Pith. sign in

REVIEW 4 major objections 5 minor 21 references

Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Systematic tuning of four inference parameters cuts math-reasoning cost by 29.4% and speeds five open LLMs by 23.9%.

desk verdict Reported gains are simulated, not measured—Section 3.5 contradicts the paper's empirical claims; not worth referee time. read the letter →

arxiv 2509.07238 v1 pith:J6UYZTCS submitted 2025-09-08 cs.LG cs.AI

Pranav Pawar , Dhwaj Jain , Varun Gupta , Kaustav Dedhia , Dashrath Kale , Sudhir Dhekane This is my paper
classification cs.LGcs.AI
keywords mathematicalreasoninghyperparameteroptimizationinferenceefficiencytemperaturetuningnucleussamplingchain-of-thoughtlargelanguagemodelscost
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 how you call an LLM—not just how you build it—can be the main lever in math-reasoning cost. It reports that tuning four exposed inference controls (temperature, maximum reasoning steps, planning interval, top-p) across five open models yields an average 29.4% reduction in tokens per correct answer and 23.9% faster inference, with accuracy unchanged at 77.8% and success rate up 25.7%. The recipe is consistent across architectures: low temperature (0.1–0.4), short reasoning depth (4–6 steps), and high nucleus sampling (0.85–0.98). A 50-problem math benchmark is used, with performance estimated via simulations and reference benchmarks—the paper discloses that no direct model inference was carried out—so the numbers are predictions of live behavior, not measurements.

What carries the argument

The central object is the four-dimensional parameter space P = {T, S, I, P}, searched by a three-phase smart sampling procedure: baseline establishment, a 15-sample grid search, and iterative refinement. Configurations are scored by a weighted objective f = 0.4·Accuracy + 0.4·Efficiency + 0.2·Speed, where efficiency is the inverse of tokens per correct answer and speed is the inverse of average inference time. The framework's multi-agent components—BaseAgent, ReActAgent, AdaptivePlanner, and CostTracker—make the search model-agnostic, with the cost-of-pass metric (tokens per correct solution) carrying the efficiency analysis.

What would settle it

Re-run the five models on the same 50-problem set with the paper's optimal configurations (e.g., DeepSeek-V3 at temperature 0.2, max steps 6, planning interval 1, top-p 0.98) using actual live inference, and compare tokens-per-correct and wall-clock inference time against default settings. If the optimized configs do not improve by roughly the reported margins, or if accuracy drops, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that inference-time hyperparameters are a controllable source of large efficiency gains in mathematical reasoning. On a 50-problem GSM8K-derived benchmark, a multi-agent framework that searches temperature (0.1–0.5), maximum reasoning steps (4–12), planning interval (1–4), and top-p (0.85–0.98) finds settings that cut tokens-per-correct-answer by an average 29.4% and average inference time by 23.9% across Qwen2.5-72B, Llama-3.1-70B, DeepSeek-V3, Mixtral-8x22B, and Yi-Lightning, with accuracy flat at 0.778 and success rate up 25.7%. The recommended regimes converge to low temperature (80% of best configs at 0.2 or below), short reasoning chains (60% at 6 steps or

Load-bearing premise

The load-bearing premise is that the simulated and reference-benchmark estimates faithfully predict real model behavior; the paper explicitly states that no direct model inference was carried out, so if live runs diverge from those estimates the reported gains would disappear.

Editorial extensions

If this is right

  • Organizations running math-reasoning workloads can adopt the recommended settings immediately, since no retraining or architecture change is required.
  • The universal patterns (low temperature, short reasoning, high top-p) give new open-weights models a starting point for tuning instead of a full search.
  • Model selection becomes sharper: DeepSeek-V3 for accuracy-critical tasks, Mixtral-8x22B for cost-critical bulk workloads, Yi-Lightning for latency-sensitive use.
  • Parameter optimization and model architecture improvements are complementary, since both dense and MoE models improved with the same tuning recipe.
  • A 100% optimization success rate suggests that default configurations generally leave efficiency on the table.

