR on a synthetic autoregressive model with known token score distributions, using increasing calibration size N and deliberately corrupted cluster labels (e.g., randomly permuting clusters before threshold optimization). If the empirical full-path coverage consistently falls below 1−α for large N by a margin that does not shrink, the guarantee fails under imperfect clustering as stated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Users of any autoregressive model can obtain a candidate set of full reasoning paths that covers the desired trajectory with probability at least 1−α, with the guarantee not weakening as the number of reasoning steps grows.
  • Per-cluster thresholds allow easy clusters to receive tight sets while preserving room for long-tail tokens, so coverage and compactness are not opposing goals.
  • The cluster-step decomposition turns global failure into traceable local terms, enabling diagnosis of which step and token type lost coverage.
  • The method is model-free: it wraps a pre-trained autoregressive model and uses only a calibration set, so it can be applied after training without modifying the model.
  • Asymptotically, the PAC bound ensures the full-sequence coverage rate converges to at least 1−α for any target α in (0,1).

Where Pith is reading between the lines

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

  • The same cluster-step decomposition could apply to other sequential decision problems, such as planning or theorem proving, where success requires a full trajectory to survive a cascade of filters.
  • The reliance on clustering quality suggests a testable extension: replacing the quantile embeddings with learned representations of token score distributions might tighten the bound or reduce the calibration sample size needed.
  • Because the theorem treats the fitted thresholds β as fixed while Algorithm 1 optimizes them on the same calibration set, a natural variant would hold out an additional fold for threshold selection, accepting a slightly larger finite-sample term in exchange for a cleaner guarantee.
  • If the guarantee holds, it implies a practical auditing protocol: tally full-path failures by cluster-step pair and compare each pair's contribution to the per-cluster slack in the bound, giving a direct diagnosis of long-tail undercoverage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes CoVeR, a conformal decoding strategy for autoregressive next-token prediction that clusters token-step distributions into local regions, learns cluster-specific quantile thresholds via a constrained optimization (Eq. 7), and claims a PAC-style coverage guarantee for full sequences at user-specified level 1-alpha. The theoretical core is Theorem 1, which bounds full-path failure probability by alpha plus positive concentration terms, based on a decomposition of non-coverage into disjoint cluster-step events. The paper also claims that unlike prior methods, the guarantee does not decay exponentially with sequence length L and that the method supports long-tail sequences through cluster-conditional calibration. No experiments or code are provided.

Significance. If the stated guarantee were valid, the paper would make a valuable contribution to conformal decoding: a length-independent, cluster-adaptive full-sequence coverage bound would improve on the (1-alpha)^L decay of dynamic conformal beam search. The decomposition of full-path failure into first-failure cluster-step events is a sensible conceptual step. However, the central theorem does not establish the paper's own PAC definition, the proof treats the data-dependent threshold optimization as fixed, and the clustering assumption is left undefined. With no experimental validation, the practical claims also remain unsubstantiated. The contribution is therefore currently a promising idea with an unproven core guarantee.

