betaFeedback

WorksComputer Science

Understanding the Out-of-Distribution Generalization of Chain-of-Thought Reasoning in LLMs

Abstract

Chain-of-Thought (CoT) reasoning substantially improves Large Language Models (LLMs), yet its out-of-distribution (OOD) generalization mechanism remains theoretically underexplored. We study CoT as compositional OOD generalization: the target task lacks complete reasoning trajectories or direct input-output pairs, while training provides only in-distribution (ID) atomic step data. We show that CoT generalizes by repeatedly applying a shared single-step predictor to compose reusable ID transitions. Theoretically, let TT denote the number of reasoning steps in the target task, nn the number of training samples per sub-task, and kk the number of output tokens per step; then the OOD task risk is controlled by the sum of the ID sub-task risks, scaled by a coefficient of order O(k)\mathcal{O}(k) that is independent of TT, although the summed risk itself may still grow with TT. A Rademacher-complexity analysis further shows that the estimation term is of order O(T/n)\mathcal{O}(\sqrt{T/n}), rather than linear in TT, due to predictor sharing under a uniform capacity control on the shared class. For autoregressive Next-Token Prediction, the final-task risk is bounded by the best-in-class step-level risk scaled by the same coefficient of order O(k)\mathcal{O}(k), plus an estimation term of order O~(kT/n)\widetilde{\mathcal{O}}\big(k\sqrt{T/n}\big). A parity-coupled construction further shows that, for the sub-task empirical risk minimizer, the expected final-task risk is at least the best in-class risk plus Ω(T/n)Ω(\sqrt{T/n}) when T≤nT\le n. Experiments on Atom-Task and GSM8K support the theory: CoT enables short-to-long and compositional generalization, while explicit state tracking is crucial. Our code is available at https://github.com/maxliang114514/CoT-OOD.

Introduction

Chain-of-Thought (CoT) reasoning has become a central mechanism for improving the reasoning ability of Large Language Models (LLMs). By prompting or training models to generate intermediate steps, CoT substantially improves mathematical and logical reasoning [1, 2, 3, 4, 5, 6, 7, 8]. Recent reasoning-oriented systems further treat step generation as an explicit training signal [9, 10, 11], shifting LLM reasoning from direct pattern matching toward structured stepwise computation [12]. Despite this empirical success, a fundamental theoretical question remains open: why can CoT-trained models generalize to complex reasoning tasks that are not directly observed during training?

Existing theoretical studies provide important insights into the expressivity and computational role of CoT. Some works show that CoT can increase the effective depth of Transformers or enable serial computation [13, 14, 15], while others analyze task decomposition, in-context learning, or localized reasoning settings [16, 17, 18]. However, most existing analyses are either qualitative, restricted to narrow or two-step settings, or do not provide explicit generalization bounds for the autoregressive Next-Token Prediction (NTP) regime commonly used by modern large-scale LLMs. As a result,

Preprint. the out-of-distribution (OOD) generalization mechanism of multi-step CoT remains theoretically underexplored.

This paper studies the mechanism of CoT through compositional OOD generalization. We consider a target task requiring a complex multi-step mapping from an input state z0z_0 to a final answer ΔzT\Delta z_T, while neither complete reasoning trajectories (z0,Δz1,…,ΔzT−1,ΔzT)(z_0, \Delta z_1, \ldots, \Delta z_{T-1}, \Delta z_T) nor direct input-output pairs (z0,ΔzT)(z_0, \Delta z_T) are available during training. Thus, the target task’s composition is OOD relative to the observed atomic transitions. We refer to this setting as compositional OOD; § 3.2 separately treats mismatched training and target step distributions (Corollary 4). Instead, the model has access to in-distribution (ID) atomic step data: samples (zi−1,Δzi)(z_{i-1}, \Delta z_i), where zi−1z_{i-1} is the reasoning state before step ii and Δzi\Delta z_i is the next incremental output. Here, TT denotes the total number of generation steps: the first T−1T-1 are intermediate reasoning updates, and ΔzT\Delta z_T is the final answer.

