REVIEW 5 major objections 7 minor 33 references
Exploring and Exploiting the Inherent Efficiency within Large Reasoning Models for Self-Guided Efficiency Enhancement
T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Large reasoning models already know how to reason more briefly, and a single steering direction can unlock that latent efficiency.
desk verdict Useful empirical finding and a promising RL reward, but the steering mechanism has an unaddressed positional mismatch and the reporting needs more rigor. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying object is the 'efficiency direction' $v_l = \mu_{\text{efficient}} - \mu_{\text{verbose}}$, a difference-in-means vector in the residual stream at the final input token, built from 1,000 shortest and 1,000 longest correct paths sampled from GSM8K. At inference the intervention is $h'_l = h_l + \lambda v_l$, applied only at the last input token and only at one selected layer; the paper reports that the optimal $\lambda$ is selected empirically. The RL counterpart is a self-reward $r(y_j) = \lambda_1 \mathbf{1}(y_j = y^*_i) - \lambda_2 \max(0, \ell(y_j) - \ell_{\text{Min\_Correct}})$, where $\ell_{\text{Min\_Correct}}$ is the shortest correct length among eight rollouts of the same problem, so the model's own best current trace defines the efficiency target and the policy is updated with GRPO. The difference-in-means vector does the work of localizing where efficient and verbose reasoning diverge; the self-reward does the work of converting that divergence into a learning signal.
What would settle it
A decisive check would be to freeze the steering layer and the lambda-selection rule in advance, then run Efficiency Steering on a held-out competition benchmark such as AIME 2024: if no lambda value shortens the average trace by at least 20 percent while keeping pass@1 within one point of the unsteered model, the claim of a universal efficiency direction fails. A second check is to recompute the steering vector from MATH Level 1 traces alone and apply it to Level 5 problems; a large accuracy drop would show the direction encodes problem ease rather than general efficiency.
Extended reading notes
Core claim
On its own terms, the paper's central discovery is that a large reasoning model already knows how to reason more briefly than it usually does, and that this latent ability is encoded in a simple, roughly linear structure. When many responses to one problem are sampled, the shortest correct trace is often less than half the length of the longest correct trace, and these two kinds of traces separate cleanly in PCA projections of the residual stream at every layer and difficulty level. Injecting the difference-of-means vector between those trace types at the final input token's hidden state, scaled by a coefficient lambda, shortens generation for positive lambda and lengthens it for negative lambda across all seven models studied; accuracy is preserved or slightly improved within a reasonable lambda range. The behavioral correlate of the representation shift is less self-reflection, namely fewer 'Wait' and 'Alternatively' markers and fewer Reflection and Transition phases, which the paper turns into a reinforcement-learning reward that penalizes any correct rollout longer than the shortest correct rollout in its batch. The author would state the claim as: reasoning efficiency is an intrinsic, steerable property of LRMs, and both representation-space steering and self-rewarded RL can unlock it.
Load-bearing premise
The load-bearing premise is that the efficiency signal extracted from a few correct examples, namely the mean difference between shortest and longest successful traces or the shortest correct trace in a rollout, generalizes to unseen problems and can be applied at one token position without suppressing reasoning the problem actually needs.
Editorial extensions
If this is right
- Efficiency Steering reduces average trace length on every model/benchmark pair tested, often by 20 to 30 percent, with pass@1 equal or slightly higher; for example, R1-Distill-Qwen-7B on MATH-500 drops from about 3,496 to 2,560 tokens while accuracy rises from 92.20 to 92.60.
- Self-Rewarded Efficiency RL removes tokens more aggressively, exceeding 50 percent on several settings such as R1-Distill-Qwen-7B on AMC and MATH-500, while accuracy changes stay small and sometimes positive.
- Most of the RL efficiency gain appears within the first 10 to 30 optimization steps, so the method is cheap; past roughly 30 steps, harder benchmarks such as AIME 2025 begin to trade accuracy for brevity.
- Training the RL objective on MATH levels 1 through 3 yields shorter reasoning on all four test benchmarks than training on levels 4 through 5, with comparable accuracy, meaning easy data transfers a brevity prior to hard problems.