major comments (5)
  1. [Section 4.5, Eq. (9) and Theorem 1] The PAC guarantee is defined as P_{Zval}[P_{X,S}[S notin C^(L)(X)] < alpha] > 1 - delta. Theorem 1 concludes, with probability at least 1-delta-zeta, P[S notin C^(L)(X)] <= alpha + sum_{l,m} [positive terms]. Because all added terms are strictly positive for finite N, this does not imply the event in Eq. (9). An upper bound of alpha + c with c>0 does not certify P[failure] < alpha. The appendix's Proposition 2 ends at P[failure] <= sum epsilon_{l,m}, and the subsequent Bernstein/Hoeffding steps make each epsilon_{l,m} an empirical estimate plus positive slack; alpha enters only through the empirical coverage constraint on D2. Thus the advertised finite-sample coverage guarantee is not supported by the stated theorem.
  2. [Section 4.5 and Algorithm 1] The proof treats the thresholds Q_l(m;beta) as fixed, but in the implemented algorithm beta is selected on the same calibration split D2 used to compute the empirical quantities in Theorem 1. Algorithm 1 explicitly adjusts beta until the coverage constraint (Eq. 7) is satisfied on D2, and the same D2 supplies \hat p_{l,m}, \hat epsilon_{l,m}, and \hat v_{l,m}. No uniform bound over the class of beta configurations is provided, and no independent hold-out set is used after optimization. The empirical Bernstein and Hoeffding inequalities in Appendix A.2 therefore do not directly apply to the algorithm's output. A corrected proof would require a separate calibration split after beta selection or a uniform generalization bound accounting for the optimization.
  3. [Section 4.5, Theorem 1 assumption] The theorem assumes that the estimated clustering map \hat h_l, for all l, 'successfully partitions' class-step pairs into distinct clusters, but this condition is never formally defined. The oracle clustering h*_l with tolerance delta is introduced in Section 4.1, yet no bound on the discrepancy between h*_l and \hat h_l is given, nor any verifiable condition that would make the decomposition into the disjoint events E_{l,m} valid for the estimated clusters. If the learned clustering is poor, the disjoint partition and the subsequent bound do not go through. This is a load-bearing gap rather than a cosmetic one.
  4. [Section 4.5, Remark 4] The paper claims a coverage rate independent of the next-token prediction length L, but the bound in Theorem 1 is a sum over l in [L] of positive terms. With finite samples each term is positive, so the right-hand side grows with L for fixed cluster count M; only asymptotically, and only if the sum of the concentration terms vanishes uniformly, could the bound be described as L-independent. As stated, the theorem gives no finite-sample guarantee that is uniform in L.
  5. [General (no experiments)] The manuscript presents no experimental results or code. The paper motivates CoVeR as a practical decoding strategy and claims to maintain a compact search space and support long-tail sequences, but none of these claims are demonstrated empirically. Even if the theory were repaired, the algorithmic claims would require evaluation on autoregressive language-model benchmarks to be substantiated.
minor comments (4)
  1. [Abstract] Typo: 'wihtin' should be 'within'.
  2. [Related Work] Typo: 'comformal' should be 'conformal'.
  3. [Eq. (1)] The quantile level in Eq. (1) appears inconsistent with the definition of Quantile: for a set C(X) = {y: score >= q_hat}, the standard conformal level is (N+1)(1-alpha)/N, not (N+1)alpha/N. Please check the intended definition of Quantile and the direction of the inequality.
  4. [Notation] Beta is denoted beta_{l,m} in some places and beta_{m,l} in Algorithm 1; the ordering is inconsistent. Also, the set T in Eq. (4) is written with set-builder notation unclear about whether the condition applies per-level.

Circularity Check

0 steps flagged

No circular derivation; the PAC guarantee is logically under-derived but not circularly defined.

full rationale

The paper's central derivation is not circular. The optimization in Eq. (7) directly enforces the empirical full-path coverage constraint 1/|I2| sum L >= 1-alpha, and Theorem 1 then returns alpha as the leading term in an upper bound on true noncoverage together with concentration slack terms. This is a standard generalization-bound structure: the empirical constraint is an input, not a disguised version of the conclusion. The positive slack terms are genuine finite-sample corrections; the bound is not the same as the constraint. Where the paper fails is logical support, not circularity: Eq. (9) defines PAC as P_{Zval}[P(failure) < alpha] > 1-delta, but Theorem 1 only establishes P(failure) <= alpha + c with c > 0, and no argument shows c <= 0. Likewise, the proof applies empirical Bernstein to classifiers whose thresholds beta were selected on the same data, and the 'successfully partition' assumption on h_hat_l is never formalized; both are correctness gaps, not circular steps. No load-bearing self-citations were found: the cited learning-theoretic tools (Park et al. 2020, empirical Bernstein) are external. Thus no circular step can be exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