Under this view, CoT does not require the model to learn the entire OOD mapping at once from scratch. Instead, it repeatedly applies a shared single-step predictor. At each step, the model reads the current state zi−1z_{i-1}, generates Δzi\Delta z_i containing kk new tokens, and updates the state by appending these tokens while keeping a fixed context length mm. CoT therefore converts a complex target mapping into a composition of reusable ID atomic transitions (Figure 1). This motivates our central question:

Figure 1. CoT as compositional OOD generalization. Left: without CoT, the model must learn a direct mapping from input z0z_0 to final answer ΔzT\Delta z_T, whose target distribution may be unavailable during training. With CoT, the same predictor is repeatedly applied to in-distribution step distributions, mapping each state zi−1z_{i-1} to update Δzi\Delta z_i. Right: a two-hop translation example solves an unseen English-to-German query by composing learned English-to-French and French-to-German steps.

When a model learns only atomic step transitions, can repeated CoT composition generalize to unseen multi-step OOD tasks, and how does the resulting error scale with the reasoning depth TT?

Contributions We answer this question by developing a generalization framework for CoT-based task decomposition. First, we prove that the risk of the complex OOD task can be upper-bounded by the sum of risks over ID sub-tasks, scaled by a coefficient of order O(k)O(k) that is independent of TT. The coefficient results from deterministic argmax decoding together with the bounded sequence loss and the local error floor (log⁡2)/k(\log2)/k that a hard-decoding (argmax) error induces. CoT thus avoids fitting the unseen composite mapping directly and composes local transitions at test time.

Second, we derive a Rademacher-complexity bound for the multi-step setting. A naive analysis that treats each reasoning step independently would yield a linear dependence on TT. In contrast, because CoT applies the same predictor across all steps, we can analyze the combined step-level sample set and obtain a sublinear generalization-gap term scaling as O(T/n)O(\sqrt{T/n}), where nn is the number of training samples per atomic step distribution.

Third, we specialize the analysis to the NTP setting of autoregressive LLMs. We directly analyze the final kk target tokens of each atomic sequence and take the N=nTN=nT complete sequences as the sampling units. The resulting final-task risk bound keeps the best in-class step-level risk explicit and adds an estimation term of order O~(kT/n)\widetilde{O}(k\sqrt{T/n}). Appendix B.10 shows the T/n\sqrt{T/n} dependence cannot be improved for the sub-task empirical risk minimizer: a construction with k=1k=1, in which the per-step errors are coupled through the XOR accumulated over the chain, makes this minimizer incur an excess of Ω(T/n)\Omega(\sqrt{T/n}) when T≤nT\le n, and carries that excess through to the final task.

We further validate our theory through controlled experiments. We introduce Atom-Task, a symbolic compositional benchmark involving multi-step sequence operations, to isolate the effect of CoT on OOD compositional generalization. Experiments show that CoT-trained models generalize from shorter or constrained training chains to longer unseen compositions, whereas models trained without intermediate reasoning fail to capture the underlying compositional structure. Ablation results further show that explicitly tracking the current reasoning state is crucial, matching our theoretical requirement that each state zi−1z_{i-1} contains sufficient information for the next atomic transition. Additional experiments on GSM8K, a benchmark of grade-school math word problems, support the same conclusion in real-world mathematical reasoning.

Preliminaries

Problem Setting We consider an autoregressive predictor with a context window size of mm. Let V\mathcal{V} be a vocabulary (a set of tokens), with nv=∣V∣n_v = |\mathcal{V}| denoting its size, and let Vm\mathcal{V}^m denote the set of all token sequences of length mm formed by tokens from V\mathcal{V}. The predictor ψ:Vm→Rnv×k\psi: \mathcal{V}^m \to\mathbb{R}^{n_v \times k} (m≫km \gg k) maps an input sequence to probability distributions over kk output positions, and belongs to a hypothesis class Ψ\Psi. Losses are computed on these soft predictions; for autoregressive generation, discrete tokens are obtained via argmax, denoted hard⁡(ψ(⋅))∈Vk\operatorname{hard}(\psi(\cdot)) \in\mathcal{V}^k.