- Neither method needs a curated short-chain dataset, an external reward model, or a fixed token budget; the shortest correct response in the model's own rollouts is the target.
Reading between the lines
- If the single-direction result holds beyond math, it suggests an inference-time dial: a deployment operator could tune lambda continuously to trade latency against accuracy per workload, without retraining or prompt engineering, and even per difficulty level.
- The same difference-in-means recipe could in principle isolate other reasoning attributes, such as verification effort, caution, or reliance on worked examples, by choosing different contrast sets; the paper's control experiment only demonstrates bidirectional control of length, so this is an extension.
- A testable follow-up is whether the efficiency direction survives after a model is fine-tuned with the RL reward; if RL collapses the representational separation, steering and RL would become alternative rather than additive tools, whereas if it persists the two could be combined.
- The shortest-correct-trace target is naturally noisy on hard problems where a model may draw the same final answer through a partially flawed path; evaluating the method on answers with verified reasoning, not just numeric equality, would separate brevity gains from shortcut exploitation.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies 'overthinking' in large reasoning models (LRMs) and argues that the capacity for concise reasoning is latent in the model's own representations and can be elicited without external supervision. It first documents that, across seven LRMs and the MATH dataset, the shortest correct reasoning traces are often less than half the length of the longest correct traces. It then claims a representational analysis (PCA of residual-stream activations at the final reasoning token) shows shortest and longest correct paths are linearly separable. Based on this, the paper introduces two methods: Efficiency Steering, a training-free activation intervention that adds a difference-in-means vector, computed from 1,000 shortest and 1,000 longest correct GSM8K traces, to the last input token at a single layer; and Self-Rewarded Efficiency RL, a GRPO variant whose reward penalizes deviations from the shortest correct rollout in a batch of 8. Experiments on four math benchmarks (MATH-500, AMC, AIME 2024, AIME 2025) report reduced lengths and preserved or improved accuracy.
Significance. If the results hold, the paper makes a useful contribution: it provides evidence that reasoning over-length is not a fixed cost, and it shows two lightweight interventions that reduce tokens by roughly 20-40% with little accuracy change across diverse model families. The breadth of the evaluation (seven backbones, four benchmarks, both training-free and RL methods) and the use of external evaluation sets are strengths. The main caveats are the unresolved positional transfer of the steering vector, the unreported per-model hyperparameters (lambda, layer, RL checkpoint), and the absence of variance estimates; these currently prevent the strong causal interpretation the paper advances. The work is likely to be of interest to the efficient-reasoning community, and the proposed methods are simple enough to be reproduced once the missing details are supplied.
major comments (5)
- [§4.2, Eqs. (2)-(4), and §4.3, Eq. (5)] The steering direction v_l is estimated from residual-stream activations at the final reasoning token, but the intervention is applied to the hidden state of the last input token only. The paper does not demonstrate that a direction estimated at the end of a long reasoning trace transfers to the final prompt-token position. This is load-bearing because if the direction does not transfer, the reported length reductions could be generic perturbation effects rather than evidence for an efficiency direction. Please add a transfer analysis, for example cosine similarity between v_l and a difference-in-means vector computed at the last input token using the same short/long trajectories, or an intervention control that applies a direction computed at the input position, and report the result.
- [§4.3, Implementation Details] The optimal steering coefficient lambda and the target layer l are not reported for any of the seven models. Since the results in Table 2 depend on these per-model choices, and Figure 5 shows that the effect of steering varies with lambda, the paper should report the exact lambda and layer used for each model, the selection criterion, and ideally a sensitivity analysis over lambda and layer.
- [Table 3 and §5.3-§5.5] The checkpoint at which the RL-trained models are evaluated is not specified. Section 5.5 shows substantial variation in accuracy across training steps (e.g., AIME 2025 drops from 39% to 32% between steps 30 and 70), so the choice of checkpoint is load-bearing for the claim that performance is preserved. State the exact number of training steps (or the stopping rule) used for Table 3, and report evaluation at multiple checkpoints.
