pith. sign in

arxiv: 2606.02248 · v1 · pith:CNEKXVJ6new · submitted 2026-06-01 · 💻 cs.CL

Geometric Latent Reasoning Induces Shorter Generations in LLMs

Shashi Kumar , Yacouba Kaloga , Petr Motlicek , Ina Kodrasi , Andrea Cavallaro This is my paper

Pith reviewed 2026-06-28 15:03 UTC · model grok-4.3

classification 💻 cs.CL
keywords latent reasoningchain-of-thoughtembedding spacelanguage modelsgeometric path approximationgeneration lengthmathematical reasoning
0
0 comments X

The pith

Replacing early explicit reasoning with continuous updates in embedding space lets models reach answers in fewer total generation steps.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper formulates latent reasoning as approximating paths through the continuous geometry of a model's token embedding space rather than generating discrete tokens at every step. It introduces a lightweight transition head that learns to output successive direction updates, using ordinary chain-of-thought text as training anchors so the path can deviate continuously from exact token points. When early reasoning steps are handled this way, the remaining generation produces correct answers after substantially fewer tokens overall on mathematical tasks. A sympathetic reader would care because the shortening appears without any explicit penalty on length, suggesting that compact continuous states can stand in for some discrete reasoning. The central result is that this geometric substitution reduces output length while preserving final correctness.

Core claim

By treating latent reasoning as a geometric path-approximation task inside pretrained embedding space, a lightweight transition head is trained to predict iterative direction updates anchored to textual chain-of-thought traces. The head permits continuous deviations from exact token embeddings, so early reasoning steps can be executed latently; the model then completes the task with substantially shorter total generations on mathematical benchmarks.

What carries the argument

The lightweight transition head that predicts iterative direction updates inside the model's token-embedding space, trained on textual chain-of-thought traces as anchors.

If this is right

  • Early explicit reasoning tokens can be replaced by continuous latent steps without any separate length objective.
  • Models reach correct answers after substantially fewer total generation steps.
  • Continuous trajectories function as compact intermediate reasoning states.
  • A tradeoff appears between the number of latent steps, final output length, and answer accuracy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could be tested on non-mathematical tasks that normally produce long explicit chains to see whether length reduction generalizes.
  • Varying the number of latent steps before switching to explicit generation might reveal an optimal budget that balances accuracy against total tokens.
  • Combining the geometric updates with existing inference-time efficiency methods could produce further reductions in compute per solved problem.

Load-bearing premise

The direction updates produced by the transition head, trained only on textual chain-of-thought examples, preserve enough semantic information to substitute for explicit token generation without degrading final answer correctness.

What would settle it

An experiment that disables the trained transition head or replaces its direction updates with random vectors in embedding space and checks whether generation lengths stay short and answers remain correct on the same mathematical benchmarks.

Figures

Figures reproduced from arXiv: 2606.02248 by Andrea Cavallaro, Ina Kodrasi, Petr Motlicek, Shashi Kumar, Yacouba Kaloga.