Reading between the lines

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

  • The quantitative figures (29.4%, 23.9%, 98%) are estimates from simulations and reference benchmarks, not measurements; the decisive check is re-running the reported optimal configurations with live inference.
  • Low temperature and short reasoning chains are likely tuned to deterministic, benchmark-style math tasks; the same settings could suppress useful diversity on open-ended or creative-generation tasks.
  • The AdaptivePlanner suggests dynamic per-problem reasoning depth, but the paper only reports static optimized settings; dynamic allocation could squeeze further gains on mixed-difficulty workloads.
  • The 15% planning-interval benefit is not isolated from the ReAct loop itself; ablating planning on and off would show whether plan revision or structured prompting produces the gain.
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

4 major / 5 minor

Summary. The paper proposes a multi-agent framework for optimizing inference hyperparameters (temperature, max reasoning steps, planning interval, top-p) of five open-source LLMs on a 50-problem math reasoning subset of GSM8K. It defines a multi-objective score f = 0.4A + 0.4E + 0.2S (Eq. 6) and reports large efficiency gains, including a 29.4% average cost reduction, 23.9% speed improvement, and 98% accuracy for DeepSeek-V3. However, Section 3.5 contains an explicit evaluation-protocol note stating that no direct model inference was carried out and that all performance figures are estimates from simulations and reference benchmarks. The paper nevertheless presents these figures as measured experimental results, applies paired t-tests and confidence intervals to them, and draws deployment conclusions from them.

Significance. If the reported results were genuine measurements, the study would provide practical, actionable guidance for LLM inference parameter tuning in mathematical reasoning, and the cross-model comparisons would be of interest to practitioners. The framework itself is clearly described and the parameter-space decomposition is reasonable. However, the central empirical claims rest on a protocol that explicitly disavows direct inference. Without a description of the simulation, its calibration, or its relationship to real model behavior, the reported accuracy, cost, and speed numbers cannot be considered evidence. The statistical validation in Section 3.7 is also internally inconsistent with the stated protocol. The contribution is therefore currently unsupported by the manuscript's own evidence.

major comments (4)
  1. [Section 3.5, Evaluation Protocol Note (after Eq. 10)] The manuscript states: 'Our testing process estimates expected performance through simulations and reference benchmarks. No direct model inference was carried out.' This directly contradicts the empirical framing of the abstract and Section 5, where Tables 5 and 7 report concrete measurements such as DeepSeek-V3's 98% accuracy, Mixtral's 361.5 tokens per correct answer, and baseline/optimized averages (565.0 vs. 418.0 tokens; 4.87 vs. 3.62 seconds). No details are given about the simulation methodology, the reference benchmarks used, or how simulated estimates are mapped to model-specific numbers. These numbers are therefore unverifiable and irreproducible. This is not a minor caveat; it invalidates the paper's main contributions.
  2. [Section 3.7 and Section 5.6] Paired t-tests, bootstrap confidence intervals, Bonferroni correction, and the reported significance values (p < 0.01, 99% confidence, Cohen's d > 0.8) require repeated, independent observations of model outputs under each configuration. If no direct model inference was carried out, as Section 3.5 states, there are no such observations. The manuscript does not explain what quantity was resampled or what distribution the t-test was applied to. The statistical-significance claims are therefore meaningless under the stated protocol and should be removed or replaced with an explicit statement that these are performance estimates, not measured outcomes with sampling noise.
  3. [Section 3.4, Eq. (6) and Section 5.2/5.5] The objective function being maximized is f = 0.4A + 0.4E + 0.2S, where E = 1/(tokens per correct answer) and S = 1/(inference time). The reported 'improvements' in cost efficiency and speed are the very quantities being optimized. Reporting 29.4% cost reduction and 23.9% speed improvement as empirical discoveries is circular unless those metrics are evaluated independently of the objective—for instance, on a held-out benchmark or under a different cost model. The paper should clarify what part of the reported gains is attributable to the objective by construction and what part reflects genuinely novel generalization.
  4. [Section 3.5 and Section 4.1] The evaluation uses only 50 carefully selected problems and 8 configurations per model, with no description of how the problems were selected or whether the set is representative. The paper later generalizes to 'universal optimization patterns' and 'production-ready configurations.' Even setting aside the simulation issue, the sample size is too small to support such broad claims. The manuscript should either temper the generality claims or provide a larger, publicly defined evaluation set and report per-configuration results.
minor comments (5)
  1. [Table 6] The formatting of the DeepSeek-V3 row appears to be missing the Planning Interval column ('0.26 1 0.98' is ambiguous). Please verify that the table is rendered correctly.
  2. [Abstract and Section 1.2] The phrase 'the first comprehensive optimization study' is overstated; no baseline comparison of prior systematic hyperparameter studies for math reasoning is provided to justify the uniqueness claim.
  3. [Section 2.4 reference [2]] Reference [2] is cited as 'Anonymous' and 'ACL ARR 2025 Submission.' Such a citation is not ideal for a formal journal submission; please replace it with a fully identified reference or remove it.
  4. [General] The paper would benefit from a clear statement at the outset that all results are simulation-based estimates, rather than burying this information in Section 3.5. This would align reader expectations and make the scope of the claims transparent.
  5. [Algorithm 1] Algorithm 1 mentions 'k=15' for smart sampling, but Section 4.1 states 8 configurations per model were used. This discrepancy should be reconciled.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported efficiency/speed gains are in-sample outcomes of the very objective used to select configurations, not independent predictions.