- [Tables 2 and 3] The paper reports point estimates for accuracy and length without variance, confidence intervals, or significance tests. Some improvements are small (e.g., GLM-Z1-32B on MATH-500: 96.00 to 96.20), so the terms 'significant' and 'improving' are not supported by the evidence. Report standard errors (e.g., bootstrap over test questions) and test the changes statistically, at least for the main claims.
- [§3.1, Figure 2] The claim that efficient and verbose reasoning paths are 'linearly separable' or exhibit a 'pronounced shift' is supported only by qualitative PCA scatter plots. Provide quantitative separation metrics (e.g., linear probe accuracy, or the ratio of between-class to within-class variance) for the layers and difficulty levels shown, and if possible for all models.
minor comments (7)
- [§4.3 Results and Analysis] The text says 'across four benchmarks and nine LRMs' and the Table 2 caption says 'nine LRMs', but the evaluation covers seven models; correct the count.
- [§4.3 Results and Analysis] The text states R1-Distill-Qwen-14B improves on AIME 2025 from 44.17 to 44.57, but Table 2 reports 42.50 to 44.17; the numbers are inconsistent.
- [§5.4 Results] The text claims R1-Distill-Qwen-7B achieves a performance gain on AMC from 79.97 to 81.63, but Table 3 reports 79.97 to 79.07; align the text with the table.
- [§5.5, Figure 6] The caption for panel (b) is labeled 'Average reasoning length over training epochs' but the x-axis is training steps; use consistent terminology.
- [§3.2] The sentence 'In Table 3, we observe...' should refer to Figure 3, which contains the lexical analysis table; fix the cross-reference.
- [§5.1, Eqs. (6)-(7)] The handling of batches in which no response is correct is described in prose but not formalized; specify the exact reward for that case, since Eq. (7) is then undefined.
- [Related Works] The claim that the methods are 'entirely free of external labels' should be tempered, since ground-truth answers are used both to select the steering trajectories and to compute the RL reward.
Circularity Check
No significant circularity: the self-referential design is explicit and the central results are evaluated on external benchmarks.
full rationale
The paper's derivation chain is empirical rather than definitional. Section 2 establishes latent efficiency by sampling multiple candidate responses per problem and comparing shortest and longest correct paths; this is an observation, not a construction. Section 3.1 computes difference-in-means directions (Eqs. 2-4) from those paths, and Section 4 applies them as steering (Eq. 5) with evaluation on held-out benchmarks (DeepMath, MATH-500, AMC, AIME). The steering vector is derived from GSM8K correct paths, not from the test sets, and the length and accuracy outcomes in Tables 2 and 3 are measured results rather than consequences of the equations by construction. Section 5's reward (Eq. 6) explicitly contains a brevity penalty, so length reduction is the intended objective; what is empirical is whether accuracy is preserved and whether the behavior generalizes. The self-referential character (using the model's own activations and rollouts) is the announced design principle, not hidden circularity. Self-citations such as Zhao et al. 2025a and 2025b appear in related-work or motivation contexts and are not load-bearing for the main claims. The notable concerns—unreported layer choice, empirical selection of lambda, and the transfer of a direction estimated at the final reasoning token to the last input token—are experimental validity issues, not circular reductions, and cannot be scored as circularity without further evidence.
Assumptions & free parameters
free parameters (6)
- Steering coefficient lambda (per model) =
not reported
- Steering target layer =
not reported
- RL training checkpoint / number of steps =
not reported
- Reward weights lambda1, lambda2 =
lambda1=1, lambda2=0.001
- Number of rollouts for min/max analysis =
8
- KL penalty coefficient and learning rate =
0.01, 1e-6
assumptions (5)
- domain assumption The linear representation hypothesis: reasoning efficiency is encoded as a single direction in the residual stream.
- domain assumption Intervening on only the final input token's hidden state at one layer is sufficient to influence the entire generated reasoning trace.
- domain assumption The shortest correct response in a batch of 8 is a meaningful efficiency target.
- domain assumption MATH/GSM8K ground-truth answers are correct and matching the final answer implies correct reasoning.
- standard math Standard transformer residual stream update equations.
invented entities (1)
-
Efficiency direction vector v_l
independent evidence
Cite this review
Pith. "Pith review of Exploring and Exploiting the Inherent Efficiency within Large Reasoning Models for Self-Guided Efficiency Enhancement." pith.science (2026). https://pith.science/paper/D7NZ5RWL
@misc{pith2026250615647,
author = {Pith},
title = {Pith review of: Exploring and Exploiting the Inherent Efficiency within Large Reasoning Models for Self-Guided Efficiency Enhancement},
year = {2026},
howpublished = {\url{https://pith.science/paper/D7NZ5RWL}},
note = {Machine review of arXiv:2506.15647}
}
read the original abstract
Recent advancements in large reasoning models (LRMs) have significantly enhanced language models' capabilities in complex problem-solving by emulating human-like deliberative thinking. However, these models often exhibit overthinking (i.e., the generation of unnecessarily verbose and redundant content), which hinders efficiency and inflates inference cost. In this work, we explore the representational and behavioral origins of this inefficiency, revealing that LRMs inherently possess the capacity for more concise reasoning. Empirical analyses show that correct reasoning paths vary significantly in length, and the shortest correct responses often suffice, indicating untapped efficiency potential. Exploiting these findings, we propose two lightweight methods to enhance LRM efficiency. First, we introduce Efficiency Steering, a training-free activation steering technique that modulates reasoning behavior via a single direction in the model's representation space. Second, we develop Self-Rewarded Efficiency RL, a reinforcement learning framework that dynamically balances task accuracy and brevity by rewarding concise correct solutions. Extensive experiments on seven LRM backbones across multiple mathematical reasoning benchmarks demonstrate that our methods significantly reduce reasoning length while preserving or improving task performance. Our results highlight that reasoning efficiency can be improved by leveraging and guiding the intrinsic capabilities of existing models in a self-guided manner.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[1]
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,
-
[3]
Language models are few-shot learners
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901,
1901
-
[6]
An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115,
-
[7]
From system 1 to system 2: A survey of reasoning large language models
Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, et al. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419,
-
[8]
Towards large reasoning models: A survey of reinforced reasoning with large language models
Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, et al. Towards large reasoning models: A survey of reinforced reasoning with large language models. arXiv preprint arXiv:2501.09686,
-
[9]
Towards reasoning era: A survey of long chain-of- thought for reasoning large language models
Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wangxiang Che. Towards reasoning era: A survey of long chain-of- thought for reasoning large language models. arXiv preprint arXiv:2503.09567, 2025a. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu...
-
[10]
15 Marthe Ballon, Andres Algaba, and Vincent Ginis. The relationship between reasoning and performance in large language models–o3 (mini) thinks harder, not longer. arXiv preprint arXiv:2502.15631,
-
[11]
Team Kimi, 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,
Show all 33 references
-
[12]
Training language models to reason efficiently
Daman Arora and Andrea Zanette. Training language models to reason efficiently. arXiv preprint arXiv:2502.04463,
-
[13]
L1: Controlling how long a reasoning model thinks with reinforcement learning
Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697,
-
[14]
Optimizing test-time compute via meta reinforcement fine- tuning
Yuxiao Qu, Matthew YR Yang, Amrith Setlur, Lewis Tunstall, Edward Emanuel Beeching, Ruslan Salakhutdinov, and Aviral Kumar. Optimizing test-time compute via meta reinforcement fine- tuning. arXiv preprint arXiv:2503.07572, 2025a. Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, K...
-
[15]
Cot-valve: Length- compressible chain-of-thought tuning
Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. Cot-valve: Length- compressible chain-of-thought tuning. arXiv preprint arXiv:2502.09601,
-
[16]
Self-training elicits concise reasoning in large language models
Tergel Munkhbat, Namgyu Ho, Seo Hyun Kim, Yongjin Yang, Yujin Kim, and Se-Young Yun. Self-training elicits concise reasoning in large language models. arXiv preprint arXiv:2502.20122,
-
[18]
Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue
Notion Blog. Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. Demystifying long chain-of-thought reasoning in llms. arXiv preprint arXiv:2502.03373,
-
[19]
Seal: Steerable reasoning calibration of large language models for free
Runjin Chen, Zhenyu Zhang, Junyuan Hong, Souvik Kundu, and Zhangyang Wang. Seal: Steerable reasoning calibration of large language models for free. arXiv preprint arXiv:2504.07986, 2025b. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł...