Figure 1
Figure 1. Figure 1: Geometric view of latent reasoning as an embedding-space trajectory. (a) Standard chain-of-thought forces reasoning through a sequence (black arrows) of exact vocabulary embeddings (purple dots). (b) GLR learns continuous displacement vectors (red arrows) to approximate these transitions. Dashed circles denote local neighborhoods where continuous states may remain meaning￾ful model inputs. (c) At inference… view at source ↗
Figure 2
Figure 2. Figure 2: Training pipeline for Geometric Latent Reasoning (GLR). Left (first forward pass): The model processes the original discrete sequence to collect both the exact (∆t k ) and predicted (∆ˆ t k ) embedding-space displacements between consecutive reasoning tokens. Right (second forward pass): Discrete thought embeddings (e t ) are replaced with continuous latent states (eˆ t ) obtained by applying the Transitio… view at source ↗
Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: GLR reduces redundant reasoning traces on SVAMP. Left: Accuracy vs. generation budget. On these simpler arithmetic problems where COT-SFT generates long traces, GLR maintains high accuracy under strict budgets (≤ 128 or 256 tokens). Right: Generation length distributions for correct answers. While COT-SFT expends hundreds of generated tokens to solve simple problems, GLR reduces the median generation lengt… view at source ↗
Figure 6
Figure 6. Figure 6: Generation length for Qwen3-1.7B GLR model at K = 0 vs. K > 0 on GSM8K. Continuous displacements drive generation efficiency. By construction, the transition head gϕ predicts continuous embedding-space dis￾placement vectors that need not coincide with exact transitions between vocabulary embed￾dings. To isolate the effect of this continuous deviation on the reasoning process, we evaluate the Qwen3-1.7B GLR… view at source ↗
Figure 7
Figure 7. Figure 7: Accuracy and generation length on AMC23. GLR reduces the median generation length on competition mathematics). 64 128 256 512 1024 2048 4096 Generation length budget (tokens) 0 20 40 60 80 100 Accuracy (%) Accuracy vs. generation length budget (OlympiadBench) CoT-SFT (0.6B) GLR-20 (0.6B) CoT-SFT (1.7B) GLR-10 (1.7B) CoT-SFT (0.6B) GLR-20 (0.6B) CoT-SFT (1.7B) GLR-10 (1.7B) Method 50 100 200 500 1000 2000 4… view at source ↗
Figure 8
Figure 8. Figure 8: Accuracy and generation length on OlympiadBench. GLR shifts the accuracy–length frontier on Olympiad-level mathematical reasoning. 64 128 256 512 1024 2048 Generation length budget (tokens) 0 20 40 60 80 100 Accuracy (%) Accuracy vs. generation length budget (MultiArith) CoT-SFT (0.6B) GLR-20 (0.6B) CoT-SFT (1.7B) GLR-10 (1.7B) CoT-SFT (0.6B) GLR-20 (0.6B) CoT-SFT (1.7B) GLR-10 (1.7B) Method 50 100 200 500… view at source ↗
Figure 9
Figure 9. Figure 9 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of total generated tokens across all answers on GSM8K. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of total generated tokens across all answers on MATH500. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Distribution of total generated tokens across all answers on AMC23. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Distribution of total generated tokens across all answers on OlympiadBench. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of total generated tokens across all answers on MultiArith. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative example of latent-to-text transition. A Qwen3-1.7B GLR-20 generation on GSM8K. After 20 latent steps inside the reasoning span, the model resumes explicit text generation mid-solution, directly using the relevant quantity and operation (54 − 20 = 34) before emitting the final answer. We display the continuous updates as {20 latent steps}; special tokens are shown in teal and latent-step placeh… view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative example of latent-to-text transition. A Qwen3-1.7B GLR-50 generation on GSM8K. We display the continuous updates as {50 latent steps}; special tokens are shown in teal and latent-step placeholders in orange. These implementations have different serving stacks and optimization levels, making direct latency comparisons confounded. A fair end-to-end speed comparison requires an optimized GLR serv… view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative example of latent-to-text transition. A Qwen3-1.7B GLR-80 generation on GSM8K. We display the continuous updates as {80 latent steps}; special tokens are shown in teal and latent-step placeholders in orange. G Future Work Future work should scale GLR to larger models and substantially larger reasoning datasets, as our current experiments are limited by compute and use only a 10K-example subset… view at source ↗
read the original abstract

Large language models solve complex problems by generating lengthy chains of explicit reasoning tokens. While effective, this makes reasoning expensive, length-sensitive, and constrained to (discrete) natural language. While latent reasoning offers a continuous alternative, determining useful structures for intermediate latent states is an open challenge. In this paper, we formulate latent reasoning as a geometric path-approximation problem within the model's pretrained token-embedding space. We introduce Geometric Latent Reasoning (GLR), which uses a lightweight transition head to predict iterative direction updates in embedding space. Using textual chain-of-thought traces as anchors, GLR learns to approximate discrete reasoning trajectories while permitting continuous deviations from exact token embeddings. Evaluations on mathematical reasoning benchmarks using Qwen3 models reveal an emergent phenomenon: geometric latent reasoning induces substantially shorter generations without an explicit length objective. By replacing early explicit reasoning with continuous latent steps, models often reach correct answers using substantially fewer total generation steps. These findings suggest that continuous trajectories act as compact intermediate reasoning states, exposing a new tradeoff between latent computation budget, output length, and accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces Geometric Latent Reasoning (GLR) as a formulation of latent reasoning as a geometric path-approximation problem in the pretrained token-embedding space of LLMs. A lightweight transition head predicts iterative direction updates, trained on textual chain-of-thought traces as anchors to allow continuous deviations from exact token embeddings. The central claim is that this induces substantially shorter generations on mathematical reasoning benchmarks with Qwen3 models as an emergent phenomenon, without an explicit length objective, by replacing early explicit reasoning steps with continuous latent steps while still reaching correct answers.

