betaFeedback

WorksComputer Science

How Does Chain-of-Thought Supervision Teach Models to Reason? A Theoretical Analysis

Abstract

Training with chain-of-thought (CoT) supervision is an effective approach for teaching models multi-step reasoning procedures. At inference time, however, models may need to continue reasoning beyond the finite teacher CoT sequences observed during training, and longer reasoning can instead degrade performance, a phenomenon known as overthinking. How the amount and structure of training-time CoT give rise to such failures remains theoretically poorly understood. To study this problem, we introduce an in-context weight prediction model trained on finite gradient descent trajectories for linear regression. Using high-dimensional asymptotic analysis, we show that the success and failure modes of long-horizon reasoning undergo phase transitions depending on the context length and the total amount of training CoT. We further distinguish two mechanisms of overthinking, error accumulation and error amplification, and show that with short training contexts, increasing the training CoT budget can worsen long-horizon error before eventually stabilizing reasoning. Experiments with fully trained linear and multi-head softmax attention further show that these qualitative behaviors persist. Our results provide a theoretical account linking properties of training-time CoT to inference-time failures and offer guidance for designing teacher CoT that supports stable reasoning.

Introduction

Popularized as a test-time prompting technique, chain-of-thought (CoT) [30] is now widely used as training supervision to improve a model’s reasoning ability [33, 8]. Whereas final-answer supervision provides only outcome-level feedback, supervising intermediate steps can help models learn general reasoning procedures. Such CoT-based supervision has produced substantial gains on complex tasks such as mathematics and logical reasoning [14, 33, 20].

During inference, a model may need to carry the reasoning procedure learned from CoT supervision beyond the number of steps shown during training to reach the correct answer. In practice, extending reasoning can instead degrade performance, a phenomenon widely known as overthinking, so additional reasoning steps do not necessarily improve accuracy [5, 26, 13]. Recent empirical studies suggest that overthinking is a problem involving the correctness and stability of the reasoning procedure itself. For example, partial acquisition of reasoning capabilities may temporarily degrade long-horizon reasoning performance [31, 32, 24]. Moreover, systems with similar short-horizon performance may exhibit sharply different long-horizon failure modes: errors may simply accumulate across reasoning steps, or they may be amplified as the system becomes increasingly confident in and builds on its earlier mistakes [7, 3]. These findings show that locally plausible steps do not guarantee global convergence or error correction under repeated inference.

Despite these observations of long-horizon failures, it remains unclear how the amount and structure of training CoT shape the stability of reasoning. In large language models, training data and model scale also affect performance, making the contribution of CoT supervision difficult to isolate experimentally [6, 19]. Consequently, observed performance offers limited guidance on how to choose CoT training data that support reasoning.

To understand how training CoT affects reasoning theoretically, we introduce an in-context weight prediction model trained on gradient descent (GD) trajectories as teacher CoT. Specifically, we use GD trajectories for linear-regression tasks as teacher CoT and train a linear attention model on each transition. Previous work has shown that sufficient CoT supervision allows a model to approximately implement multi-step GD [16], but a theoretical understanding of long-horizon success conditions and failure regimes remains limited. More recently, phase transitions between successful inference and overthinking have been identified in a setting without training-time CoT (Takanami & Pehlevan, 2026). We extend this framework to finite teacher CoT trajectories and ask whether recursive application of the learned model can reproduce the correct update rule beyond the training horizon. Through this analysis, we characterize how the number of training tasks, context length, and teacher CoT length affect the inference dynamics.