Everything central rests on the cluster-step decomposition, the assumption that fitted clusters match the oracle, and the in-sample optimization of beta. These are not externally validated, and the proof does not account for their uncertainty.

free parameters (4)
  • beta quantile levels beta_{m,l} = optimized on D2 via Algorithm 1
    Each cluster-step threshold is a learned quantile level; the full-path coverage constraint is enforced on calibration data, and the same values enter Theorem 1's bound.
  • cluster count M = not specified
    Number of clusters per step is chosen by hand; no selection criterion or sensitivity analysis is given.
  • regularization weights lambda_{l,m} = not specified
    Weights controlling the local non-coverage trade-off are set by hand; no recipe is provided.
  • clustering parameters T, gamma, bucket width = examples only
    Quantile levels T, split ratio gamma, and step-bucket width are user-specified; T=(0.5,...,0.9) is an example, but no default or tuning guidance is given.
axioms (5)
  • domain assumption Exchangeability / i.i.d. of calibration and test samples
    Standard conformal prediction assumption invoked throughout Section 3 and Theorem 1; not verified.
  • domain assumption Autoregressive cascading: if a prefix is excluded, all continuations are excluded
    Used in Proposition 2's disjoint decomposition; true for the defined set recursion but restricts the class of decoders.
  • ad hoc to paper Clustering h_hat successfully partitions class-step pairs
    Stated as an assumption in Theorem 1 with no formal success condition, no metric, and no bound on clustering error.
  • ad hoc to paper Greedy optimization converges to a beta satisfying the empirical coverage constraint
    Algorithm 1's while loops and random trade-offs are not proven to converge to a feasible or near-optimal solution; Theorem 1 assumes the resulting beta.
  • standard math Empirical Bernstein and Hoeffding inequalities
    Used in Appendix A.2 to bound local errors and cluster probabilities; standard concentration results.
invented entities (1)
  • oracle clustering map h*_l with tolerance delta no independent evidence
    purpose: Defines the formal local partition that the estimated h_hat is assumed to approximate
    The oracle is never estimated and no rate for the gap between h* and h_hat is provided; it appears in the event definition in Section 4.1 but not in the implementable algorithm.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CoVeR: Conformal Calibration for Versatile and Reliable Autoregressive Next-Token Prediction." pith.science (2026). https://pith.science/paper/UOG3K2OU

@misc{pith2026250904733,
  author       = {Pith},
  title        = {Pith review of: CoVeR: Conformal Calibration for Versatile and Reliable Autoregressive Next-Token Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOG3K2OU}},
  note         = {Machine review of arXiv:2509.04733}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Autoregressive pre-trained models combined with decoding methods have achieved impressive performance on complex reasoning tasks. While mainstream decoding strategies such as beam search can generate plausible candidate sets, they often lack provable coverage guarantees, and struggle to effectively balance search efficiency with the need for versatile trajectories, particularly those involving long-tail sequences that are essential in certain real-world applications. To address these limitations, we propose \textsc{CoVeR}, a novel model-free decoding strategy wihtin the conformal prediction framework that simultaneously maintains a compact search space and ensures high coverage probability over desirable trajectories. Theoretically, we establish a PAC-style generalization bound, guaranteeing that \textsc{CoVeR} asymptotically achieves a coverage rate of at least $1 - \alpha$ for any target level $\alpha \in (0,1)$.

Figures

Figures reproduced from arXiv: 2509.04733 by Dacheng Tao, Shunyu Liu, Yingjie Wang, Yongcheng Jing, Yuzhu Chen.

