Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read By amplifying a few last-layer activations, the paper elicits long self-correcting reasoning chains at test time with no training.

desk verdict A plausible training-free trick that moves reasoning accuracy a few points, but the mechanism claims and self-reflection metric need tighter controls before I'd trust the story. read the letter →

arxiv 2505.17697 v1 pith:OYG7BYHP submitted 2025-05-23 cs.CL cs.LG

Zekai Zhao , Qi Liu , Kun Zhou , Zihan Liu , Yifei Shao , Zhiting Hu , Biwei Huang This is my paper
classification cs.CLcs.LG
keywords longchain-of-thoughtactivationcontrolMLPanalysisself-reflectiontraining-freeelicitationparameter-efficientfine-tuningreasoningbenchmarksinterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

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 tries to establish that long chain-of-thought (CoT) reasoning — the slow, self-correcting deliberation seen in recent 'thinking' models — is a latent, steerable property of base language models rather than a capability that must be produced by expensive reinforcement learning or fine-tuning on distilled data. The authors show that a small set of high-impact activations in the last few MLP layers governs long-form reasoning attributes such as output length and self-reflection, and that amplifying those activations while inserting 'wait' tokens elicits long CoTs without any training. The activation dynamics around the trigger token follow a sharp rise and then a logarithmic decay, which they fit with the analytic function $f(t) = a - b\log(t+c)$ and apply at inference time as a multiplier on the selected activations. On complex math and science benchmarks, this training-free intervention raises both accuracy and self-reflection rates across several model families and sizes. A parameter-efficient fine-tuning variant that updates only a last-layer amplification module and low-rank adapters (1.51% of parameters) matches or exceeds full fine-tuning, supporting the paper's claim that this reasoning ability is concentrated in a small, identifiable set of parameters.

What carries the argument

The load-bearing object is the sparse set of high-impact MLP activations in the final few layers, identified by contrastive activation analysis: the authors feed question–response pairs that differ in correctness, length, and self-reflection, rank neurons by average activation difference, and keep the top 150–200. The second piece is an analytic activation schedule, $A' = A\cdot(1+\alpha f(t))$ with $f(t) = a - b\log(t+c)$, which reproduces the sharp rise and logarithmic decay observed after the 'wait' token; fitting its three coefficients on a few contrastive examples lets the intervention be applied to any base model. The third piece is the reflection trigger: a rule that inserts 'wait' whenever the previous sentence contains enough digits, gated by a cool-down window. Together these form EELo-CoT, a training-free controller that steers the model's reasoning style at test time.

What would settle it

Run the same activation-selection procedure on length-matched contrastive pairs (long and correct versus long and incorrect, both with reflection phrases) and check whether the top-ranked activations still coincide with those found by the original pairs; if the neuron sets diverge almost completely, the claimed mechanism is an artifact of correlated length and correctness. A companion check is to replace the 'wait' token with an arbitrary token at matched positions while keeping the fitted amplification schedule; if accuracy and reflection gains persist, the trigger word is not special and the reflection effect is driven purely by the amplification.

Watch

Extended reading notes

Core claim

The central discovery is that a handful of MLP-neuron activations in the final layers of a language model acts as a control panel for long-CoT behavior: output length, self-reflection frequency, and, to a substantial degree, answer accuracy all track the values of these activations. Contrastive examples — correct, long, self-reflecting responses versus incorrect, short, non-reflecting ones — pick out roughly 150–200 key neurons, and simply multiplying their values by a constant factor while inserting a 'wait' token after sentences containing several digits raises self-reflection rates and accuracy on Math500, AMC23, and GPQA. The activation time series around the trigger token follows a sharp rise then a logarithmic decay, captured by the fitted family $f(t) = a - b\log(t+c)$, so the intervention is training-free: at inference the model amplifies its own activations according to the fitted curve. The same localization supports a training recipe in which only the last-layer amplification gate and a low-rank adapter set are learned, updating 1.51% of parameters. The paper interprets these results as evidence that the long-CoT capability mostly pre-exists in base models and that post-training, which often leaves these activations 'dead', mainly suppresses or fails to surface it.

Load-bearing premise

The load-bearing premise is that the contrastive pairs — correct, long, self-reflecting responses versus incorrect, short, non-reflecting ones — isolate a single reusable 'long-CoT ability' trait; if the selected activations actually encode answer difficulty, domain style, or plain verbosity instead, the intervention's gains could come from a confound rather than a reasoning mechanism.

Editorial extensions

If this is right

  • The training-free controller improves accuracy and self-reflection rate on Math500, AMC23, and GPQA across base models of 1.5B, 3B, 7B, and 32B parameters, and on a different 8B model family, with no gradient updates.
  • The parameter-efficient training variant, updating 1.51% of parameters, reaches or exceeds full fine-tuning and LoRA on the same benchmarks, indicating that long-CoT ability is localized enough to be acquired by a small amplification module.
  • Because only a few contrastive examples are needed to fit the coefficients, the activation-control recipe transfers to new models and datasets at near-zero marginal cost.
  • The observation that instruct models' key activations are inactive or 'dead' provides an explanation for why over-post-trained models are harder to adapt to long-CoT behavior through fine-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the identified activations are a true style governor, the same contrastive procedure could be run with inverted signs to suppress reflection or verbosity, a test the paper leaves implicit.
  • The digit-count trigger is a heuristic; a learned or searched trigger token, or an intervening classifier over the hidden state, could make the reflection timing more robust on non-mathematical tasks.
  • Because the contrastive pairs vary correctness, length, and reflection together, a cleaner test of the mechanism would use length-matched pairs and compare whether the same neurons encode reflection independently of verbosity.
  • The method's success on a model family not tuned for mathematics suggests the mechanism may be generic to decoder-only transformers; the paper's stated next steps — multimodal models, planning tasks, and non-transformer architectures — would directly test that generality.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes EELo-CoT, a training-free method for eliciting long chain-of-thought (CoT) reasoning from base language models. It identifies a small set of MLP activations by contrasting correct, long, self-reflective responses with incorrect, short, non-reflective responses, then amplifies those activations at inference time according to a fitted logarithmic decay function and inserts “wait” tokens to trigger self-reflection. The authors report substantial increases in self-reflection rate and modest, model-dependent accuracy gains on Math500, AMC23, and GPQA, and they also present a parameter-efficient fine-tuning variant that trains only 1.51% of parameters. The paper claims that long-CoT-related activations concentrate in the last few layers, are sparse, and follow predictable dynamics.

Significance. If the mechanism were established, this would be a significant result: it would offer a training-free route to slow-thinking behavior, a lightweight fine-tuning target, and public code and data for reproducibility. The accuracy improvements are non-circular evidence that the intervention can be useful, and the observation that a small number of activations dominates the effect is worth investigating. However, the current evidence does not isolate a reusable “long-CoT ability”: the activation-selection contrasts are confounded, the self-reflection metric is partly mechanical, and the reported gains lack error bars and consistency checks. The paper is therefore best viewed as an interesting empirical intervention whose mechanistic interpretation requires additional control experiments. The authors' limitations section is candid about several of these issues, but the main claims go beyond what the current experiments support.

major comments (5)
  1. [Section 3.2 (Implementation Details), Section 2.1] The contrastive groups used to identify key activations differ simultaneously in correctness, output length, and the presence of reflection phrases: Group 1 is correct, exceeds 1000 tokens, and contains self-reflection, while Group 2 is incorrect, shorter than 1000 tokens, and contains no reflection. The per-neuron activation difference that selects the top-150/200 activations therefore cannot be attributed specifically to long-CoT reasoning; it may encode answer difficulty, verbosity, formatting, or other correlated factors. Since the central causal claim—that a small set of last-layer activations governs long-form reasoning—depends on attribute-specific selection, the paper should add matched control experiments, for example length-matched and difficulty-matched contrasts that vary only the reflection behavior, and should compare the selected activations against randomly chosen activations or activations selected by a difficulty-matched contrast.
  2. [Section 2.2, Table 1, Section 3.1] The self-reflection rate is computed by matching phrases including “wait” and “let me double check”, while the intervention literally inserts “wait” tokens into the generated text. The large reflection-rate increases (e.g., from 10.2% to 49.4% or from 73.8% to 90.6% in Table 1) are therefore partly a definitional artifact. The paper should report a self-reflection rate that excludes intervention-inserted tokens, or otherwise separate model-initiated reflection from externally triggered reflection, before claiming that the method “activates” self-reflection.
  3. [Table 1, Table 4, Table 5] Accuracy gains are small and inconsistent across models and benchmarks, and no error bars, number of seeds, or significance tests are provided. For example, Qwen2-7B-base on Math500 improves from 30.80% to 31.20%, while Qwen2.5-7B-Math-base on AMC23 shows no gain (65.00% to 65.00%) and LLaMA-3.1-8B on Math500 goes from 9.8% to 11.8%. Several baseline rows decrease accuracy relative to the base model. Without repeated runs and uncertainty quantification, the claim that EELo-CoT reliably improves accuracy and outperforms the baselines is not supported. At minimum, the authors should report multiple seeds and confidence intervals, and ideally pass@k estimates for the accuracy comparisons.
  4. [Section 3.1, Eq. (1), Section 2.3] The analytic decay function f(t)=a−b·log(t+c) is presented with fitted coefficients a=0.17, b=0.033, c=−0.997, but no details are given about the fitting procedure: the number of trajectories, the token window, the train/test split, or the goodness of fit. With c close to −1, the function is nearly singular at small t, so it may overfit noise rather than capture a robust dynamical pattern. The paper also does not test whether this logarithmic form is preferable to a constant amplification or a linear alternative. In addition, the abstract states “exponential decay” while Eq. (1) and Finding-6 describe a logarithmic decay; the terminology should be aligned. A sensitivity analysis of the fitted coefficients and their influence on downstream accuracy is needed.
  5. [Section 4, Table 3] The parameter-efficient fine-tuning comparison is reported as single accuracy numbers without repetitions or uncertainty estimates, and the baselines are not matched for capacity: LoRA uses rank 256, while EELo-CoT uses rank 64 LoRA plus an additional amplification module. The claim that EELo-CoT “outperforms full LoRA fine-tuning” with 1.51% parameters would be strengthened by a rank-matched LoRA baseline and by reporting variance across seeds or hyperparameter settings, especially since the differences on Math500 (91.60 vs. 90.20) and GPQA (70.02 vs. 69.19) are small.
minor comments (4)
  1. [Section 3.2] The implementation details state that the number of activations is 150 with an “amplification factor set as 4”, but Section 2.2 uses factors 1.2, 1.4, and 1.6 and Eq. (2) defines a scaling factor α; the relationship between these quantities should be clarified.
  2. [Sections 2.2, 3.2, and 4.1] The number of selected key activations is inconsistent across the paper: top-200 in Section 2.2, 150 in Section 3.2, and n=100 in Section 4.1. The authors should state whether these are different experimental choices and justify the differences.
  3. [Table 4] The caption of Table 4 says “Length is the average number of word counts”, while all other length columns are measured in tokens; the units should be made consistent.
  4. [Section 3.3 and Figure captions] In Table 1, the row label “Our Rule” is used in the text but the table row says “+ EELo-CoT (Ours)”; the terminology should be unified throughout.

Circularity Check

1 steps flagged · score 6.0 of 10

Self-reflection gains are partly tautological: the metric counts 'wait' tokens while the method inserts them; accuracy gains are the only non-circular evidence.

  1. self definitional [Section 2.2 (Analysis Setup) and Section 3.1 (Forcing Reflection after Reasoning)]
    ""The self-reflection rate is computed by computing the percentage of responses that contain special reflection tokens and phrases, e.g., 'wait' and 'let me double check'." ... "Once we detect k or more digits, we insert a 'wait' token in the starting position of the next sentence.""

    The operational definition of the headline metric is the presence of reflection phrases including 'wait'; the intervention is the literal insertion of the token 'wait' whenever a digit-count trigger fires. Therefore any response that receives the intervention automatically contains 'wait' and is counted as self-reflective, independent of whether the model actually engages in self-reflection. The reported jumps (e.g., Qwen2-7B-base 3.20% to 65.20% on Math500; Qwen2.5-7B-base 10.2% to 49.4%) are guaranteed by the measurement definition, not by the discovered activation mechanism. This is a self-definitional reduction: the outcome 'self-reflection' is defined in terms of the exact intervention applied.

full rationale

Score 6, not higher, because the paper contains independent content: the top-k activation selection is fitted to contrastive data and the accuracy gains (e.g., 45.00% to 57.50% on AMC23 and 30.30% to 35.86% on GPQA for Qwen2.5-7B-base) are not forced by the reflection metric. However, a central dependent variable—self-reflection rate—is defined by the presence of the very token the method inserts, so the headline 'significantly increased self-reflection rates' reduces by construction. The analytic-function coefficients are fitted to activation trajectories and then applied as an intervention; that is transparent curve-fitting rather than circularity. The confounded contrastive design (correct/long/reflective vs incorrect/short/non-reflective) is a validity threat, but it does not make the activation selection equivalent to its inputs by definition. No load-bearing self-citation was found.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central method rests on several manually chosen components: a contrastive pair definition, a threshold for selecting activations, a fixed log-decay functional form with fitted coefficients, an amplification scale, a digit-count trigger, and a cool-down window. None of these are derived from first principles; they are calibrated on data and evaluated on held-out benchmarks, which limits but does not eliminate their ad hoc character.

free parameters (7)
  • Log-decay coefficients a, b, c = a=0.17, b=0.033, c=-0.997 for Qwen2.5-7B-base
    Fitted to observed activation trajectories around 'wait' tokens; used in Eq. (1) to set the amplification schedule.
  • Amplification scaling factor alpha = 4 for training-free experiments; 1.2, 1.4, 1.6 explored in analysis
    Controls intervention intensity in Eq. (2); chosen by hand without sensitivity analysis.
  • Number of key activations n = 150 for training-free, 100 for training module; top-200 in analysis
    Manual selection of how many neurons to amplify.
  • Activation selection threshold = >4
    Activations whose average contrastive difference exceeds 4 are labeled long-CoT related.
  • Digit-count trigger k = 5
    If the last sentence has at least 5 digits, insert a 'wait' token.
  • Cool-down window = 4 sentences
    Temporarily disables forced reflection to avoid meaningless loops.
  • LoRA ranks and beta = LoRA rank 64 on first 63 layers; activation module width 100; beta trainable
    Hyperparameters for the parameter-efficient fine-tuning module.
assumptions (5)
  • domain assumption Contrastive filtering (positive: correct, over 1000 tokens, reflective; negative: incorrect, shorter, non-reflective) isolates long-CoT ability as a single attribute.
    Invoked in Section 3.2 to construct training pairs; if false, the identified activations encode confounds like difficulty or verbosity.
  • domain assumption MLP activations in the last layers are causally responsible for reasoning style, not merely correlated with it.
    The intervention in Section 2.2 assumes amplifying these activations will elicit long-CoT behavior.
  • ad hoc to paper The 'wait' token acts as a general trigger for self-reflection across models and tasks.
    Used as the trigger in Sections 2.2 and 3.1; no independent evidence it is optimal or general.
  • ad hoc to paper A logarithmic decay function f(t)=a-b*log(t+c) captures the essential activation dynamics.
    Chosen after visual inspection in Finding-6; no derivation or goodness-of-fit reported.
  • ad hoc to paper Digit count in the last sentence is a valid proxy for when self-reflection is useful.
    The forcing-reflection rule in Section 3.1 inserts 'wait' when the last sentence contains k or more digits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models." pith.science (2026). https://pith.science/paper/OYG7BYHP

@misc{pith2026250517697,
  author       = {Pith},
  title        = {Pith review of: Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OYG7BYHP}},
  note         = {Machine review of arXiv:2505.17697}
}
read the original abstract

Despite the remarkable reasoning performance, eliciting the long chain-of-thought (CoT) ability in large language models (LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers largely governs long-form reasoning attributes, such as output length and self-reflection. By simply amplifying these activations and inserting "wait" tokens, we can invoke the long CoT ability without any training, resulting in significantly increased self-reflection rates and accuracy. Moreover, we find that the activation dynamics follow predictable trajectories, with a sharp rise after special tokens and a subsequent exponential decay. Building on these insights, we introduce a general training-free activation control technique. It leverages a few contrastive examples to identify key activations, and employs simple analytic functions to modulate their values at inference time to elicit long CoTs. Extensive experiments confirm the effectiveness of our method in efficiently eliciting long CoT reasoning in LLMs and improving their performance. Additionally, we propose a parameter-efficient fine-tuning method that trains only a last-layer activation amplification module and a few LoRA layers, outperforming full LoRA fine-tuning on reasoning benchmarks with significantly fewer parameters. Our code and data are publicly released.

Figures

Figures reproduced from arXiv: 2505.17697 by the authors.

Figure 1
Figure 1. (a) Sparse Activations when processing Long CoT 1.0 1.2 1.4 1.6 Amplification factor 60 65 70 Accuracy (%) Constant Intervene Wait Token Insert [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 5
Figure 5. (b) Activation Patterns of base and long CoT LLMs 0 500 1000 Token 0 5 10 15 Activation Value Qwen2.5-7B-Instruct [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 7
Figure 7. The overall framework of the proposed Long-CoT elicitation method. Based on the [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figures from the paper (2 more)
Figure 9
Figure 9. Figure 9: From the figure, the average value of top 100 activations around the "wait" token follow a [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 8
Figure 8. Figure 8: Wait Token Activa￾tion Trend from the R1 distilled Qwen Model 50 0 50 100 Position relative to token 0.000 0.025 0.050 0.075 0.100 Average activation 'Wait' Token [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Logit Arithmetic Elicits Long Reasoning Capabilities Without Training

    cs.CL 2025-07 conditional novelty 4.0 of 10

    ThinkLogit blends logits from a small reasoning guider into a frozen 32B model, improving math pass@1 by up to 29% without training the large model.

Reference graph

Works this paper leans on

58 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    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

    arXiv 2025

  2. [2]

    Qwq: Reflect deeply on the boundaries of the unknown

    Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm. github.io/blog/qwq-32b-preview/ , 2024. Published November 28, 2024; accessed May 15, 2025

    work page 2024

  3. [3]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    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

    arXiv 2024

  4. [4]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599, 2025

    arXiv 2025

  5. [5]

    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, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  6. [6]

    Redstar: Does scaling long-cot data unlock better slow-reasoning systems?arXiv preprint arXiv:2501.11284, 2025

    Haotian Xu, Xing Wu, Weinong Wang, Zhongzhi Li, Da Zheng, Boyuan Chen, Yi Hu, Shijia Kang, Jiaming Ji, Yingying Zhang, et al. Redstar: Does scaling long-cot data unlock better slow-reasoning systems?arXiv preprint arXiv:2501.11284, 2025

    arXiv 2025

  7. [7]

    Demystifying long chain-of-thought reasoning in llms.arXiv preprint arXiv:2502.03373, 2025

    Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. Demystifying long chain-of-thought reasoning in llms.arXiv preprint arXiv:2502.03373, 2025

    arXiv 2025

  8. [8]

    Distilling system 2 into system 1.arXiv preprint arXiv:2407.06023, 2024

    Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. Distilling system 2 into system 1.arXiv preprint arXiv:2407.06023, 2024

    arXiv 2024

Show all 58 references
  1. [9]

    Rlhf deciphered: A critical analysis of reinforcement learning from human feedback for llms.arXiv preprint arXiv:2404.08555, 2024

    Shreyas Chaudhari, Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande, and Bruno Castro da Silva. Rlhf deciphered: A critical analysis of reinforcement learning from human feedback for llms.arXiv preprint arXiv:2404.08555, 2024

    2024 arXiv

  2. [10]

    Rrhf: Rank responses to align language models with human feedback without tears.arXiv preprint arXiv:2304.05302, 2023

    Zheng Yuan, Hongyi Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang. Rrhf: Rank responses to align language models with human feedback without tears.arXiv preprint arXiv:2304.05302, 2023

    2023 arXiv

  3. [11]

    Language models are multilingual chain-of-thought reasoners.arXiv preprint arXiv:2210.03057, 2022

    Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush V osoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, et al. Language models are multilingual chain-of-thought reasoners.arXiv preprint arXiv:2210.03057, 2022

    arXiv

  4. [12]

    Activation addition: Steering language models without optimization

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. arXiv e-prints, pages arXiv–2308, 2023

    2023

  5. [13]

    Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space.arXiv preprint arXiv:2203.14680, 2022

    Mor Geva, Avi Caciularu, Kevin Ro Wang, and Yoav Goldberg. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space.arXiv preprint arXiv:2203.14680, 2022

    2022 arXiv

  6. [14]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. NeurIPS, 2021

    2021

  7. [15]

    Qwen2.5: A party of foundation models, September 2024

    Qwen Team. Qwen2.5: A party of foundation models, September 2024

    2024

  8. [16]

    Overtrained language models are harder to fine-tune.arXiv preprint arXiv:2503.19206, 2025

    Jacob Mitchell Springer, Sachin Goyal, Kaiyue Wen, Tanishq Kumar, Xiang Yue, Sadhika Malladi, Graham Neubig, and Aditi Raghunathan. Overtrained language models are harder to fine-tune.arXiv preprint arXiv:2503.19206, 2025. 10

    2025 arXiv

  9. [17]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. InFirst Conference on Language Modeling, 2024

    2024

  10. [18]

    Llms can easily learn to reason from demonstrations structure, not content, is what matters!arXiv preprint arXiv:2502.07374, 2025

    Dacheng Li, Shiyi Cao, Tyler Griggs, Shu Liu, Xiangxi Mo, Eric Tang, Sumanth Hegde, Kourosh Hakhamaneshi, Shishir G Patil, Matei Zaharia, et al. Llms can easily learn to reason from demonstrations structure, not content, is what matters!arXiv preprint arXiv:2502.07374, 2025

    2025 arXiv

  11. [19]

    Limo: Less is more for reasoning.arXiv preprint arXiv:2502.03387, 2025

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning.arXiv preprint arXiv:2502.03387, 2025

    2025 arXiv

  12. [20]

    Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    2021 arXiv

  13. [21]

    Efficient memory management for large language model serving with pagedattention

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of the 29th Symposium on Operating Systems Principles, pa...

    2023

  14. [22]

    Gpt-4 technical report.ArXiv, abs/2303.08774, 2023

    OpenAI. Gpt-4 technical report.ArXiv, abs/2303.08774, 2023

    2023 arXiv

  15. [23]

    The claude 3 model family: Opus, sonnet, haiku

    Anthropic. The claude 3 model family: Opus, sonnet, haiku. Technical report, Anthropic, March 2024

    2024

  16. [24]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.ArXiv, abs/2403.05530, 2024

    Google Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.ArXiv, abs/2403.05530, 2024

    2024 arXiv

  17. [25]

    Qwen3 Technical Report.arXiv e-prints, page arXiv:2505.09388, May 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

    2025 arXiv

  18. [26]

    The claude 3 model family: Opus, sonnet, haiku

    Anthropic. The claude 3 model family: Opus, sonnet, haiku. https://www-cdn.anthropic. com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf, March 2024. Model Card

    2024

  19. [27]

    Gemini 2.5: Our most intelligent ai model

    Koray Kavukcuoglu. Gemini 2.5: Our most intelligent ai model. https://blog.google/ technology/google-deepmind/gemini-model-thinking-updates-march-2025/ #gemini-2-5-thinking, March 2025. Blog post

    2025

  20. [28]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730...

    2022

  21. [29]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023

    2023

  22. [30]

    Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023. 11

    2023

  23. [31]

    Solving quantitative reasoning problems with language models.ArXiv, abs/2206.14858, 2022

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Venkatesh Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. Solving quantitative reasoning problems with languag...

    2022 arXiv

  24. [32]

    Jiang, Jia Deng, Stella Biderman, and Sean Welleck

    Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics.ArXiv, abs/2310.10631, 2023

    2023 arXiv

  25. [33]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, R. X. Xu, Jun-Mei Song, Mingchuan Zhang, Y . K. Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.ArXiv, abs/2402.03300, 2024

    2024 arXiv

  26. [34]

    Singhal, Shekoofeh Azizi, Tao Tu, Said Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Kumar Tanwani, Heather J

    K. Singhal, Shekoofeh Azizi, Tao Tu, Said Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Kumar Tanwani, Heather J. Cole-Lewis, Stephen J. Pfohl, P A Payne, Martin G. Seneviratne, Paul Gamble, Chris Kelly, Nathaneal Scharli, Aakanksha Chowdhery, P. A. Mans- field, Bla...

    2022

  27. [35]

    Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic

    Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony S. Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. Galactica: A large language model for science.ArXiv, abs/2211.09085, 2022

    2022 arXiv

  28. [36]

    Pal: Program-aided language models.arXiv preprint arXiv:2211.10435, 2022

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. Pal: Program-aided language models.arXiv preprint arXiv:2211.10435, 2022

    2022 arXiv

  29. [37]

    Contrastive decoding: Open-ended text generation as optimization

    Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. InAnnual Meeting of the Association for Computational Linguistics, 2022

    2022

  30. [38]

    Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992, 2023

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhit- ing Hu. Reasoning with language model is planning with world model.arXiv preprint arXiv:2305.14992, 2023

    2023 arXiv

  31. [39]

    rstar-math: Small llms can master math reasoning with self-evolved deep thinking.arXiv preprint arXiv:2501.04519, 2025

    Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. rstar-math: Small llms can master math reasoning with self-evolved deep thinking.arXiv preprint arXiv:2501.04519, 2025

    2025 arXiv

  32. [40]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    2022 arXiv

  33. [41]

    Improve mathematical reasoning in language models by automated process supervision.arXiv preprint arXiv:2406.06592, 2024

    Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Meiqi Guo, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, et al. Improve mathematical reasoning in language models by automated process supervision.arXiv preprint arXiv:2406.06592, 2024

    2024 arXiv

  34. [42]

    An investigation of neuron activation as a unified lens to explain chain-of-thought eliciting arithmetic reasoning of llms.arXiv preprint arXiv:2406.12288, 2024

    Daking Rai and Ziyu Yao. An investigation of neuron activation as a unified lens to explain chain-of-thought eliciting arithmetic reasoning of llms.arXiv preprint arXiv:2406.12288, 2024

    2024 arXiv

  35. [43]

    Unlocking general long chain-of-thought reasoning capabilities of large language models via representation engineering.arXiv preprint arXiv:2503.11314, 2025

    Xinyu Tang, Xiaolei Wang, Zhihao Lv, Yingqian Min, Wayne Xin Zhao, Binbin Hu, Ziqi Liu, and Zhiqiang Zhang. Unlocking general long chain-of-thought reasoning capabilities of large language models via representation engineering.arXiv preprint arXiv:2503.11314, 2025

    2025 arXiv

  36. [44]

    Adaptive group policy optimization: Towards stable training and token-efficient reasoning.arXiv preprint arXiv:2503.15952, 2025

    Chen Li, Nazhou Liu, and Kai Yang. Adaptive group policy optimization: Towards stable training and token-efficient reasoning.arXiv preprint arXiv:2503.15952, 2025

    2025

  37. [45]

    Hybrid group relative policy optimization: A multi-sample approach to enhancing policy optimization.arXiv preprint arXiv:2502.01652, 2025

    Soham Sane. Hybrid group relative policy optimization: A multi-sample approach to enhancing policy optimization.arXiv preprint arXiv:2502.01652, 2025. 12

    2025 arXiv

  38. [46]

    Expanding rl with verifiable rewards across diverse domains.arXiv preprint arXiv:2503.23829, 2025

    Yi Su, Dian Yu, Linfeng Song, Juntao Li, Haitao Mi, Zhaopeng Tu, Min Zhang, and Dong Yu. Expanding rl with verifiable rewards across diverse domains.arXiv preprint arXiv:2503.23829, 2025

    2025 arXiv

  39. [47]

    Opencodereasoning: Advancing data distillation for competitive coding.arXiv preprint arXiv:2504.01943, 2025

    Wasi Uddin Ahmad, Sean Narenthiran, Somshubra Majumdar, Aleksander Ficek, Siddhartha Jain, Jocelyn Huang, Vahid Noroozi, and Boris Ginsburg. Opencodereasoning: Advancing data distillation for competitive coding.arXiv preprint arXiv:2504.01943, 2025

    2025 arXiv

  40. [48]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations.arXiv preprint arXiv:2312.08935, 2023

    Peiyi Wang, Lei Li, Zhihong Shao, RX Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations.arXiv preprint arXiv:2312.08935, 2023

    2023 arXiv

  41. [49]

    Generative verifiers: Reward modeling as next-token prediction.arXiv preprint arXiv:2408.15240, 2024

    Lunjun Zhang, Arian Hosseini, Hritik Bansal, Mehran Kazemi, Aviral Kumar, and Rishabh Agarwal. Generative verifiers: Reward modeling as next-token prediction.arXiv preprint arXiv:2408.15240, 2024

    2024 arXiv

  42. [50]

    Code to think, think to code: A survey on code-enhanced reasoning and reasoning-driven code intelligence in llms.arXiv preprint arXiv:2502.19411, 2025

    Dayu Yang, Tianyang Liu, Daoan Zhang, Antoine Simoulin, Xiaoyi Liu, Yuwei Cao, Zhaopu Teng, Xin Qian, Grey Yang, Jiebo Luo, et al. Code to think, think to code: A survey on code-enhanced reasoning and reasoning-driven code intelligence in llms.arXiv preprint arXiv:2502.19411, 2025

    2025 arXiv

  43. [51]

    Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

    Yifan Du, Zikang Liu, Yifan Li, Wayne Xin Zhao, Yuqi Huo, Bingning Wang, Weipeng Chen, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. Virgo: A preliminary exploration on reproducing o1-like mllm.arXiv preprint arXiv:2501.01904, 2025

    2025 arXiv

  44. [52]

    Towards best practices of activation patching in language models: Metrics and methods.arXiv preprint arXiv:2309.16042, 2023

    Fred Zhang and Neel Nanda. Towards best practices of activation patching in language models: Metrics and methods.arXiv preprint arXiv:2309.16042, 2023

    2023 arXiv

  45. [53]

    Transformer feed-forward layers are key-value memories.arXiv preprint arXiv:2012.14913, 2020

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories.arXiv preprint arXiv:2012.14913, 2020

    2012 arXiv

  46. [54]

    Knowledge neurons in pretrained transformers.arXiv preprint arXiv:2104.08696, 2021

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers.arXiv preprint arXiv:2104.08696, 2021

    2021 arXiv

  47. [55]

    Discovering latent knowledge in language models without supervision.arXiv preprint arXiv:2212.03827, 2022

    Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt. Discovering latent knowledge in language models without supervision.arXiv preprint arXiv:2212.03827, 2022

    2022 arXiv

  48. [56]

    Qwen2.5-vl technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Han...

    2025 arXiv

  49. [57]

    Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024

    Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. Osworld: Benchmarking multimodal agents for op...

    2024

  50. [58]

    Wait, I see the mistake here,

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 13 Table 4: The evaluation results of our method using Llama-3.1-8B-base on Math500 and GPQA. Acc is the accuracy rate measured in percentage, Length...

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.