Under the CoT paradigm, the model does not predict the final answer directly via ψ(z0)\psi(z_0), but generates a chain of TT reasoning steps. We write ⋅^\hat{\cdot} for model-generated quantities and reserve the unhatted symbol for the ground truth. At step ii, the incremental output is Δz^i=hard⁡(ψ(z^i−1))\Delta\hat{z}_i = \operatorname{hard}(\psi(\hat{z}_{i-1})), and the state updates by appending these kk tokens while discarding the first kk (sliding window):

z^i=z^i−1[k+1:m]⊕Δz^i(⊕ denotes sequence concatenation),Δz^T=hard⁡(ψ(z^T−1)),\hat{z}_i = \hat{z}_{i-1}[k+1:m] \mathbin{\oplus} \Delta\hat{z}_i \qquad(\mathbin{\oplus}\text{ denotes sequence concatenation}), \qquad\Delta\hat{z}_T = \operatorname{hard}(\psi(\hat{z}_{T-1})),

thus maintaining the sequence length mm for all intermediate states, with z^0=z0\hat{z}_0 = z_0; the true state follows the same recursion with Δzi\Delta z_i in place of Δz^i\Delta\hat{z}_i, i.e., zi=zi−1[k+1:m]⊕Δziz_i = z_{i-1}[k+1:m] \mathbin{\oplus} \Delta z_i.

For a given complex task, any instance (z0,ΔzT)(z_0, \Delta z_T) with z0∈Vmz_0 \in\mathcal{V}^m and ΔzT∈Vk\Delta z_T \in\mathcal{V}^k satisfies (z0,ΔzT)∼D(z_0, \Delta z_T) \sim\mathcal{D}. There exists a ground-truth mapping ψ∗\psi_* (not necessarily within Ψ\Psi) such that hard⁡(ψ∗(zi−1))=Δzi\operatorname{hard}(\psi_*(z_{i-1})) = \Delta z_i for all ii, implying ψ∗\psi_* generates the CoT sequence (Δz1,…,ΔzT)(\Delta z_1, \ldots, \Delta z_T) from z0z_0 to yield the target output ΔzT\Delta z_T.

Define the state transition ψˉ:Vm→Vm\bar{\psi} : \mathcal{V}^m \to\mathcal{V}^m as ψˉ(z):=z[k+1:m]⊕hard⁡(ψ(z))\bar{\psi}(z) := z[k+1:m] \mathbin{\oplus} \operatorname{hard}(\psi(z)), representing the input state for the next step (note ψˉ\bar{\psi} uses hard predictions for transitions). To distinguish from ψ\psi, let Fψ:Vm→Rnv×kF_{\psi} : \mathcal{V}^m \to\mathbb{R}^{n_v \times k} denote the composite CoT mapping Fψ(z0):=ψ(ψˉ(T−1)(z0))F_{\psi}(z_0) := \psi(\bar{\psi}^{(T-1)}(z_0)), which encapsulates the TT-step reasoning process (note FψF_{\psi} applies ψ\psi to the final state, outputting soft predictions for loss computation). Our learning goal is to minimize the population risk:

RD(Fψ)=∫L(Fψ(z0),ΔzT) dP(z0,ΔzT)=∫L(ψ(ψˉ(T−1)(z0)),ΔzT) dP(z0,ΔzT),R_{\mathcal{D}}(F_{\psi}) = \int\mathcal{L}(F_{\psi}(z_0), \Delta z_T)\,dP(z_0,\Delta z_T) = \int\mathcal{L}(\psi(\bar{\psi}^{(T-1)}(z_0)), \Delta z_T)\,dP(z_0,\Delta z_T),

where PP is the probability measure associated with D\mathcal{D} and L:Rnv×k×Rnv×k→R\mathcal{L} : \mathbb{R}^{n_v \times k} \times\mathbb{R}^{n_v \times k} \to\mathbb{R} is the sequence-level loss function that compares the model’s soft predictions (in Rnv×k\mathbb{R}^{n_v \times k}) with the ground-truth output ΔzT\Delta z_T (converted from discrete tokens to one-hot vectors in Rnv×k\mathbb{R}^{n_v \times k}); the formal definition is provided in Section 3. This aims to identify a ψ∈Ψ\psi\in\Psi approximating ψ∗\psi_*, thereby generating the CoT sequence from z0z_0 to yield ΔzT\Delta z_T.

Notations ψˉ(T)\bar{\psi}^{(T)} denotes the TT-fold composition ψˉ∘⋯∘ψˉ\bar{\psi} \circ\cdots\circ\bar{\psi} (TT times in total). The final prediction is extracted via ψ(ψˉ(T−1)(z0))\psi(\bar{\psi}^{(T-1)}(z_0)). For a sequence zz, z[a:b]z[a:b] denotes the subsequence from index aa to bb (inclusive), and z[i]z[i] denotes its ii-th component (or token). The operator ⊕\oplus denotes the concatenation of two sequences. Other key notations are summarized in Table 4 in Appendix B.