Figure 1
Figure 1. Figure 1: Construction process of COVER. First, we compute the conformal scores of all length-l calibration sequences for constructing quantile embeddings in Eq. (4), where the colored segments represent the labels appearing in the sequence at step l. Notably, some labels, such as the orange ones, belong to long-tail categories. Second, we treat the conformal scores of all length-l sequences that share a same token … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    Learn then test: Calibrating predictive algorithms to achieve risk control

    Anastasios N Angelopoulos, Stephen Bates, Emmanuel J Cand \`e s, Michael I Jordan, and Lihua Lei. Learn then test: Calibrating predictive algorithms to achieve risk control. arXiv preprint arXiv:2110.01052, 2021

    Pith/arXiv arXiv 2021

  2. [2]

    Uncommonsense: Informative negative knowledge about everyday concepts

    Hiba Arnaout, Simon Razniewski, Gerhard Weikum, and Jeff Z Pan. Uncommonsense: Informative negative knowledge about everyday concepts. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 37--46, 2022

    work page 2022

  3. [3]

    Conformal prediction for reliable machine learning: theory, adaptations and applications

    Vineeth Balasubramanian, Shen-Shyang Ho, and Vladimir Vovk. Conformal prediction for reliable machine learning: theory, adaptations and applications. Newnes, 2014

    work page 2014

  4. [4]

    Efficient beam search for large language models using trie-based decoding

    Brian J Chan, Jui-Hung Cheng, Mao Xun Huang, Chao-Ting Chen, and Hen-Hsen Huang. Efficient beam search for large language models using trie-based decoding. arXiv preprint arXiv:2502.00085, 2025

    arXiv 2025

  5. [5]

    Say what you mean! large language models speak too positively about negative commonsense knowledge

    Jiangjie Chen, Wei Shi, Ziquan Fu, Sijie Cheng, Lei Li, and Yanghua Xiao. Say what you mean! large language models speak too positively about negative commonsense knowledge. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 9890--9908. Association for Computational Linguistics, 2023

    work page 2023

  6. [6]

    Relating neural text degeneration to exposure bias

    Ting-Rui Chiang and Yun-Nung Chen. Relating neural text degeneration to exposure bias. In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 228--239. Association for Computational Linguistics, 2021

    work page 2021

  7. [7]

    Conformal autoregressive generation: Beam search with coverage guarantees

    Nicolas Deutschmann, Marvin Alberts, and Mar \' a Rodr \' guez Mart \' nez. Conformal autoregressive generation: Beam search with coverage guarantees. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024

    work page 2024

  8. [8]

    Class-conditional conformal prediction with many classes

    Tiffany Ding, Anastasios Angelopoulos, Stephen Bates, Michael Jordan, and Ryan J Tibshirani. Class-conditional conformal prediction with many classes. In Advances in neural information processing systems, pages 64555--64576, 2023

    work page 2023

  9. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Large language models can self-improve

    Jiaxin Huang, Shixiang Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. Large language models can self-improve. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 1051--1068, 2023

    work page 2023

  11. [11]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Large language models pass the turing test

    Cameron R Jones and Benjamin K Bergen. Large language models pass the turing test. arXiv preprint arXiv:2503.23674, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    In search of the long-tail: Systematic generation of long-tail knowledge via logical rule induced search, 2024

    Huihan Li, Zeyi Liao, Yuting Ning, Siyuan Wang, Xiang Lorraine Li, Ximing Lu, Faeze Brahman, Wenting Zhao, Yejin Choi, and Xiang Ren. In search of the long-tail: Systematic generation of long-tail knowledge via logical rule induced search, 2024

    work page 2024

  14. [14]

    Negative expertise

    Marvin Minsky. Negative expertise. 1997

    work page 1997

  15. [15]

    Openai o3 system card

    OpenAI. Openai o3 system card. Technical Report, 2025

    work page 2025

  16. [16]

    Pac confidence sets for deep neural networks via calibrated prediction

    Sangdon Park, Osbert Bastani, Nikolai Matni, and Insup Lee. Pac confidence sets for deep neural networks via calibrated prediction. In Proceedings of the International Conference on Learning Representations, 2020

    work page 2020

  17. [17]