Our contributions are as follows:

  • We derive a phase diagram in the asymptotic limit showing that context length and the total amount of training CoT divide long-horizon reasoning into four regimes: exact GD recovery, bounded error, exponential divergence, and quadratic divergence. (§ 4)

  • We show that reasoning can extend beyond the finite-length teacher CoT. The number of pretraining tasks required for this behavior decreases as the amount of teacher CoT increases, while shorter contexts require more teacher CoT. (§ 5.1)

  • We identify two distinct mechanisms of overthinking: error accumulation, where errors introduced at each reasoning step remain uncorrected, and error amplification, where unstable reasoning dynamics repeatedly magnify existing errors, leading to quadratic and exponential MSE growth, respectively. (§ 5.2)

  • We show that with short contexts, increasing the training CoT budget can initially degrade long-horizon performance before the model learns to reason correctly. (§ 5.3)

  • We empirically show that the same qualitative mechanisms persist in more complex architectures, including fully trained linear attention and multi-head softmax attention. (§ 6)

Our results provide a unified theoretical understanding of the role of finite teacher CoT, appropriate CoT design as training data, and the mechanisms that generate failure.

Related Work

CoT supervision and overthinking. CoT data are increasingly used as training signals to improve reasoning and distill it into smaller models [15, 20, 2, 11, 35, 23]. At the same time, longer reasoning or recurrent computation can reduce performance, a phenomenon often described as overthinking [5, 7, 26, 13, 3]. However, most existing work focuses on improving training methods or observing trained models, leaving unclear how the properties of training CoT determine test-time behavior. We study this relation in a solvable model and characterize when training CoT supports correct inference and when it makes the test-time reasoning unstable.

Theoretical studies of CoT. Theoretical studies of CoT have mainly focused on its inference-time benefits, showing that CoT can enable efficient reasoning on otherwise difficult tasks [1, 18, 16]. However, the conditions under which such learning fails remain poorly understood. To study this question theoretically, recent work extends analytically tractable models of in-context linear regression [9, 4, 29, 34, 22] to iterative CoT, investigating how limited pretraining data or context and task mismatch can lead to overthinking [17] (Takanami & Pehlevan 2026). However, these studies mainly analyze the test-time behavior of trained models and do not characterize the effect of providing finite teacher CoT as training data. We therefore focus on training-time CoT and ask how far a model can continue to reason correctly beyond the reasoning steps observed during training and how this generalization fails.

Model

We study an in-context weight-prediction model for linear regression. In this model, training CoT consists of GD trajectories, and the pre-trained model is recursively applied as test-time CoT to reproduce GD in context on new tasks.

Data generation

We consider MM linear-regression tasks indexed by μ=1,…,M\mu= 1,\ldots,M. For each task μ\mu, we prepare the training data and teacher CoT trajectories as follows.

Training tasks. For each of the MM tasks, we prepare LL labeled examples {xμ,l,yμ,l}l=1,…,L\{x_{\mu,l}, y_{\mu,l}\}_{l=1,\ldots,L}, given by

Xμ:=[xμ,1,…,xμ,L]∈RD×L,yμ:=XμTwμ+εμ∈RL,(1)X_\mu:= [x_{\mu,1},\ldots,x_{\mu,L}] \in\mathbb{R}^{D \times L}, \qquad y_\mu:= X_\mu^{\mathsf{T}}w_\mu+ \varepsilon_\mu\in\mathbb{R}^{L}, \tag*{(1)}

where xμ,l∼N(0D,ID)x_{\mu,l} \sim\mathcal{N}(0_D,I_D), wμ∼N(0D,ID/D)w_\mu\sim\mathcal{N}(0_D,I_D/D), and εμ∼N(0L,σ2IL)\varepsilon_\mu\sim\mathcal{N}(0_L,\sigma^2 I_L), independently across all tasks and examples. We denote the task context by Qμ:=(Xμ,yμ)Q_\mu:= (X_\mu,y_\mu).

CoT as training data. We prepare a KK-step trajectory of gradient descent (GD) {sμ,k}k=0,…,K\{s_{\mu,k}\}_{k=0,\ldots,K} with step size η>0\eta> 0 on the empirical squared loss. Starting from zero, its states satisfy

sμ,k+1:=sμ,k−ηLXμ(XμTsμ,k−yμ)(k=0,…,K−1),(2)s_{\mu,k+1} := s_{\mu,k} - \frac{\eta}{L}X_\mu\left(X_\mu^{\mathsf{T}}s_{\mu,k}-y_\mu\right) \qquad(k=0,\ldots,K-1), \tag*{(2)}

and sμ,0:=0Ds_{\mu,0} := 0_D.