References

  1. [1]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022.arxiv.org/abs/2201.11903
  2. [2]Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MathBench: Building math generational models through hybrid instruction tuning. In The Twelfth International Conference on Learning Representations, 2024.arxiv.org/abs/2309.05653
  3. [3]Longhui Yu, Weisen Jiang, Han Shi, Jincheng YU, Zhengying Liu, Yu Zhang, James Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. In The Twelfth International Conference on Learning Representations, 2024.arxiv.org/abs/2309.12284
  4. [4]Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9426–9439, 2024.
  5. [5]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.arxiv.org/abs/2402.03300
  6. [6]Fangxu Yu. Flow of reasoning: Efficient training of llm policy with divergent thinking. arXiv (Cornell University), 2024.
  7. [7]Namgyu Ho, Laura Schmid, and Se-Young Yun. Large language models are reasoning teachers. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pages 14852–14882, 2023.
  8. [8]Seungone Kim, Se Joo, Doyoung Kim, Joel Jang, Seonghyeon Ye, Jamin Shin, and Minjoon Seo. The cot collection: Improving zero-shot and few-shot learning of language models via chain-of-thought fine-tuning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12685–12708, 2023.arxiv.org/abs/2305.14045
  9. [9]Team OpenAI. Learning to reason with llms. OpenAI Blog, 2024.
  10. [10]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shiring Ma, Xiao Bi, et al. Deepseek-rl: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.
  11. [11]Kanishk Gandhi, Ayush K Chakravarthy, Anikait Singh, Nathan Lile, and Noah Goodman. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. In Second Conference on Language Modeling, 2025.arxiv.org/abs/2503.01307
  12. [12]Akshara Prabhakar, Thomas L Griffiths, and R Thomas McCoy. Deciphering the factors influencing the efficacy of chain-of-thought: Probability, memorization, and noisy reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 3710–3724, 2024.
  13. [13]Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36:70757–70798, 2023.arxiv.org/abs/2305.15408
  14. [14]William Merrill and Ashish Sabharwal. The expressive power of transformers with chain of thought. In The Twelfth International Conference on Learning Representations, 2024.arxiv.org/abs/2310.07923
  15. [15]Zhiyuan Li, Hong Liu, Denny Zhou, and Tengyu Ma. Chain of thought empowers transformers to solve inherently serial problems. In The Twelfth International Conference on Learning Representations, 2024.arxiv.org/abs/2402.12875
  16. [16]Chenxiao Yang, Zhiyuan Li, and David Wipf. Chain-of-thought provably enables learning the (otherwise) unlearnable. In The Thirteenth International Conference on Learning Representations, 2025.
  17. [17]Alireza Amiri Bavandpour, Xinting Huang, Mark Rofin, and Michael Hahn. Lower bounds for chain-of-thought reasoning in hard-attention transformers. In International Conference on Machine Learning, pages 3274–3306. PMLR, 2025.
  18. [18]Ben Prystawski, Michael Li, and Noah Goodman. Why think step by step? reasoning emerges from the locality of experience. Advances in Neural Information Processing Systems, 36:70926–70947, 2023.arxiv.org/abs/2304.03843
  19. [19]Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms. Cambridge university press, 2014.
  20. [20]Liyuan Liu, Biqin Song, Zhibin Pan, Chuanwu Yang, Chi Xiao, and Weifu Li. Gradient learning under tilted empirical risk minimization. Entropy, 24(7):956, 2022.DOI
  21. [21]Yuyang Deng, Junyuan Hong, Jiayu Zhou, and Mehrdad Mahdavi. On the generalization ability of unsupervised pretraining. In International Conference on Artificial Intelligence and Statistics, pages 4519–4527. PMLR, 2024.arxiv.org/abs/2403.06871
  22. [22]Liyuan Liu, Hong Chen, Chi Xiao, and Weifu Li. The consistency analysis of gradient learning under independent covariate shift. Neurocomputing, 635:129883, 2025.DOI
  23. [23]Zhihao Li, Xue Jiang, Liyuan Liu, Xuelin Zhang, Hong Chen, and Feng Zheng. On the generalization ability of next-token-prediction pretraining. In Forty-second International Conference on Machine Learning, 2025.