Significance. If the central claim holds under rigorous evaluation, the work would be significant for identifying an emergent tradeoff between latent computation budget, output length, and accuracy. The geometric framing and use of continuous trajectories as compact intermediate states could provide a new mechanism for efficient reasoning beyond discrete token generation.

major comments (3)
  1. [Abstract] Abstract and method description: no equations, training procedure, or loss function are provided for the transition head, nor any description of how it is trained or evaluated on CoT anchors. This prevents checking whether continuous deviations preserve semantic content or simply allow early termination before the answer head.
  2. [Abstract] Abstract: the claim that models 'often reach correct answers using substantially fewer total generation steps' is unsupported by any quantitative results, accuracy metrics, error bars, or comparison to baselines. Without these, it is impossible to determine whether length reduction preserves correctness or arises from non-semantic shortcuts.
  3. [Method] The setup trains the transition head only to predict next-token direction vectors from textual CoT traces, yet allows free continuous deviation at inference. No mechanism (regularization, manifold constraint, or consistency loss) is described to ensure deviations remain on a semantically valid reasoning path, so shorter generations may reflect reduced token emission rather than genuine latent computation.
minor comments (1)
  1. [Abstract] The abstract would be clearer with a brief statement of the model sizes, benchmark names, and observed length reduction factors.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below, providing clarifications from the full manuscript and indicating where revisions will strengthen the presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract and method description: no equations, training procedure, or loss function are provided for the transition head, nor any description of how it is trained or evaluated on CoT anchors. This prevents checking whether continuous deviations preserve semantic content or simply allow early termination before the answer head.

    Authors: The abstract is intentionally concise. The full Method section (Section 3) defines the transition head as a lightweight network predicting direction vectors d_t = W * h_t + b, with the training objective being the squared L2 loss between predicted directions and the embedding differences extracted from CoT token sequences. The head is trained by supervising on these anchor trajectories to approximate discrete steps while permitting continuous interpolation at inference. We will revise the abstract to include a one-sentence description of the loss and CoT supervision. revision: partial

  2. Referee: [Abstract] Abstract: the claim that models 'often reach correct answers using substantially fewer total generation steps' is unsupported by any quantitative results, accuracy metrics, error bars, or comparison to baselines. Without these, it is impossible to determine whether length reduction preserves correctness or arises from non-semantic shortcuts.

    Authors: The abstract summarizes the central empirical finding. Section 4 reports the full quantitative results on mathematical reasoning benchmarks with Qwen3 models, including per-task accuracy, mean generation lengths with standard deviations across multiple runs, and direct comparisons against standard CoT prompting and other baselines. These experiments show accuracy is maintained while total steps decrease. We will incorporate a brief summary of the key metrics into the abstract. revision: partial

  3. Referee: [Method] The setup trains the transition head only to predict next-token direction vectors from textual CoT traces, yet allows free continuous deviation at inference. No mechanism (regularization, manifold constraint, or consistency loss) is described to ensure deviations remain on a semantically valid reasoning path, so shorter generations may reflect reduced token emission rather than genuine latent computation.

    Authors: The CoT anchor supervision directly grounds the learned directions in the semantic structure of the pretrained embedding space, so that iterative updates follow trajectories that were originally derived from valid reasoning sequences. At inference the model applies the same direction predictions without token emission until the answer head is triggered. While no additional explicit regularization term is introduced, the emergent preservation of accuracy alongside reduced length provides evidence that the latent steps perform useful computation rather than mere early stopping. We will expand the Method section with a dedicated paragraph clarifying this inference procedure and the role of anchor-based training. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation is self-contained.

full rationale

The paper formulates GLR as a geometric path-approximation problem and trains a transition head on textual CoT traces to learn direction updates, with shorter generations reported as an emergent observation rather than a fitted or self-defined quantity. No equations, self-citations, or ansatzes are quoted that reduce the central claims (latent steps substituting for explicit tokens, length reduction without explicit objective) to inputs by construction. The method's training anchors and evaluation on benchmarks remain independent of the reported length effect, satisfying the criteria for a non-circular finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract only; no explicit free parameters, axioms, or invented entities are stated. The transition head itself is an added learned component whose parameters are presumably fitted but not quantified.

pith-pipeline@v0.9.1-grok · 5724 in / 1006 out tokens · 18057 ms · 2026-06-28T15:03:58.342149+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 20 canonical work pages · 14 internal anchors

  1. [1]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  2. [2]

    MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning

    Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning.arXiv preprint arXiv:2309.05653, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Training Large Language Models to Reason in a Continuous Latent Space

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space.arXiv preprint arXiv:2412.06769, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Softcot: Soft chain-of-thought for efficient reasoning with llms.arXiv preprint arXiv:2502.12134, 2025

    Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. Softcot: Soft chain-of-thought for efficient reasoning with llms.arXiv preprint arXiv:2502.12134, 2025

    work page arXiv 2025

  6. [6]

    CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. Codi: Compressing chain-of-thought into continuous space via self-distillation.arXiv preprint arXiv:2502.21074, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    arXiv preprint arXiv:2502.03275 , year=

    DiJia Su, Hanlin Zhu, Yingchen Xu, Jiantao Jiao, Yuandong Tian, and Qinqing Zheng. Token assorted: Mixing latent and text tokens for improved language model reasoning.arXiv preprint arXiv:2502.03275, 2025

    work page arXiv 2025

  8. [8]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    s1: Simple test-time scaling

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute opti- mally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  13. [13]

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

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023. 10

    2023

  14. [14]

    Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting.Advances in Neural Information Processing Systems, 36:74952–74965, 2023

    Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting.Advances in Neural Information Processing Systems, 36:74952–74965, 2023

    2023

  15. [15]

    Measuring Faithfulness in Chain-of-Thought Reasoning

    Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, et al. Measuring faithfulness in chain-of-thought reasoning.arXiv preprint arXiv:2307.13702, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Implicit chain of thought reasoning via knowledge distillation.arXiv preprint arXiv:2311.01460, 2023

    Yuntian Deng, Kiran Prasad, Roland Fernandez, Paul Smolensky, Vishrav Chaudhary, and Stuart Shieber. Implicit chain of thought reasoning via knowledge distillation.arXiv preprint arXiv:2311.01460, 2023

    work page arXiv 2023

  17. [17]

    Softcot++: Test-time scaling with soft chain-of-thought reasoning.arXiv preprint arXiv:2505.11484, 2025

    Yige Xu, Xu Guo, Zhiwei Zeng, and Chunyan Miao. Softcot++: Test-time scaling with soft chain-of-thought reasoning.arXiv preprint arXiv:2505.11484, 2025

    work page arXiv 2025

  18. [18]

    Soft thinking: Unlocking the reasoning potential of llms in continuous concept space.arXiv preprint arXiv:2505.15778, 2025

    Zhen Zhang, Xuehai He, Weixiang Yan, Ao Shen, Chenyang Zhao, Shuohang Wang, Yelong Shen, and Xin Eric Wang. Soft thinking: Unlocking the reasoning potential of llms in continuous concept space.arXiv preprint arXiv:2505.15778, 2025

    work page arXiv 2025

  19. [19]

    Hybrid latent reasoning via reinforcement learning.arXiv preprint arXiv:2505.18454, 2025

    Zhenrui Yue, Bowen Jin, Huimin Zeng, Honglei Zhuang, Zhen Qin, Jinsung Yoon, Lanyu Shang, Jiawei Han, and Dong Wang. Hybrid latent reasoning via reinforcement learning.arXiv preprint arXiv:2505.18454, 2025

    work page arXiv 2025

  20. [20]

    Open r1: A fully open reproduction of deepseek-r1, January 2025

    Hugging Face. Open r1: A fully open reproduction of deepseek-r1, January 2025

    2025

  21. [21]

    Openr1-math-220k

    Anton Lozhkov, Hynek Kydlíˇcek, Loubna Ben Allal, Guilherme Penedo, Edward Beeching, Quentin Gallouédec, Nathan Habib, Lewis Tunstall, and Leandro von Werra. Openr1-math-220k. https://huggingface.co/datasets/open-r1/OpenR1-Math-220k, 2025

    2025

  22. [22]

    Human-level control through deep reinforcement learning.nature, 518(7540):529–533, 2015

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning.nature, 518(7540):529–533, 2015

    2015

  23. [23]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  24. [24]

    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, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  25. [25]

    Measuring Mathematical Problem Solving With the MATH Dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  26. [26]

    Solving general arithmetic word problems

    Subhro Roy and Dan Roth. Solving general arithmetic word problems. InProceedings of the 2015 conference on empirical methods in natural language processing, pages 1743–1752, 2015

    2015

  27. [27]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, et al. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Pape...

    2024

  29. [29]

    Arkil Patel, Satwik Bhattamishra, and Navin Goyal. Are nlp models really able to solve simple math word problems? InProceedings of the 2021 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 2080–2094, 2021. 11

    2021

  30. [30]

    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 Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. The languag...

    2024

  31. [31]

    HuggingFace's Transformers: State-of-the-art Natural Language Processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface’s transform- ers: State-of-the-art natural language processing.arXiv preprint arXiv:1910.03771, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  32. [32]

    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 Systems Principles, 2023. 12 A Hyperparameters Both the COT-SFT baseline and our propo...

    2023