Pre-training setting

During pre-training, the model is trained to predict the next step of the teacher CoT, sμ,k+1s_{\mu,k+1}, from the context and the teacher CoT at step kk, i.e., (Qμ,sμ,k)(Q_\mu,s_{\mu,k}). Specifically, the input data (prompt), architecture, and loss function are defined as follows.

Prompt construction. At training step kk, the prompt contains the original LL labeled examples and a slot for the current teacher state:

Eμ,k:=(Xμ0Dyμ00D×Lsμ,k0LT1)=(xμ,1⋯xμ,L0Dyμ,1⋯yμ,L00D⋯0Dsμ,k0⋯01)∈R(2D+2)×(L+1).(3)E_{\mu,k} := \begin{pmatrix} X_\mu& 0_D \\ y_\mu& 0 \\ 0_{D\times L} & s_{\mu,k} \\ 0_L^{\mathsf{T}} & 1 \end{pmatrix} = \begin{pmatrix} x_{\mu,1} & \cdots& x_{\mu,L} & 0_D \\ y_{\mu,1} & \cdots& y_{\mu,L} & 0 \\ 0_D & \cdots& 0_D & s_{\mu,k} \\ 0 & \cdots& 0 & 1 \end{pmatrix} \in\mathbb{R}^{(2D+2)\times(L+1)}. \tag*{(3)}

The first LL columns encode labeled examples. The final column is a state slot, and the marker 1 in the last row distinguishes it from an example token. The prompt for predicting sμ,k+1s_{\mu,k+1} contains the task context QμQ_\mu and the current state sμ,ks_{\mu,k}, but no earlier states. Thus, conditional on the task context, the one-step predictor is Markovian. Masking or downweighting earlier reasoning tokens in this way can improve length generalization [1, 21].

Linear attention. For a prompt with LL context examples, we use a one-layer linear self-attention model with a residual connection, defined by

Atten⁡(E;θ):=E+1L(WVE)(WKE)TWQE.(4)\operatorname{Atten}(E;\theta) := E + \frac{1}{L}(W_VE)(W_KE)^{\mathsf{T}}W_QE. \tag*{(4)}

Here the trainable parameters are θ=(WQ,WK,WV)\theta= (W_Q,W_K,W_V), where WQ,WK,WV∈R(2D+2)×(2D+2)W_Q,W_K,W_V \in\mathbb{R}^{(2D+2)\times(2D+2)} are the query, key, and value matrices, respectively. We read the state slot of the final column using

Readout⁡(E):=ED+2:2D+2,L+1,end.(5)\operatorname{Readout}(E) := E_{D+2:2D+2,L+1,\mathrm{end}}. \tag*{(5)}

The next-step predictor of the weight vector is therefore

s^μ,k+1:=Readout⁡(Atten⁡(Eμ,k;θ)).(6)\hat{s}_{\mu,k+1} := \operatorname{Readout}\left(\operatorname{Atten}(E_{\mu,k};\theta)\right). \tag*{(6)}

Learning. We consider the squared one-step prediction error over all teacher transitions in pretraining data:

L(θ):=1MK∑μ=1M∑k=0K−1∥s^μ,k+1−sμ,k+1∥22.(7)\mathcal{L}(\theta) := \frac{1}{MK}\sum_{\mu=1}^{M}\sum_{k=0}^{K-1}\left\|\hat{s}_{\mu,k+1}-s_{\mu,k+1}\right\|_{2}^{2}. \tag*{(7)}

Effective model

For analytical simplicity, we impose the following constraints on the trainable parameters:

WV:=(∗∗∗∗∗∗∗∗ID0DID0D∗∗∗∗),WK⊤WQ:=(∗∗A0D∗∗0D⊤0∗∗0D×D0D∗∗0D⊤0),(8)W_V := \begin{pmatrix} * & * & * & * \\ * & * & * & * \\ I_D & 0_D & I_D & 0_D \\ * & * & * & * \end{pmatrix}, \qquad W_K^{\top}W_Q := \begin{pmatrix} * & * & A & 0_D \\ * & * & 0_D^{\top} & 0 \\ * & * & 0_{D\times D} & 0_D \\ * & * & 0_D^{\top} & 0 \end{pmatrix}, \tag*{(8)}

where the rows from top to bottom and the columns from left to right are both partitioned into blocks of dimensions (D,1,D,1)(D,1,D,1). Here ∗* denotes a block that does not affect the readout, while A∈RD×DA \in\mathbb{R}^{D\times D} is the only learnable parameter. This restriction is less severe than it may appear: each constrained block is justified by one of three considerations: (i) the rotational symmetry of the data, which motivates choosing the fixed blocks so that they do not select preferred directions; (ii) the fact that fixing it entails no loss of generality in the setting studied here; or (iii) preventing the model from advancing its inference without consulting the context (see Appendix § 11 for details). These reductions concentrate the nontrivial parameter governing inference in the single matrix AA. Moreover, gradient-flow analyses of related models show that training from suitable initializations leads to analogous block structures (Huang et al., [16]; Javanmard et al., [17]). Under this parameterization, Equation (6) is given by

s^μ,k+1=sμ,k+1LXμ(Xμ⊤Asμ,k+ηyμ).(9)\hat{s}_{\mu,k+1}=s_{\mu,k}+\frac{1}{L}X_\mu\left(X_\mu^{\top}As_{\mu,k}+\eta y_\mu\right). \tag*{(9)}

and the effective learning problem derived from Equation (7) is

Leff(A):=1MK∑μ=1M∑k=0K−1∥s^μ,k+1−sμ,k+1∥22+λD∥A∥F2.(10)\mathcal{L}_{\mathrm{eff}}(A):=\frac{1}{MK}\sum_{\mu=1}^{M}\sum_{k=0}^{K-1}\left\|\hat{s}_{\mu,k+1}-s_{\mu,k+1}\right\|_{2}^{2}+\frac{\lambda}{D}\left\|A\right\|_{F}^{2}. \tag*{(10)}