    Probabilistic conformal prediction with approximate conditional validity

    Vincent Plassier, Alexander Fishkov, Mohsen Guizani, Maxim Panov, and Eric Moulines. Probabilistic conformal prediction with approximate conditional validity. In Proceedings of the International Conference on Learning Representations, 2025

    work page 2025

  18. [18]

    Jaakkola, and Regina Barzilay

    Victor Quach, Adam Fisch, Tal Schuster, Adam Yala, Jae Ho Sohn, Tommi S. Jaakkola, and Regina Barzilay. Conformal language modeling. In Proceedings of the International Conference on Learning Representations, 2024

    work page 2024

  19. [19]

    Conformal nucleus sampling

    Shauli Ravfogel, Yoav Goldberg, and Jacob Goldberger. Conformal nucleus sampling. In Findings of the Association for Computational Linguistics, pages 27--34, 2023

    work page 2023

  20. [20]

    Allen Z. Ren, Anushri Dixit, Alexandra Bodrova, Sumeet Singh, Stephen Tu, Noah Brown, Peng Xu, Leila Takayama, Fei Xia, Jake Varley, Zhenjia Xu, Dorsa Sadigh, Andy Zeng, and Anirudha Majumdar. Robots that ask for help: Uncertainty alignment for large language model planners. In 7th Annual Conference on Robot Learning, 2023

    work page 2023

  21. [21]

    A tutorial on conformal prediction

    Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction. Journal of Machine Learning Research, 9 0 (3), 2008

    2008

  22. [22]

    Improved Beam Search for Hallucination Mitigation in Abstractive Summarization

    Arvind Krishna Sridhar and Erik Visser. Improved beam search for hallucination mitigation in abstractive summarization. arXiv preprint arXiv:2212.02712, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [23]

    A survey of reasoning with foundation models

    Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, et al. A survey of reasoning with foundation models. arXiv preprint arXiv:2312.11562, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Less likely brainstorming: Using language models to generate alternative hypotheses

    Liyan Tang, Yifan Peng, Yanshan Wang, Ying Ding, Greg Durrett, and Justin F Rousseau. Less likely brainstorming: Using language models to generate alternative hypotheses. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, page 12532, 2023

    work page 2023

  25. [25]

    Conformal prediction under covariate shift

    Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal prediction under covariate shift. In Advances in neural information processing systems, 2019

    work page 2019

  26. [26]

    Self-instruct: Aligning language models with self-generated instructions

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022

    Pith/arXiv arXiv 2022

  27. [27]

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

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, 2022

    work page 2022

  28. [28]

    Self-evaluation guided beam search for reasoning

    Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. Self-evaluation guided beam search for reasoning. In Advances in Neural Information Processing Systems, pages 41618--41650, 2023

    work page 2023

  29. [29]

    Language-informed beam search decoding for multilingual machine translation

    Yilin Yang, Stefan Lee, and Prasad Tadepalli. Language-informed beam search decoding for multilingual machine translation. In Findings of the Association for Computational Linguistics, pages 15761--15772, 2024

    work page 2024

  30. [30]

    Calibrating sequence likelihood improves conditional language generation

    Yao Zhao, Mikhail Khalman, Rishabh Joshi, Shashi Narayan, Mohammad Saleh, and Peter J Liu. Calibrating sequence likelihood improves conditional language generation. In Proceedings of the International Conference on Learning Representations, 2023

    work page 2023

  31. [31]

    RICA : Evaluating robust inference capabilities based on commonsense axioms

    Pei Zhou, Rahul Khanna, Seyeon Lee, Bill Yuchen Lin, Daniel Ho, Jay Pujara, and Xiang Ren. RICA : Evaluating robust inference capabilities based on commonsense axioms. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7560--7579, 2021

    work page 2021

  32. [32]

    Deductive beam search: Decoding deducible rationale for chain-of-thought reasoning

    Tinghui Zhu, Kai Zhang, Jian Xie, and Yu Su. Deductive beam search: Decoding deducible rationale for chain-of-thought reasoning. In First Conference on Language Modeling, 2024

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.