  24. [24]Andreas Maurer, Massimiliano Pontil, and Bernardino Romera-Paredes. The benefit of multitask representation learning. Journal of Machine Learning Research, 17(81):1–32, 2016.arxiv.org/abs/1505.06279
  25. [25]Xinhao Yao, Ruifeng Ren, Yun Liao, Lizhong Ding, and Yong Liu. Compositional generalization from learned skills via cot training: A theoretical and structural analysis for reasoning. In The Fourteenth International Conference on Learning Representations, 2026.arxiv.org/abs/2502.04667
  26. [26]Amirhesam Abedsoltan, Huaqing Zhang, Kaiyue Wen, Hongzhou Lin, Jingzhao Zhang, and Mikhail Belkin. Task generalization with autoregressive compositional structure: Can learning from d tasks generalize to d^T tasks? arXiv preprint arXiv:2502.08991, 2025.arxiv.org/abs/2502.08991
  27. [27]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021.arxiv.org/abs/2110.14168
  28. [28]Shunyu Yao, Dian Yu, Jeffrey Zhao, Izak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822, 2023.arxiv.org/abs/2305.10601
  29. [29]Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawksi, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 17682–17690, 2024.DOI
  30. [30]Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research, 2023.arxiv.org/abs/2211.12588
  31. [31]Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2203.11171
  32. [32]Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems, 36:46534–46594, 2023.arxiv.org/abs/2303.17651
  33. [33]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459–9474, 2020.arxiv.org/abs/2005.11401
  34. [34]Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, et al. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2205.10625
  35. [35]Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2210.03493
  36. [36]Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476–15488, 2022.arxiv.org/abs/2203.14465
  37. [37]Nirmit Joshi, Gal Vardi, Adam Block, Surbhi Goel, Zhiyuan Li, Theodor Misiakiewicz, and Nathan Srebro. A theory of learning with autoregressive chain of thought. In The Thirty Eighth Annual Conference on Learning Theory, pages 3161–3212. PMLR, 2025.arxiv.org/abs/2503.07932
  38. [38]Jianhao Huang, Zixuan Wang, and Jason D Lee. Transformers learn to implement multi-step gradient descent with chain of thought. In The Thirteenth International Conference on Learning Representations, 2025.arxiv.org/abs/2502.21212
  39. [39]Boshi Wang, Xiang Yue, Yu Su, and Huan Sun. Grokking of implicit reasoning in transformers: A mechanistic journey to the edge of generalization. Advances in Neural Information Processing Systems, 37:95238–95265, 2024.DOI
  40. [40]Eden Biran, Daniela Gottesman, Sohee Yang, Mor Geva, and Amir Globerson. Hopping too late: Exploring the limitations of large language models on multi-hop queries. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14113–14130, 2024.
  41. [41]Steve Hanneke, Idan Mehalel, and Shay Moran. Sample complexity of autoregressive reasoning: Chain-of-thought vs. end-to-end. arXiv preprint arXiv:2604.12013, 2026.arxiv.org/abs/2604.12013
  42. [42]Gokul Swamy, Sanjiban Choudhury, Wen Sun, Zhiwei Steven Wu, and J Andrew Bagnell. All roads lead to likelihood: The value of reinforcement learning in fine-tuning. arXiv preprint arXiv:2503.01067, 2025.arxiv.org/abs/2503.01067
  43. [43]Amrith Setlur, Nived Rajaraman, Sergey Levine, and Aviral Kumar. Scaling test-time compute without verification or rl is suboptimal. In International Conference on Machine Learning, pages 54058–54094. PMLR, 2025.arxiv.org/abs/2502.12118
  44. [44]Yang Yue, Zhiqi Chen, Rui Li, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025.arxiv.org/abs/2504.13837
  45. [45]Darsh J Shah, Peter Rushton, Somanshu Singla, Mohit Parmar, Kurt Smith, Yash Vanjani, Ashish Vaswani, Adarsh Chaluv ar aju, Andrew Hojel, Andrew Ma, et al. Rethinking reflection in pre-training. arXiv preprint arXiv:2504.04022, 2025.arxiv.org/abs/2504.04022

Paper details

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

views · 0

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents