Pith. sign in

REVIEW 5 major objections 6 minor 55 references

Who Reasons in the Large Language Models?

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 LLM reasoning may trace to one linear layer: o_proj.

desk verdict A provocative but under-evidenced claim that o_proj carries reasoning; the SfN toolkit is worth knowing, but the central ranking claim needs better controls. read the letter →

arxiv 2505.20993 v1 pith:4BCUFMQX submitted 2025-05-27 cs.CL cs.AI

Jie Shao , Jianxin Wu This is my paper
classification cs.CLcs.AI
keywords LLMinterpretabilityoutputprojectiono_projreasoninglocalizationStethoscopeforNetworksparameter-efficientfine-tuningattentionmechanismmathematical
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

This paper tries to establish that reasoning ability in a well-trained large language model is not spread across the whole network but is dominated by one component: the output projection (o_proj) of the Transformer's multi-head self-attention. It introduces a diagnostic suite called Stethoscope for Networks (SfN), with four probes that compare a base chat model with its reasoning fine-tune: weight-difference analysis, module swapping, selective freezing during fine-tuning, and module destruction. Across Qwen and DeepSeek model families and sizes from 1.5B to 70B, the probes converge on the same picture: o_proj carries the reasoning-related change, while conversational fluency depends on the other attention projections and the MLP. If the hypothesis holds, reasoning can be trained or transplanted by touching a small fraction of parameters, which would make reasoning fine-tuning much faster and suggest that o_proj acts as a modular plugin.

What carries the argument

The central mechanism is the output projection (o_proj), the linear layer that maps multi-head attention outputs back into the residual stream of each Transformer block. Around it, Stethoscope for Networks (SfN) is a suite of four diagnostic operations: the $\Delta$ Stethoscope compares per-module $\ell^2$ weight differences and relative-change distributions between base and reasoning fine-tune; the Merge Stethoscope swaps modules from the reasoning model into the base model without any fine-tuning; the Freeze Stethoscope fine-tunes only selected modules; and the Destruction Stethoscope zeros, reinitializes, or removes modules to see which are needed for conversation. The load-bearing signature is that o_proj's weight shift is the largest and uniquely bimodal, and that swapping or tuning only o_proj transfers reasoning behavior while destroying it does not break conversation.

What would settle it

Re-run the Freeze Stethoscope on a reasoning dataset whose responses are length-matched to the base model's chat outputs, and also merge a randomly perturbed o_proj with the same $\ell^2$ norm as the real one; if the o_proj advantage vanishes in either test, the reasoning-localization claim is an artifact of output-length shift or of any large weight change, not of o_proj specifically.

Watch

Extended reading notes

Core claim

In the paper's own terms, the claim is that for an LLM that reasons well, the output projection o_proj in the multi-head self-attention module is the single or at least the most important module dominating reasoning ability. The evidence: the $\ell^2$ norm of weight change between a base model and its reasoning-distilled version is largest for o_proj (second largest at 1.5B, largest at 14B, 32B, and 70B), and only o_proj shows a bimodal distribution of relative weight changes. Merging only the o_proj weights of DeepSeek-R1-Distill-Qwen-1.5B into Qwen2.5-Math-1.5B makes the resulting model solve AIME 2024 questions the base cannot, while swapping q/k/v projections or the MLP degrades output; fine-tuning only o_proj and layernorm on 14B and 32B bases gives reasoning scores close to full fine-tuning; and zeroing or reinitializing o_proj leaves conversational output at level III while damaging the MLP collapses it. The paper also conjectures a division of labor: o_proj is for reasoning, all other modules are for fluent conversation.

Load-bearing premise

The argument rests on treating the weight difference between a base model and its reasoning fine-tune as a faithful map of where reasoning is stored; if the large o_proj changes are mostly a byproduct of optimization dynamics, response-length shifts, or normalization mismatches between models, the localization conclusion does not follow.

Editorial extensions

If this is right

  • Fine-tuning only o_proj plus normalization on a 14B or 32B base reaches AIME 2024 and Math 500 scores close to full-parameter fine-tuning, with roughly three times faster training at 32B and far fewer trainable parameters.
  • At 1.5B scale, replacing only o_proj in a chat-capable base with the reasoning model's o_proj lifts AIME 2024 accuracy from 0.067 to 0.200 with no further training; replacing q/k/v projections or the MLP instead degrades the output.
  • A chat model and a reasoning model built on the same base differ mainly in o_proj, layernorm, embeddings, and the LM head, about 10% of the model, so the two could be stored as one model with two small module sets.
  • The Destruction Stethoscope results imply o_proj is not needed for fluent conversation, so one could swap reasoning-specific o_proj layers in and out without harming chat quality.

Reading between the lines

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

  • A natural reading beyond the paper: o_proj may act as a readout that projects attention-selected information into the residual stream in a form that supports multi-step reasoning, which would explain why q/k/v and the MLP matter more for fluent, context-appropriate talk.
  • The Delta Stethoscope's weight shifts are computed between models that also differ in average response length; a controlled test with length-matched responses would separate reasoning-related weight changes from longer-generation effects.
  • If o_proj is a plugin, then parameter-efficient fine-tuning should concentrate most of its benefit in this layer; this predicts that low-rank adapters placed only on o_proj will match or exceed adapters placed elsewhere.
  • The reasoning-distillation pipeline studied here is SFT-based; whether the same localization appears after reinforcement-learning training is left open and is the most direct extension.
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 hypothesizes that, in LLMs that reason well, the output projection (o_proj) of each Transformer attention block is the single most important module for reasoning, while other modules primarily support fluent dialogue. To support this, the authors introduce Stethoscope for Networks (SfN), a suite of four diagnostic tools: the Delta Stethoscope (weight-difference analysis between a base and a reasoning-distilled model), the Merge Stethoscope (replacing modules in a base model with those from a reasoning model), the Freeze Stethoscope (fine-tuning only selected parameter subsets), and the Destruction Stethoscope (zeroing, reinitializing, or removing modules and observing conversational output). Experiments are conducted on Qwen/DeepSeek-R1-Distill model pairs and on Qwen2.5-14B/32B-Instruct with the s1K dataset. The reported evidence includes larger o_proj weight shifts in the Delta analysis, improved AIME 2024 performance after merging only o_proj, competitive reasoning performance when fine-tuning only o_proj plus normalization, and qualitative preservation of conversation when o_proj is destroyed.

Significance. If the central hypothesis were established, the result would be practically important: it would suggest that reasoning ability can be improved by tuning a very small parameter subset, enabling faster and more memory-efficient fine-tuning, and it would offer a modular view of LLM capabilities. The paper has clear strengths: it uses openly available models and benchmarks, reports results across several model sizes, combines multiple intervention-based diagnostics, and is unusually candid about the limitations of some of its evidence. The Merge and Freeze experiments are genuine interventions rather than purely correlational analyses, which is a step beyond many interpretability studies. However, the paper's strongest claim—that o_proj is the single or most important reasoning module—is not yet supported by the experimental design, because the Freeze experiments lack the controls needed to rank modules, and the quantitative evidence is based on very small benchmark sizes without variance estimates. The contribution is therefore best viewed as an interesting and well-articulated hypothesis with suggestive preliminary evidence, not as a demonstrated conclusion.

major comments (5)
  1. [Section 2.3, Table 2] The Freeze Stethoscope, which is the paper's main constructive evidence for Hypothesis 1, lacks the ablations needed to support a claim that o_proj is the 'single or at least the most important' reasoning module. The comparison F1 (Emb+Head) versus F2 (Emb+Head+o_proj) shows that adding o_proj to the tunable set improves AIME 2024 scores, but the paper never tunes q_proj, k_proj, v_proj, or any MLP component alone, nor a parameter-matched random subset of weights. Without such controls, the improvement from F1 to F2 could reflect additional trainable capacity or optimization dynamics rather than o_proj-specific function. Similarly, the fact that F3 (adding q/k/v_proj) does not improve over F2 does not establish that o_proj is more important than each of q_proj, k_proj, and v_proj individually, since the three projections are unfrozen together and could have offsetting effects. To support the ranking in Hypothesis 1, the authors should add single-module and parameter-matched baseline fine-tuning experiments.
  2. [Section 2.3, Table 2 and Section 2.2, Table 1] The quantitative evidence for the Freeze and Merge Stethoscopes rests on very small benchmark sizes with no variance estimates. AIME 2024 contains 30 problems, so the reported differences are small in absolute terms: in Table 2, F1 at 0.200 versus F2 at 0.367 corresponds to roughly 2 versus 6 correct answers, and in Table 1, model A at 0.067 versus M1 at 0.200 corresponds to roughly 2 versus 6 correct answers. Without repeated runs with different random seeds, confidence intervals, or a per-question breakdown, these differences are not statistically robust. The authors should report variance or supplement AIME with larger reasoning benchmarks to demonstrate that the observed improvements are not driven by a handful of questions.
  3. [Section 2.1, Assumption 1] The Delta Stethoscope is correlational, and the paper's own Section 2.3 acknowledges that normalization-parameter mismatches can confound weight-difference analyses at larger scales. The larger o_proj weight shift and its bimodal relative-change distribution are suggestive, but they do not by themselves establish that o_proj causally encodes reasoning; the shift could reflect optimization dynamics, output-length differences between the base and distilled models, or other training artifacts. The paper would be strengthened by a concrete control, such as applying the Delta analysis to a model fine-tuned to produce longer or more verbose outputs without improved reasoning, and showing that o_proj does not show the same distinctive pattern in that case.
  4. [Section 3.1, Table 3] The Destruction Stethoscope results, which support Conjecture 1 about the division of labor between o_proj and other modules, are based on a single conversation, as the paper itself states: 'the results are not statistics computed in many different experiments—it only reflects the conversation illustrated in Figure 4.' This is too fragile a basis for a general conjecture about conversational ability. The authors should evaluate destruction effects on a larger set of diverse dialogue prompts and, ideally, report a quantitative measure such as response coherence scores, perplexity, or task-specific metrics, rather than a single qualitative output-level judgment.
  5. [Section 2.3, Assumption 3 and Hypothesis 1] Hypothesis 1 is stated as o_proj being 'the single or at least the most important module that dominates' reasoning, but the paper never defines what 'importance' means operationally or compares o_proj against all other individual modules in a controlled way. The Freeze Stethoscope compares grouped modules (Emb+Head, o_proj, q/k/v/o_proj, all), and the Merge Stethoscope in Table 1 compares grouped replacements ({q,k,v}_proj and mlp) rather than individual modules. As a result, the evidence does not yet rule out the possibility that, for example, q_proj alone or a specific MLP projection is equally or more important for reasoning. The authors should either weaken the claim to 'o_proj is among the important modules' or provide per-module ablations that justify the 'most important' ranking.
minor comments (6)
  1. [Section 2.3, Table 2 caption] The footnote states that 'Without tuning these components, finetuning failed to converge,' but the table caption does not mention that F4 corresponds to full fine-tuning while Assumption 3 describes only o_proj and normalization tuning; this should be clarified to avoid confusion about which model is 'B' in Assumption 3.
  2. [Section 2.1, Figure 3] The figure caption says the vertical axis represents frequency, but the histograms do not show numeric axis labels; adding tick labels and a shared legend would improve interpretability.
  3. [Section 2.2, text after Table 1] There is a typo: 'has neverbeing finetuned' should be 'has never been fine-tuned.'
  4. [Section 3.1, paragraph on destructors] The ReInit destructor uses Gaussian noise with std=0.02, but no rationale is given for this choice; a brief justification or a sensitivity check would be helpful.
  5. [Appendix C] The first sentence says 'We report appropriate information regarding the statistical significance of our experiments,' but the following sentences state that the paper does not focus on p-values and provides no error bars or confidence intervals; this wording should be revised to avoid overstating the statistical backing.
  6. [Section 4, first bullet] The claim that tuning only o_proj can lead to 'Fast and better reasoning LLMs' is presented as a practical implication, but the experiments only show that o_proj tuning is competitive or slightly worse than full fine-tuning on the tested benchmarks; the wording should be tempered to match the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central o_proj claim is supported by fresh interventions on external benchmarks, and remaining gaps are experimental-control limitations, not definitional reductions.

full rationale

The paper's derivation chain is not circular. The Delta Stethoscope (Section 2.1) observes weight shifts and formulates a hypothesis; the Merge Stethoscope (Section 2.2) replaces o_proj and measures AIME 2024; the Freeze Stethoscope (Section 2.3) is a fresh fine-tuning intervention on a different base model and external benchmarks, with F2 versus F1 isolating the added o_proj tuning; and the Destruction Stethoscope (Section 3.1) manipulates modules and observes conversational output levels. None of these steps defines o_proj importance in terms of the conclusion itself: the Freeze experiment is not fitted to the Delta weight-shift data, and the benchmarks are independent of the module selection. The acknowledged limitations in the Conclusions (limited model families, qualitative destruction assessments, and lack of a theoretical account) are experimental and explanatory gaps rather than circular reductions. While the absence of parameter-matched controls for tuning q/k/v/MLP alone weakens the 'single most important module' claim, that is an underdetermination concern, not a self-referential or definitional equivalence. There is also no load-bearing self-citation chain: the cited Qwen, DeepSeek-R1, and s1 works are independent external sources. Therefore the paper receives a circularity score of 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim depends on four interpretative assumptions, one per stethoscope, rather than on fitted numeric parameters. The only hand-chosen experimental settings are the destruction layer range and the reinitialization noise scale. There are no invented entities.

free parameters (2)
  • Destruction layer range for conversation probes = blocks 5-30 of Qwen2.5-32B
    Chosen empirically so outputs do not all degrade to level I; this range influences which modules appear important in Table 3 (Section 3.2).
  • ReInit Gaussian standard deviation = 0.02
    Arbitrary destruction noise scale chosen for the ReInit destructor; affects severity but not the central reasoning claim (Section 3.1).
assumptions (5)
  • domain assumption Weight difference w(B)-w(A) contains essential information about the source of reasoning ability (Assumption 1).
    Underlies the Delta Stethoscope and motivates focus on o_proj; not derived and could be confounded by optimization dynamics or normalization shifts (Section 2.1).
  • ad hoc to paper If merging only o_proj from B into A yields stronger reasoning, then o_proj is crucial for reasoning (Assumption 2).
    Methodological assumption specific to the Merge Stethoscope; direct parameter substitution is treated as surgically meaningful (Section 2.2).
  • ad hoc to paper If fine-tuning only o_proj and normalization from A yields strong reasoning, then o_proj is crucial (Assumption 3).
    Underlies the Freeze Stethoscope; relies on frozen modules already providing the needed computation (Section 2.3).
  • ad hoc to paper Destruction of a module indicates the module's importance for conversation (Assumption 4).
    Underlies the Destruction Stethoscope; assumes zeroing or reinitializing a module tests only that module's role (Section 3.1).
  • domain assumption DeepSeek-R1-Distill-Qwen models are valid reasoning-enhanced fine-tunes of the corresponding Qwen base models.
    The whole comparison depends on B being obtained from A by fine-tuning; only then does w(B)-w(A) isolate reasoning acquisition (Section 2.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Who Reasons in the Large Language Models?." pith.science (2026). https://pith.science/paper/4BCUFMQX

@misc{pith2026250520993,
  author       = {Pith},
  title        = {Pith review of: Who Reasons in the Large Language Models?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BCUFMQX}},
  note         = {Machine review of arXiv:2505.20993}
}
read the original abstract

Despite the impressive performance of large language models (LLMs), the process of endowing them with new capabilities--such as mathematical reasoning--remains largely empirical and opaque. A critical open question is whether reasoning abilities stem from the entire model, specific modules, or are merely artifacts of overfitting. In this work, we hypothesize that the reasoning capabilities in well-trained LLMs are primarily attributed to the output projection module (oproj) in the Transformer's multi-head self-attention (MHSA) mechanism. To support this hypothesis, we introduce Stethoscope for Networks (SfN), a suite of diagnostic tools designed to probe and analyze the internal behaviors of LLMs. Using SfN, we provide both circumstantial and empirical evidence suggesting that oproj plays a central role in enabling reasoning, whereas other modules contribute more to fluent dialogue. These findings offer a new perspective on LLM interpretability and open avenues for more targeted training strategies, potentially enabling more efficient and specialized LLMs.

Figures

Figures reproduced from arXiv: 2505.20993 by the authors.

Figure 1
Figure 1. Stethoscope for Networks. SfN is a framework designed to identify which components of an LLM give rise to specific abilities. By comparing weight changes and observing behaviors under controlled module merging, tuning, or destruction, SfN provides interpretable insights into the origin of capabilities like reasoning. A definitive answer to any of the above questions will be extremely valuable to guiding the future d… view at source ↗
Figure 2
Figure 2. Per-module L2 distance of linear weights between models A and B. Notably, the o_proj module shows the second-largest change in 1.5B models, and the largest in 14B and 32B models, highlighting its potential importance for reasoning. Similar trends are observed in 7B and 8B models (see appendix). Qwen2.5-32B vs. DeepSeek-R1-Distill-Qwen-32B Qwen2.5-14B vs. DeepSeek-R1-Distill-Qwen-14B self_attn.q_proj self_attn.k_proj… view at source ↗
Figure 3
Figure 3. Layer-wise distribution of relative weight changes between models A and B. While most modules display a unimodal distribution, the o_proj module uniquely exhibits a bimodal distribution, highlighting its distinctive behavior. Consistent patterns are observed across models of other sizes, with detailed results provided in the appendix. Therefore, it is reasonable to guess that the output projection o_proj plays a piv… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Four levels of responses generated by the LLM. From level I to level IV, the model exhibits stronger language organization and logical reasoning skills. Each example includes a question (e.g., a math problem from AIME or a typical user-issued request) and the correspon…
Figure 5
Figure 5. Figure 5: Examples of outputs generated by merged models. Only M1 produces both a valid reasoning process and the correct answer. 2.3 The Freeze Stethoscope As models A and B scale up (e.g., to 7B parameters), merging components such as q,k,v_proj or mlp still results in signifi…
Figure 6
Figure 6. Figure 6: Training loss curves for fine-tuning Qwen2.5-14B,32B-Instruct on reasoning tasks. Different models unfreeze different sets of parameters, as detailed in [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Per-module L2 distance of linear weights between models A and B. Notably, the o_proj module shows the largest in 7B, 8B and 70B models, highlighting its potential importance for reasoning. Llama3.3-70B-Instruct vs. DeepSeek-R1-Distill-Llama-70B Llama-3.1-8B vs. DeepSee…
Figure 8
Figure 8. Figure 8: Layer-wise distribution of relative weight changes between models A and B. While most modules display a unimodal distribution, the o_proj module uniquely exhibits a bimodal distribution, highlighting its distinctive behavior. establish the robustness of our findings. T…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 4 canonical work pages

  1. [1]

    Physics of language models: Part 3.1, knowledge storage and extraction.arXiv preprint arXiv:2309.14316, 2023

    Zeyuan Allen-Zhu and Yuanzhi Li. Physics of language models: Part 3.1, knowledge storage and extraction.arXiv preprint arXiv:2309.14316, 2023

    arXiv 2023

  2. [2]

    Physics of language models: Part 3.2, knowledge manipula- tion.arXiv preprint arXiv:2309.14402, 2023

    Zeyuan Allen-Zhu and Yuanzhi Li. Physics of language models: Part 3.2, knowledge manipula- tion.arXiv preprint arXiv:2309.14402, 2023

    arXiv 2023

  3. [3]

    Physics of language models: Part 3.3, knowledge capacity scaling laws.arXiv preprint arXiv:2404.05405, 2024

    Zeyuan Allen-Zhu and Yuanzhi Li. Physics of language models: Part 3.3, knowledge capacity scaling laws.arXiv preprint arXiv:2404.05405, 2024

    arXiv 2024

  4. [4]

    Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    arXiv 2016

  5. [5]

    Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    arXiv 2023

  6. [6]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    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, 2020

    1901

  7. [7]

    Palm: Scaling language modeling with pathways.arXiv preprint arXiv:2204.02311, 2022

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Nina Mielke, Alec Radford, et al. Palm: Scaling language modeling with pathways.arXiv preprint arXiv:2204.02311, 2022

    arXiv 2022

  8. [8]

    Deep reinforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom B Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. InAdvances in neural information processing systems, volume 30, 2017

    work page 2017

Show all 55 references
  1. [9]

    Scaling instruction-finetuned language models

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Xin Wang, Xingyu Yuan, Adams Yu, Sharan Narang, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022

    2022 arXiv

  2. [10]

    The language model evaluation harness, 07 2024

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang S...

    2024

  3. [11]

    Transformer feed-forward layers are key-value memories.arXiv preprint arXiv:2012.14913, 2021

    Mor Geva, Tal Schuster, and Jonathan Berant. Transformer feed-forward layers are key-value memories.arXiv preprint arXiv:2012.14913, 2021

    2012 arXiv

  4. [12]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  5. [13]

    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

    2025 arXiv

  6. [14]

    Deepseek-coder: When the large language model meets programming–the rise of code intelligence.arXiv preprint arXiv:2401.14196, 2024

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming–the rise of code intelligence.arXiv preprint arXiv:2401.14196, 2024

    2024 arXiv

  7. [15]

    A structural probe for finding syntax in word representations

    John Hewitt and Christopher D Manning. A structural probe for finding syntax in word representations. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Pape...

    2019

  8. [16]

    Transformer quality in linear time

    Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In International conference on machine learning, pages 9099–9117. PMLR, 2022. 12

    2022

  9. [17]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jia- jun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report.arXiv preprint arXiv:2409.12186, 2024

    2024 arXiv

  10. [18]

    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

    2024 arXiv

  11. [19]

    Aime 2024 dataset

    Maxwell Jia. Aime 2024 dataset. https://huggingface.co/datasets/Maxwell-Jia/ AIME_2024, 2024

    2024

  12. [20]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

    2023

  13. [21]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  14. [22]

    Natural questions: a benchmark for question answering research.Transactions of the Association for Computational Linguistics, 7:453–466, 2019

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research.Transactions of the Association for Computat...

    2019

  15. [23]

    Gonzalez, Hao Zhang, and Ion Stoica

    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 lan- guage model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating System...

    2023

  16. [24]

    Locating and editing factual associations in gpt.Advances in neural information processing systems, 35:17359–17372, 2022

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt.Advances in neural information processing systems, 35:17359–17372, 2022

    2022

  17. [25]

    s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393, 2025

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling.arXiv preprint arXiv:2501.19393, 2025

    2025 arXiv

  18. [26]

    Abstractive text sum- marization using sequence-to-sequence rnns and beyond.arXiv preprint arXiv:1602.06023, 2016

    Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. Abstractive text sum- marization using sequence-to-sequence rnns and beyond.arXiv preprint arXiv:1602.06023, 2016

    2016 arXiv

  19. [27]

    Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization.arXiv preprint arXiv:1808.08745, 2018

    Shashi Narayan, Shay B Cohen, and Mirella Lapata. Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization.arXiv preprint arXiv:1808.08745, 2018

    2018 arXiv

  20. [28]

    Training language models to follow instructions with human feedback.arXiv preprint arXiv:2203.02155, 2022

    Long Ouyang, Jeff 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.arXiv preprint arXiv:2203.02155, 2022

    2022 arXiv

  21. [29]

    Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    2019

  22. [30]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    2023

  23. [31]

    Zero: Memory optimiza- tions toward training trillion parameter models

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: Memory optimiza- tions toward training trillion parameter models. InSC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–16. IEEE, 2020

    2020

  24. [32]

    Squad: 100,000+ questions for machine comprehension of text.arXiv preprint arXiv:1606.05250, 2016

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for machine comprehension of text.arXiv preprint arXiv:1606.05250, 2016. 13

    2016 arXiv

  25. [33]

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters

    Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 3...

    2020

  26. [34]

    The mechanistic basis of data dependence and abrupt learning in an in-context classification task.arXiv preprint arXiv:2312.03002, 2023

    Gautam Reddy. The mechanistic basis of data dependence and abrupt learning in an in-context classification task.arXiv preprint arXiv:2312.03002, 2023

    2023 arXiv

  27. [35]

    Tool- former: Language models can teach themselves to use tools.arXiv preprint arXiv:2302.04761, 2023

    Timo Schick, Ananya Dwivedi-Yu, Roberta Raileanu, Saghar Hosseini, Murray Chadwick, Gau- rav Mishra, Siddharth Karamcheti, Neil Houlsby, Aravind Elangovan, Mike Lewis, et al. Tool- former: Language models can teach themselves to use tools.arXiv preprint arXiv:2302.04761, 2023

    2023 arXiv

  28. [36]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  29. [37]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    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, 2024

    2024 arXiv

  30. [38]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

    2024

  31. [39]

    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

    2025 arXiv

  32. [40]

    QwQ-32B: Embracing the Power of Reinforcement Learning, March 2025

    Qwen Team. QwQ-32B: Embracing the Power of Reinforcement Learning, March 2025

    2025

  33. [41]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  34. [42]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  35. [43]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  36. [44]

    Analyzing the structure of attention in a transformer language model.arXiv preprint arXiv:1906.04284, 2019

    Jesse Vig and Yonatan Belinkov. Analyzing the structure of attention in a transformer language model.arXiv preprint arXiv:1906.04284, 2019

    1906 arXiv

  37. [45]

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

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

    2022 arXiv

  38. [46]

    Self-instruct: Aligning language models with self-generated instruc- tions.arXiv preprint arXiv:2212.10560, 2022

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instruc- tions.arXiv preprint arXiv:2212.10560, 2022

    2022 arXiv

  39. [47]

    Emergent abilities of large language models.arXiv preprint arXiv:2206.07682, 2022

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.arXiv preprint arXiv:2206.07682, 2022

    2022 arXiv

  40. [48]

    Chain-of-thought prompting elicits reasoning in large language models.arXiv preprint arXiv:2201.11903, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models.arXiv preprint arXiv:2201.11903, 2022

    2022 arXiv

  41. [49]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain 14 Gugger, M...

    2020

  42. [50]

    Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144, 2016

    Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation.arXiv preprint arXiv:1609.08144, 2016

    2016 arXiv

  43. [51]

    Qwen3 technical report, 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

  44. [52]

    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, 2024

    2024 arXiv

  45. [53]

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, et al. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement.arXiv preprint arXiv:2409.12122, 2024

    2024 arXiv

  46. [54]

    Simplerl-zoo: Investigating and taming zero reinforcement learning for open base models in the wild, 2025

    Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, and Junxian He. Simplerl-zoo: Investigating and taming zero reinforcement learning for open base models in the wild, 2025

    2025

  47. [55]

    Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625, 2023

    Denny Zhou, Dale Schuurmans, Xuezhi Wang, Ed Chi, and Quoc V Le. Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625, 2023. 15

    2023 arXiv

Pith tools

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