-
[22]
Detox: Toxic subspace projection for model editing
Rheeya Uppaal, Apratim De, Yiting He, Yiquao Zhong, and Junjie Hu. Detox: Toxic subspace projection for model editing. arXiv preprint arXiv:2405.13967,
-
[23]
Adasteer: Your aligned llm is inherently an adaptive jailbreak defender
Weixiang Zhao, Jiahe Guo, Yulin Hu, Yang Deng, An Zhang, Xingyu Sui, Xinyang Han, Yanyan Zhao, Bing Qin, Tat-Seng Chua, et al. Adasteer: Your aligned llm is inherently an adaptive jailbreak defender. arXiv preprint arXiv:2504.09466, 2025a. Nora Belrose. Diff-in-means concept e...
-
[24]
The geometry of truth: Emergent linear structure in large language model representations of true/false datasets
Samuel Marks and Max Tegmark. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824,
-
[25]
Linear representations of sentiment in large language models
Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154,
-
[26]
Refusal in language models is mediated by a single direction
17 Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717,
-
[27]
Training verifiers to solve math word problems
Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168,
-
[28]
Deepmath-103k: A large-scale, challenging, de- contaminated, and verifiable mathematical dataset for advancing reasoning
Zhiwei He, Tian Liang, Jiahao Xu, Qiuzhi Liu, Xingyu Chen, Yue Wang, Linfeng Song, Dian Yu, Zhenwen Liang, Wenxuan Wang, et al. Deepmath-103k: A large-scale, challenging, de- contaminated, and verifiable mathematical dataset for advancing reasoning. arXiv preprint arXiv:2504.11456,
-
[29]
Deepseekmath: Pushing the limits of mathematical reasoning in open language models
Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300,
-
[30]
Openrlhf: An easy-to-use, scalable and high-performance rlhf framework
Jian Hu, Xibin Wu, Zilin Zhu, Xianyu, Weixun Wang, Dehao Zhang, and Yu Cao. Openrlhf: An easy-to-use, scalable and high-performance rlhf framework. arXiv preprint arXiv:2405.11143,
-
[31]
Harnessing the reasoning economy: A survey of efficient reasoning for large language models
Rui Wang, Hongru Wang, Boyang Xue, Jianhui Pang, Shudong Liu, Yi Chen, Jiahao Qiu, Derek Fai Wong, Heng Ji, and Kam-Fai Wong. Harnessing the reasoning economy: A survey of efficient reasoning for large language models. arXiv preprint arXiv:2503.24377,
-
[32]
A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond
Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, et al. A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond. arXiv preprint arXiv:2503.21614, 2025b. Sicheng Fen...
-
[33]
Trade-offs in large reasoning models: An empirical analysis of deliberative and adaptive reasoning over foundational capabilities
Weixiang Zhao, Xingyu Sui, Jiahe Guo, Yulin Hu, Yang Deng, Yanyan Zhao, Bing Qin, Wanx- iang Che, Tat-Seng Chua, and Ting Liu. Trade-offs in large reasoning models: An empirical analysis of deliberative and adaptive reasoning over foundational capabilities. arXiv preprint arXi...
-
[2017]
Linguistic regularities in continuous space word representations
Tomáš Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word representations. In Proceedings of the 2013 conference of the north american chapter of the association for computational linguistics: Human language technologies, pages 746–751,
2013
-
[2020]
The llama 3 herd of models
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783,
-
[2021]
Reasoning models know when they’re right: Probing hidden states for self-verification
16 Anqi Zhang, Yulin Chen, Jane Pan, Chen Zhao, Aurojit Panda, Jinyang Li, and He He. Reasoning models know when they’re right: Probing hidden states for self-verification. arXiv preprint arXiv:2504.05419,
-
[2023]
Representation engineering: A top-down approach to ai transparency
Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405,
-
[2024]
Gemma 2: Improving open language models at a practical size
Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118,
-
[2025]
Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al
URL https://openai.com/index/ o3-mini-system-card . 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 ...
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.