References

  1. [1]Emmanuel Abbe, Samy Bengio, Aryo Lotfi, Colin Sandon, and Omid Saremi. How far can transformers reason? the globality barrier and inductive scratchpad. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds.), Advances in Neural Information Processing Systems, volume 37, pp. 27850–27895. Curran Associates, Inc., 2024. doi: 10.52202/079017-0874. URL https://proceedings.neurips.cc/paper_files/paper/2024/file/3107e4bdb658c79053d7ef59cbc804dd-Paper-Conference.pdf.DOI
  2. [2]Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=3zKtaqxThW.arxiv.org/abs/2306.13649
  3. [3]Conrad Ainslie, Pedram Hassanzadeh, Michael W. Mahoney, and Ashesh Chattopadhyay. Eigenanalysis framework for autoregressive neural emulators of multi-scale chaotic dynamics, 2026. URL https://arxiv.org/abs/2608.16084.
  4. [4]Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learning algorithm is in-context learning? investigations with linear models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=0q0X4H8YΝ4I.arxiv.org/abs/2211.15661
  5. [5]Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end algorithm synthesis with recurrent networks: Extrapolation without overthinking. In Advances in Neural Information Processing Systems, volume 35, pp. 20232–20242, 2022. doi: 10.52202/068431-1471. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/7f7031bde58ad59d83941dfa7d975aa-Abstract-Conference.html.DOI
  6. [6]Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, Usvsn Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar Van Der Wal. Pythia: A suite for analyzing large language models across training and scaling. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp. 2397–2430. PMLR, 2023. URL https://proceedings.mlr.press/v202/biderman23a.html.arxiv.org/abs/2304.01373
  7. [7]Adam Block, Dylan Foster, Akshay Krishnamurthy, Max Simchowitz, and Cyril Zhang. Butterfly effects of SGD noise: Error amplification in behavior cloning and autoregression. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=CgPs0419TQ.arxiv.org/abs/2310.11428
  8. [8]Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53, 2024. URL http://jmlr.org/papers/v25/23-0870.html.
  9. [9]Shivam Garg, Dimitris Tsipras, Percy Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes. In Advances in Neural Information Processing Systems, volume 35, pp. 30583–30598, 2022. doi: 10.52202/068431-2217. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/c529dba08a146ea8d6cf715ae8930cbe-Abstract-Conference.html.
  10. [10]Jonas Geiping, Sean Michael McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. Scaling up test-time compute with latent reasoning: A recurrent depth approach. In Advances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-1380. URL https://proceedings.neurips.cc/paper_files/paper/2025/hash/3b01972cf31e6fa0fe29e4b8b5c2a0a1-Abstract-Conference.html.DOI
  11. [11]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645(8081):633–638, 2025. doi: 10.1038/s41586-025-09422-z. URL https://www.nature.com/articles/s41586-025-09422-z.DOI
  12. [12]Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. In The Second Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=Itxz7S4Ip3.arxiv.org/abs/2412.06769
  13. [13]Michael Hassid, Gabriel Synnaeve, Yossi Adi, and Roy Schwartz. Don’t overthink it. preferring shorter thinking chains for improved LLM reasoning, 2025. URL https://arxiv.org/abs/2505.17813.
  14. [14]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), pp. 14852–14882. Association for Computational Linguistics, 2023. doi: 10.18653/v1/2023.acl-long.830. URL https://aclanthology.org/2023.acl-long.830/.DOI
  15. [15]Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. In Findings of the Association for Computational Linguistics: ACL 2023, pp. 8003–8017. Association for Computational Linguistics, 2023. doi: 10.18653/v1/2023.findings-acl.507. URL https://aclanthology.org/2023.findings-acl.507/.DOI
  16. [16]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. URL https://openreview.net/forum?id=r3DF5sOo5B.arxiv.org/abs/2502.21212
  17. [17]Adel Javanmard, Baharan Mirzasoleiman, and Vahab Mirrokni. Understanding the role of training data in test-time scaling. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=Y9fIDNa2nJ.arxiv.org/abs/2510.03605
  18. [18]Juno Kim and Taiji Suzuki. Transformers probably solve parity efficiently with chain of thought. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=n2N8ysDYop.arxiv.org/abs/2410.08633
  19. [19]Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, et al. DataComp-LM: In search of the next generation of training sets for language models. In Advances in Neural Information Processing Systems, volume 37, 2024. doi: 10.52202/079017-0455. URL https://proceedings.neurips.cc/paper_files/paper/2024/hash/19e4ea30dde58259665db3758885e412-Abstract-Datasets_and_Benchmarks_Track.html. Datasets and Benchmarks Track.
