Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Resa: Transparent Reasoning Models via SAEs

T0 review · 5 major / 6 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 SAE-Tuning elicits 1.5B-model reasoning from question-answer pairs alone, matching RL-trained models for roughly $1.

desk verdict A cheap, open, and potentially useful recipe for CoT-free reasoning elicitation whose headline mechanistic claim—SAE feature transfer—is not actually tested by the experiments as designed. read the letter →

arxiv 2506.09967 v2 pith:2DROM234 submitted 2025-06-11 cs.CL cs.AIcs.LG

Shangshang Wang , Julian Asilis , Ömer Faruk Akgül , Enes Burak Bilgin , Ollie Liu , Deqing Fu , Willie Neiswanger This is my paper
classification cs.CLcs.AIcs.LG
keywords sparseautoencodersSAE-Tuningreasoningelicitationchain-of-thought-freesupervisionlow-rankadaptersKLdivergencedistillationfeatureextractioninterpretability
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 strong reasoning can be elicited from a small base language model without reinforcement learning and without chain-of-thought traces, by first training a sparse autoencoder on verified question-answer pairs and then using that frozen autoencoder to guide supervised fine-tuning. The claimed payoff is cost and transparency: a 1.5B model reaches about the same average score as its RL-trained counterpart for roughly $1 and 20 minutes of training. The authors also argue that the extracted reasoning ability generalizes across datasets, transfers across models in the same family as a portable adapter, and leaves a layer-wise feature distribution that predicts reasoning performance.

What carries the argument

The load-bearing object is the sparse autoencoder (SAE) inserted into the target model at a fixed layer, combined with the $\mathrm{KL}$-divergence objective. The SAE is a dictionary of 65,536 features trained with a Top-$k$ encoder ($k=32$) to reconstruct the source model's layer-12 MLP activations; freezing it and requiring the target model's output distribution to match with and without its insertion forces the target's LoRA parameters to become compatible with features that encode reasoning. The <think>...</think> wrapper around a bare final answer is the trigger hypothesized to activate those features, and the trained SAE is removed at test time, leaving the reasoning ability in the model's own parameters.

What would settle it

Train the identical LoRA setup with the same $\mathrm{KL}$ objective but replace the frozen SAE with a randomly initialized, frozen autoencoder of the same shape, or with an identity map; if the resulting model matches the reported gains, the reasoning-feature transfer story is wrong, while a collapse toward the 39% CoT-free SFT baseline would support the claim that the SAE's learned features are doing the work.

Watch

Extended reading notes

Core claim

SAE-Tuning transfers reasoning ability from a source model to a target model using only CoT-free verified question-answer data. In Stage I, an SAE is trained on activations of a chosen source layer, capturing latent features triggered by a template that places the final answer inside <think> tokens. In Stage II, the frozen SAE is inserted after the same layer of the target model and low-rank adapters are trained to minimize the $\mathrm{KL}$ divergence between the target's output distribution with and without the SAE; after training, the SAE is removed. The central result is that this procedure applied end-to-end to the base R1-Distill model as its own source (Resa-STILL-v5) scores 48.06% average on six reasoning benchmarks, essentially matching the RL-trained Tina-STILL at 48.16%, while a lightly RL-trained source yields peaks such as 49.46%.

Load-bearing premise

The mechanism assumes that writing a bare final answer inside <think> tokens activates the source model's latent reasoning features, and that matching the output distribution with a frozen SAE inserted transfers those features into the target's LoRA parameters; no experiment isolates whether gains come from feature transfer rather than from the $\mathrm{KL}$ objective acting as a regularizer.

Editorial extensions

If this is right

  • Standard SFT on the same CoT-free data without the SAE scores only 39.00% average, so the SAE is presented as the active ingredient rather than the answer-only template alone.
  • Training an SAE from scratch on the trigger dataset is about as effective as fine-tuning a pre-trained SAE, so SAE-Tuning needs no pre-existing SAE and no expensive SAE pre-training corpus.
  • Reasoning features extracted from the STILL dataset transfer to other datasets (DeepScaleR, Open-S1, II-Thought, OpenR1), reaching performance on par with models trained via RL on those datasets.
  • Adapters trained on Qwen or Qwen-Math can be attached to R1-Distill at test time without retraining, yielding scores around 47.5-47.9%, supporting the decomposition strong reasoning equals abstract reasoning ability plus foundational knowledge.
  • The layer-wise count of reasoning features forms a tri-modal distribution whose Gaussian mixture components align with the Gaussian mixture of achieved reasoning performance, offering a data-driven heuristic for choosing the SAE hookpoint layer.

Reading between the lines

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

  • The $\mathrm{KL}$ objective itself could be acting as a regularizer; a decisive control would replace the frozen SAE with a randomly initialized frozen autoencoder of the same shape, or with an identity map, and check whether the reported gains survive.
  • The reasoning-as-adapter result suggests that reasoning ability and factual knowledge may be composable independently; a natural extension is testing whether the same adapter transfers to larger models or to models outside the architecture family when the SAE dimensions can be matched.
  • The tri-modal GMM correlation is an aggregate pattern, and the paper notes single-layer feature counts do not predict performance; a predictive test would use the identified layer-clusters to select hookpoints on held-out data and measure whether the resulting Resa models outperform layer-choice by heuristic.
  • If the mechanism is genuinely about surfacing latent reasoning features, SAE-Tuning could plausibly extend to other verifiable skills beyond math and science, such as code correctness or instruction following, using the same answer-only data format.
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 / 6 minor

Summary. The paper introduces SAE-Tuning, a two-stage method for eliciting reasoning abilities in 1.5B language models. Stage I trains a sparse autoencoder (SAE) on the activations of a source model at a chosen layer, using verified chain-of-thought-free (CoT-free) question-answer data formatted with a <think>[answer]</think> template. Stage II freezes this SAE, inserts it into a target model, and trains LoRA adapters to minimize the KL divergence (Eq. 8) between the target's output distributions with and without the SAE, after which the SAE is discarded. The paper claims that SAE-Tuning matches or approaches RL-trained counterparts on six reasoning benchmarks at roughly $1 and under 20 minutes of training, that the extracted reasoning abilities generalize across datasets and transfer as modular adapters across models in the Qwen family, and that layer-wise counts of SAE features activated at think tokens align with layer-wise reasoning performance, providing transparency.

Significance. If the empirical results hold, the cost and data-efficiency claims are significant for democratizing reasoning-model research, and the open release of code, weights, and training logs is a concrete strength. The modular-adapter result, where LoRA adapters trained on Qwen or Qwen-Math transfer to R1-Distill without retraining, is a potentially useful finding for compositional reuse. However, the interpretive layer of the paper—feature transfer and transparency—is currently under-supported: the mechanism in Eq. (8) is not isolated by controls, and the Section 5 GMM analysis does not establish prediction. The headline quantitative claims are also built on small benchmarks, best-checkpoint selection, and a cost baseline that is not matched to the direct RL counterpart. The work will interest the reasoning and interpretability communities, but the evidence as presented supports the empirical method more strongly than it supports the feature-transfer and transparency interpretations.

major comments (5)
  1. [Section 3.2, Figure 1, Table 2] The headline efficiency gains (">2000x cost, >450x time") are computed against STILL-3-1.5B-Preview and DeepScaleR-1.5B-Preview, not against the RL-trained counterparts (Tina-STILL, Tina-DeepScaleR) to which the performance comparison is actually made. By Figure 1's own numbers, Tina-STILL costs $60 and 60 hours, so Resa-STILL-v5's $1 and 0.33 hours represent roughly a 60x cost reduction and a 180x time reduction against the matched counterpart, not the >2000x and >450x stated in the Abstract. Please report the efficiency comparison against the direct RL counterpart and qualify the headline claims accordingly.
  2. [Section 2.1, Eq. (8), Table 4] The central mechanistic claim that Stage II transfers learned reasoning features is untested. The only non-SAE control (STILL-CoT-free-SFT, Table 3, 39.00 avg) differs from SAE-Tuning in both the presence of the SAE and the loss function (answer-label cross-entropy vs. the KL divergence in Eq. 8). Moreover, Resa-STILL-v2 with a pretrained SAE that never saw the trigger data reaches 44.99 avg, capturing 3.07 of the 9.06-point gap over SFT. Without controls using a randomly initialized SAE, an identity-plus-noise perturbation with matched reconstruction error, or a pure KL-consistency baseline without any SAE, the observed gains could be attributed to a consistency regularizer rather than to the SAE's learned reasoning features. Please add such controls or substantially soften the feature-transfer interpretation.
  3. [Section 5, Table 6, Figure 3] The transparency analysis fits a 3-component GMM to the 26 layer-wise feature counts and to the 26 layer-wise performance scores, then reports that the fitted means and weights align. This does not establish prediction. Both distributions are indexed by the same ordered layer indices, so similar component locations are expected for any smooth trend; the feature counts are small integers (0-5), the performance scores are means of six small benchmarks with no uncertainty, and a 3-component GMM has roughly eight free parameters on 26 points. The text also inconsistently states that the feature-count GMM is fit to the base model's counts (Section 5) while Table 6's caption says the counts are for Tina-STILL. Please specify the exact input data, provide a permutation test or held-out layers to show the alignment exceeds chance, and give bootstrap confidence intervals on the GMM means. As written, the claim that the structure is a robust predictor is not supported.
  4. [Section 3.2, Tables 3-4, Appendix B] All reported Resa scores are best-checkpoint selections, and checkpoint-to-checkpoint variation is large (e.g., Resa-DeepScaleR-v1 ranges from 40.87 to 48.95 across checkpoints in Appendix Table 9). The headline comparison of 48.06 vs. 48.16 is within the sampling noise of the 30-problem AIME sets, and no seeds or error bars are reported. In addition, the Abstract mixes configurations: the 43.33 AIME24 and 90 AMC23 examples come from Resa-STILL-Tina-50-step (Trained-from-Scratch), which uses a lightly RL-trained source, not the base-model source used for the >97% retention claim. Please specify a fixed checkpoint-selection rule, report bootstrap confidence intervals or per-benchmark standard errors, and align each Abstract claim with the exact configuration in Table 4.
  5. [Section 2.1, Section 5] The paper states in Section 2.1 that the <think>[Answer]</think> template is hypothesized to activate the source model's latent reasoning abilities and promises a detailed analysis in Section 5, but Section 5 does not ablate the template in the training pipeline; it only counts SAE features that fire on think tokens in a separate prompt. The role of the thinking tokens is therefore untested, which matters because the method's CoT-free claim depends on this design choice. Please include an ablation that removes or alters the think tokens in the trigger/elicitation data, or explicitly acknowledge that the template's contribution is not isolated.
minor comments (6)
  1. [Eq. (3)] Equation (3) contains a grammatical error: "we only the top k features" should read "we keep only the top k features."
  2. [Section 2.1] The sentence "In the following, We now formalize" has a capitalization and duplication issue; please rephrase.
  3. [Table 4] The column header "GPQAMinerva" in Table 4 is missing a space; it should read "GPQA" and "Minerva" as separate columns.
  4. [Figure 1] Figure 1 is difficult to parse as rendered, with unclear axis labeling and bar values; please redraw with explicit units and legible value labels.
  5. [References] The reference "Intelligent Internet. II-Thought: A large-scale, high-quality reasoning dataset, 2025" and the in-text citation "Internet, 2025" are incomplete; please provide the full dataset documentation or a URL.
  6. [Section 5] The phrase "we cut off the first and final layer" is informal; consider "we exclude the first and final layers."

Circularity Check

1 steps flagged · score 4.0 of 10

Main SAE-Tuning results are externally benchmarked and self-contained; only the Section 5 transparency claim is partially circular, where a performance GMM is fit to the same 26 models whose performance the feature-count structure is said to predict.

  1. fitted input called prediction [Section 5, 'Feature Counts v.s. Reasoning Performance Correlation' (after Table 6)]
    "we fit a 3-component Gaussian Mixture Model (3-GMM) to both distributions: (1) the a priori reasoning feature counts from the base model, and (2) the final reasoning scores from our 26 Resa models. ... This suggests that ... the overall structure of how reasoning is organized into three distinct layer-clusters within the model is a robust predictor of how performance will be distributed."

    The performance GMM is fitted to the same 26 Resa-model average scores that the feature-count structure is claimed to predict. The reported 'close structural alignment' is therefore an in-sample comparison of two fitted distributions, not an out-of-sample prediction. Calling the feature distribution a 'robust predictor' and recommending layer-cluster selection based on it is a post hoc fit to the outcome: the predicted quantity (the performance distribution) was used to fit the very GMM whose alignment is then offered as evidence. No held-out validation is provided, so the 'prediction' reduces to a descriptive comparison of fits on the same dataset.

full rationale

The paper's central performance claims (Tables 3, 4, and 5) are benchmarked against external RL-trained baselines and held-out reasoning suites (AIME, AMC, MATH500, GPQA, Minerva); these results do not reduce to the method's inputs by construction. The Stage II objective (Eq. 8) is a KL-consistency loss between the target model's own distributions with and without a frozen SAE, and the absence of a null-SAE KL control means the mechanism is not cleanly identified; however, this is a correctness/identification weakness, not an equation-level circularity. The one genuinely circular element is the Section 5 transparency analysis: a 3-GMM is fit to the 26 layer-wise performance scores, and then the feature-count distribution is said to 'predict' that same performance distribution. Because the performance GMM is fit to the outcome it is later claimed to be predicted by, the structural alignment is an in-sample post hoc comparison rather than a validated predictor. This partial circularity affects the transparency contribution, while the efficient-elicitation and modularity results retain independent external-benchmark support, yielding a moderate overall score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central efficiency claim rests on several free choices (layer 12, best checkpoint selection, 3-GMM fit) and unproven domain assumptions about what the SAE actually transfers. The main invented construct, 'reasoning features', has no independent validation.

free parameters (3)
  • SAE hookpoint layer = 12 (of 28)
    Chosen by heuristic that middle layers are crucial for reasoning; the layer ablation in Table 6 shows layers 15-18 perform better, so this default is not optimized and affects results.
  • Best checkpoint selection per model = varies, e.g., Resa-STILL-v1 step 2000, Resa-DeepScaleR-v1 step 2500
    Headline results report the peak checkpoint across 4-10 evaluated steps, without seed averaging; this inflates reported performance and is a free choice affecting the central comparison.
  • 3-GMM component parameters = means 4.9/14.5/22.7 and 5.6/15.1/23.0; weights 41/37/22% and 39/37/24%
    Fitted to the layer-wise feature counts and the 26 Resa layer-performance scores, then compared; this is a post hoc fit, not a held-out prediction, and the number of components is assumed.
assumptions (4)
  • domain assumption Reasoning abilities are already latent in the base model and can be elicited without new knowledge.
    This 'elicitation hypothesis' is taken from prior work (Zhao et al. 2025, Wang et al. 2025b) and is the premise for expecting SAE-Tuning to surface reasoning without CoT traces (Section 3.2).
  • ad hoc to paper The KL divergence objective, training the target model to match its own output distribution with and without a frozen SAE, transfers source reasoning features into the target's LoRA parameters.
    This is the core mechanism of SAE-Tuning (Eq. 8). The paper provides control experiments (standard SFT fails) but no mechanistic evidence that feature transfer, rather than regularization, is what improves performance.
  • ad hoc to paper Inserting a final answer between <think> and </think> tokens triggers the source model's latent reasoning features.
    Section 2.1 hypothesizes the template activates latent reasoning; no evidence distinguishes token-specific features from reasoning features.
  • ad hoc to paper A 3-component Gaussian Mixture Model is the right structural model for both feature counts and performance across layers.
    Section 5 fits a 3-GMM to both distributions without justifying the number of components or testing alternative models; the 'close alignment' result depends on this choice.
invented entities (1)
  • 'Reasoning features' as SAE features activated exclusively at <think> and </think> tokens
    purpose: Operationalize latent reasoning ability for the transparency analysis.
    No external validation that these features are reasoning-specific; Table 6 shows the best-performing layer (18) has 0 such features, undermining the construct's predictive value.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Resa: Transparent Reasoning Models via SAEs." pith.science (2026). https://pith.science/paper/2DROM234

@misc{pith2026250609967,
  author       = {Pith},
  title        = {Pith review of: Resa: Transparent Reasoning Models via SAEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DROM234}},
  note         = {Machine review of arXiv:2506.09967}
}
abstract

How cost-effectively can we elicit strong reasoning in language models by leveraging their underlying representations? We answer this question with Resa, a family of 1.5B reasoning models trained via a novel and efficient sparse autoencoder tuning (SAE-Tuning) procedure. This method first trains an SAE to capture reasoning abilities from a source model, and then uses the trained SAE to guide a standard supervised fine-tuning process to elicit such abilities in a target model, all using verified question-answer data without any reasoning traces. Notably, when applied to certain base models before further RL post-training, SAE-Tuning retains >97% of its RL-trained counterpart's reasoning performance while reducing training costs by >2000x to roughly \$1 and training time by >450x to around 20 minutes. Furthermore, when applied to lightly RL-trained models (e.g., within 1 hour on 2 GPUs), it enables reasoning performance such as 43.33% Pass@1 on AIME24 and 90% Pass@1 on AMC23 for only around \$1 additional cost. Surprisingly, the reasoning abilities extracted via SAEs are potentially both generalizable and modular. Generality means abilities extracted from one dataset still elevate performance on a larger and overlapping corpus. Modularity means abilities extracted from Qwen or Qwen-Math can be attached to the R1-Distill model at test time, without any retraining, and yield comparable gains. Extensive ablations validate these findings and all artifacts are fully open-sourced.

Figures

Figures reproduced from arXiv: 2506.09967 by the authors.

Figure 1
Figure 1. Comparison between Example Resa Models and Baselines [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Two-Stage Pipeline of SAE-Tuning The procedure begins with SAE training (Left), where an SAE is trained to capture reasoning features from a source model with a trigger dataset. During SAE-guided SFT (Right), the trained SAE is then frozen and inserted into a target model. An elicitation dataset is used to guide a SFT process to elicit the reasoning abilities in the target model. Notably, the trigger and elicitation… view at source ↗
Figure 3
Figure 3. Reasoning Feature Extraction (Left) This shows the layer-wise feature counts of the base R1-Distill model. (Middle) This shows the layer-wise feature counts of the Tina-STILL model. (Right) This shows the reasoning performance of the trained Resa models with different layer-wise SAEs when Tina-STILL is the source model. Model Name AIME24 AIME25 AMC23 MATH500 GPQA Minerva Avg. Feature Cnts. Resa-STILL-2nd-Layer 26.67… view at source ↗

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. From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning

    cs.LG 2026-06 unverdicted novelty 6.5 of 10

    SFT supplies entangled compositional traces of atomic skills and routing modules; RL identifies those modules and enables recombination on novel compositions outside the SFT support.

Reference graph

Works this paper leans on

29 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [3]

    DanielBeaglehole, AdityanarayananRadhakrishnan, EnricBoix-Adserà, andMikhailBelkin

    URLhttps://arxiv.org/abs/2503.00177. DanielBeaglehole, AdityanarayananRadhakrishnan, EnricBoix-Adserà, andMikhailBelkin. Towarduniversal steering and monitoring of ai models,

    arXiv

  2. [4]

    Matthew Chen, Joshua Engels, and Max Tegmark

    URLhttps://arxiv.org/abs/2502.03708. Matthew Chen, Joshua Engels, and Max Tegmark. Low-rank adapting models for sparse autoencoders,

    arXiv

  3. [5]

    Low-Rank Adapting Models for Sparse Autoencoders

    URLhttps://arxiv.org/abs/2501.19406. Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V. Le, Sergey Levine, and Yi Ma. Sft memorizes, rl generalizes: A comparative study of foundation model post-training,

    work page Pith review arXiv

  4. [6]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey

    URLhttps://arxiv.org/abs/2501.17161. Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models,

    arXiv

  5. [8]

    DeepSeek-AI

    URLhttps://arxiv.org/abs/2503.16219. DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning,

    arXiv

  6. [9]

    Clémentine Fourrier, Nathan Habib, Hynek Kydlíček, Thomas Wolf, and Lewis Tunstall

    URLhttps://arxiv.org/abs/2501.12948. Clémentine Fourrier, Nathan Habib, Hynek Kydlíček, Thomas Wolf, and Lewis Tunstall. Lighteval: A lightweight framework for llm evaluation,

    arXiv

  7. [10]

    15 Resa: Transparent Reasoning Models via SAEs Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu

    URLhttps://arxiv.org/abs/2505.14071. 15 Resa: Transparent Reasoning Models via SAEs Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders,

    arXiv

  8. [12]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt

    URLhttps://arxiv.org/abs/2506.04178. 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,

    arXiv

Show all 29 references
  1. [14]

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E

    URLhttps://arxiv.org/abs/2503.09532. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of Symposium on O...

    arXiv

  2. [15]

    URL https://arxiv.org/abs/2501.19393. 16 Resa: Transparent Reasoning Models via SAEs Kyle O’Brien, David Majercak, Xavier Fernandes, Richard Edgar, Blake Bullwinkel, Jingya Chen, Harsha Nori, Dean Carignan, Eric Horvitz, and Forough Poursabzi-Sangde. Steering language model re...

    arXiv

  3. [16]

    Charles O’Neill and Thang Bui

    URLhttps://arxiv.org/abs/2411.11296. Charles O’Neill and Thang Bui. Sparse autoencoders enable scalable and reliable circuit identification in language models,

    arXiv

  4. [17]

    NinaPanickssery,NickGabrieli,JulianSchulz,MegTong, EvanHubinger,andAlexanderMattTurner

    URLhttps://arxiv.org/abs/2405.12522. NinaPanickssery,NickGabrieli,JulianSchulz,MegTong, EvanHubinger,andAlexanderMattTurner. Steering Llama 2 via contrastive activation addition,

    arXiv

  5. [18]

    URLhttps://arxiv.org/abs/2312.06681. Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, ...

    arXiv

  6. [19]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman

    URLhttps://arxiv.org/abs/2412.15115. 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. InProceedings of Conference on Language Modeling (COLM),

    arXiv

  7. [20]

    Shangshang Wang and Willie Neiswanger

    URLhttps://arxiv.org/abs/2402.03300. Shangshang Wang and Willie Neiswanger. LLM reasoning: Curated insights,

    arXiv

  8. [21]

    Shangshang Wang, Julian Asilis, Ömer Faruk Akgül, Enes Burak Bilgin, Ollie Liu, and Willie Neiswanger

    URL https: //shangshangwang.notion.site/llm-reasoning. Shangshang Wang, Julian Asilis, Ömer Faruk Akgül, Enes Burak Bilgin, Ollie Liu, and Willie Neiswanger. Tina: Tiny reasoning models via lora, 2025a. URLhttps://arxiv.org/abs/2504.15777. Yiping Wang, Qing Yang, Zhiyuan Zeng,...

    arXiv

  9. [22]

    URLhttps://arxiv.org/abs/2501.17148. 17 Resa: Transparent Reasoning Models via SAEs Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, Chenyang Shao, Yuwei Yan, Qinglong Yang, Yiwen Song, Sijia...

    arXiv

  10. [23]

    URLhttps://arxiv.org/abs/2501.09686. An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and Zhenru Zhang. Qwen2.5-math technical report: To...

    arXiv

  11. [24]

    Rosie Zhao, Alexandru Meterez, Sham Kakade, Cengiz Pehlevan, Samy Jelassi, and Eran Malach

    URLhttps://arxiv.org/abs/2409.12122. Rosie Zhao, Alexandru Meterez, Sham Kakade, Cengiz Pehlevan, Samy Jelassi, and Eran Malach. Echo chamber: Rl post-training amplifies behaviors learned in pretraining,

    arXiv

  12. [25]

    18 Resa: Transparent Reasoning Models via SAEs Appendix A

    URLhttps://arxiv.org/ abs/2504.07912. 18 Resa: Transparent Reasoning Models via SAEs Appendix A. Full Hyperparameter We show our default choice of hyperparameter in Table

    arXiv

  13. [26]

    Checkpoint Steps AIME24 AIME25 AMC23 MATH500 GPQA Minerva Avg. 1000 20.00 33.33 75.00 82.60 29.78 33.33 45.67 1500 33.33 23.33 75.00 82.80 30.88 30.81 46.03 2000 33.33 33.33 75.00 83.80 29.41 28.79 47.28 2500 30.00 23.33 77.50 84.20 26.47 33.33 45.81 Table 8: Performance of Re...

  14. [27]

    Checkpoint Steps AIME24 AIME25 AMC23 MATH500 GPQA Minerva Avg. 1000 23.33 23.33 72.50 85.40 30.51 34.85 44.99 1500 20.00 20.00 75.00 81.40 29.78 30.30 42.75 2000 23.33 23.33 67.50 83.60 29.78 35.86 43.90 2500 23.33 26.67 67.50 83.00 25.74 34.34 43.43 Table 12: Performance of R...

    2000

  15. [28]

    Checkpoint Steps AIME24 AIME25 AMC23 MATH500 GPQA Minerva Avg. 500 33.33 30.00 80.00 84.00 29.41 35.86 48.77 1000 33.33 23.33 70.00 84.80 29.41 32.83 45.62 1500 26.67 13.33 67.50 83.80 31.25 36.87 43.24 2000 26.67 23.33 70.00 83.20 28.68 34.34 44.37 2500 33.33 16.67 70.00 82.4...

    2000

  16. [29]

    Checkpoint Steps AIME24 AIME25 AMC23 MATH500 GPQA Minerva Avg. 1000 26.67 20.00 70.00 85.00 29.41 32.32 43.90 1500 26.67 16.67 70.00 82.40 26.84 31.31 42.31 2000 26.67 30.00 80.00 83.20 29.78 37.37 47.84 2500 16.67 23.33 67.50 86.00 32.35 34.34 43.37 Table 36: Performance of R...

    2000

  17. [2018]

    URLhttps://arxiv.org/abs/1610.01644. Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martín Blázquez, Guilherme Penedo, Lewis Tunstall, Andrés Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémen...

    arXiv

  18. [2021]

    Hugging Face

    URLhttps: //arxiv.org/abs/2103.03874. Hugging Face. Open r1: A fully open reproduction of deepseek-r1, January

    arXiv

  19. [2023]

    Quy-Anh Dang and Chris Ngo

    URLhttps://arxiv.org/abs/2309.08600. Quy-Anh Dang and Chris Ngo. Reinforcement learning for reasoning in small LLMs: What works and what doesn’t,

    arXiv

  20. [2024]

    URLhttps://arxiv.org/abs/ 2406.04093. Etash Guha, Ryan Marten, Sedrick Keh, Negin Raoof, Georgios Smyrnis, Hritik Bansal, Marianna Nezhurina, Jean Mercat, Trung Vu, Zayne Sprague, Ashima Suvarna, Benjamin Feuer, Liangyu Chen, Zaid Khan, Eric Frankel, Sachin Grover, Caroline Ch...

    arXiv

  21. [2025]

    Anthropic

    URLhttps://arxiv.org/abs/2502.02737. Anthropic. Towards monosemanticity: Decomposing language models with dictionary learning.Transformer Circuits Thread,

    arXiv

Pith tools

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