  20. [20]Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun (eds.), International Conference on Learning Representations, volume 2024, pp. 39578–39601, 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/file/aca97732e30bcf303bc22ac3924fd16-Paper-Conference.pdf.arxiv.org/abs/2305.20050
  21. [21]Bingbin Liu, Jordan T. Ash, Surbhi Goel, Akshay Krishnamurthy, and Cyril Zhang. Transformers learn shortcuts to automata. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=De4FYqjFueZ.arxiv.org/abs/2210.10749
  22. [22]Yue M. Lu, Mary Letey, Jacob A. Zavatone-Veth, Anindita Maiti, and Cengiz Pehlevan. Asymptotic theory of in-context learning by linear attention. Proceedings of the National Academy of Sciences, 122(28):e2502599122, 2025. doi: 10.1073/pnas.2502599122. URL https://doi.org/10.1073/pnas.2502599122.
  23. [23]Zihan Pengmei, Costas Mavromatis, Zhengyuan Shen, Yunyi Zhang, Vassilis N. Ioannidis, and Huzefa Rangwala. The kinetics of reasoning: How chain-of-thought shapes learning in transformers?, 2025. URL https://arxiv.org/abs/2510.25791.
  24. [24]Qihan Ren, Peng Wang, Ruikun Cai, Shuai Shao, Dadi Guo, Yuejin Xie, Yafu Li, Quanshi Zhang, Xia Hu, Jing Shao, and Dongrui Liu. Rethinking generalization in reasoning SFT: A conditional analysis on optimization, data, and model capability, 2026. URL https://arxiv.org/abs/2604.06628.
  25. [25]Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. CODI: Compressing chain-of-thought into continuous space via self-distillation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 677–693. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.emnlp-main.36. URL https://aclanthology.org/2025.emnlp-main.36/.
  26. [26]Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. Between underthinking and overthinking: An empirical study of reasoning length and correctness in LLMs, 2025. URL https://arxiv.org/abs/2505.00127.
  27. [29]Johannes Von Oswald, Eyyvind Niklasson, Ettore Randazzo, João Sacramento, Alexander Mordvintsev, Andrey Zhmoginov, and Max Vladymyrov. Transformers learn in-context by gradient descent. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp. 35151–35174. PMLR, 2023. URL https://proceedings.mlr.press/v202/von-oswald23a.html.arxiv.org/abs/2212.07677
  28. [30]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pp. 24824–24837, 2022. doi: 10.52202/068431-1800. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html.DOI
  29. [31]Jason Wei, Najoung Kim, Yi Tay, and Quoc V. Le. Inverse scaling can become U-shaped. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 15580–15591. Association for Computational Linguistics, 2023. doi: 10.18653/v1/2023.emnlp-main.963. URL https://aclanthology.org/2023.emnlp-main.963/.
  30. [32]Tung-Yu Wu and Melody Lo. U-shaped and inverted-U scaling behind emergent abilities of large language models. In The Thirteenth International Conference on Learning Representations, 2025. URL https://proceedings.iclr.cc/paper_files/paper/2025/hash/f696200311b445af686c3ebd87ade1d7-Abstract-Conference.html.arxiv.org/abs/2410.01692
  31. [33]Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. STaR: Bootstrap-ping reasoning with reasoning. In Advances in Neural Information Processing Systems, volume 35, pp. 15476–15488, 2022. doi: 10.52202/068431-1126. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/639a9a172c044fbb64175b5fad42e9a5-Abstract-Conference.html.
  32. [34]Yedi Zhang, Aaditya K. Singh, Peter E. Latham, and Andrew M. Saxe. Training dynamics of in-context learning in linear attention. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 76047–76087. PMLR, 2025. URL https://proceedings.mlr.press/v267/zhang25br.html.arxiv.org/abs/2501.16265
  33. [35]Dawei Zhu, Xiyu Wei, Guangxiang Zhao, Wenhao Wu, Haosheng Zou, Junfeng Ran, Xun Wang, Lin Sun, Xiangzheng Zhang, and Sujian Li. Chain-of-thought matters: Improving long-context language models with reasoning path supervision. In Findings of the Association for Computational Linguistics: EMNLP 2025, pp. 3197–3211. Association for Computational Linguistics, 2025a. doi: 10.18653/v1/2025.findings-emnlp.170. URL https://aclanthology.org/2025.findings-emnlp.170/.DOI
  34. [36]Hanlin Zhu, Shibo Hao, Zhiting Hu, Jiantao Jiao, Stuart Russell, and Yuandong Tian. Reasoning by superposition: A theoretical perspective on chain of continuous thought. In Advances in Neural Information Processing Systems, volume 38, 2025b. doi: 10.52202/085713-2670. URL https://proceedings.neurips.cc/paper_files/paper/2025/hash/72c363c2a573ca2128bd176d3317696b-Abstract-Conference.html.DOI

Paper details

vlnc:W_vw5cvnte
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

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

  • Computer Science

Generative AI tools

View statement
Looking for related papers…
Contents