  1. fitted input called prediction [Section 3.4 Eq. (6); Section 5.5 Table 7]
    "max θ∈P f(θ, M) = α·A(θ, M) + β·E(θ, M) + γ·S(θ, M) ... The weights α=0.4, β=0.4, and γ=0.2 were determined through preliminary experiments to keep accuracy intact while improving efficiency. ... Table 7: Cost Efficiency 565.0 tokens 418.0 tokens +29.4%"

    The configurations are selected as argmax of f, which explicitly rewards E (efficiency) and S (speed). The headline claims—29.4% cost reduction and 23.9% speed improvement—are exactly the E and S components of the objective being maximized. Any chosen configuration is selected precisely because it scores higher on these components; reporting them as discovered gains is reporting the selection criterion, not an independent empirical result. The weight choice (β=0.4, γ=0.2) further bakes efficiency/speed priority into the selection.

  2. self definitional [Algorithm 1 and Section 4.1]
    "Algorithm 1: θ∗ ← arg max_{θ∈G} f(θ, M) ... Section 4.1: 'Evaluation Scale: 50 problems per configuration across 5 models.'"

    The same 50-problem set D is used both to select θ* (Evaluate(M,θ,D)) and to report the final improvements in Table 7. Thus the 'optimized' metrics are the objective values on the training/evaluation set, not predictions on held-out problems. Success is defined as f(θ*,D) > f(θ_default,D), so '100% optimization success rate' is a restatement of the argmax definition rather than an out-of-sample validation.

full rationale

The paper's central empirical claims are further undermined by Section 3.5's explicit statement that 'Our testing process estimates expected performance through simulations and reference benchmarks. No direct model inference was carried out.' That is a validity/correctness problem more than circularity. The circularity is that the optimization procedure selects configurations by maximizing an objective whose components are the very efficiency and speed metrics later reported as gains, and the evaluation is performed on the same 50-problem set used for selection. Consequently, the reported improvements reduce by construction to the selection rule. The specific magnitudes and the accuracy-maintained result are not logically forced, but the direction of the headline efficiency/speed improvements is baked into the objective. No load-bearing self-citation or imported uniqueness theorem appears. Score 6 reflects one or more predictions reducing by construction.

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

The entire central claim rests on simulated performance, an unvalidated objective function, and a small unlisted benchmark. No external data or code is provided.

free parameters (3)
  • alpha (accuracy weight) = 0.4
    Chosen through 'preliminary experiments' (Section 3.4) to keep accuracy intact; directly sets the objective function and thus the selected optimal configs.
  • beta (efficiency weight) = 0.4
    Chosen through 'preliminary experiments' to favor efficiency; directly ensures efficiency gains are prioritized in the optimization outcome.
  • gamma (speed weight) = 0.2
    Chosen through 'preliminary experiments' to downweight speed; affects the trade-off and therefore which configs are reported as optimal.
assumptions (3)
  • domain assumption Simulations and reference benchmarks accurately estimate real model performance.
    Explicitly stated in Section 3.5: 'No direct model inference was carried out.' All results depend on this without validation.
  • ad hoc to paper The weighted objective function f = 0.4A + 0.4E + 0.2S reflects production deployment priorities.
    Introduced in Section 3.4 to formalize the optimization target; the weights are chosen by the authors and are not derived from any external standard.
  • domain assumption The 50-problem benchmark is representative of mathematical reasoning tasks.
    Stated in Section 3.1 as '50 representative math problems' with no justification or list of problems.
invented entities (1)
  • Multi-agent optimization framework (BaseAgent, ReActAgent, AdaptivePlanner, CostTracker)
    purpose: Orchestrate parameter search and evaluation across models
    Described at a high level in Section 3.1 but no code or implementation artifacts are provided, so its existence and functionality cannot be independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning." pith.science (2026). https://pith.science/paper/J6UYZTCS

@misc{pith2026250907238,
  author       = {Pith},
  title        = {Pith review of: Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J6UYZTCS}},
  note         = {Machine review of arXiv:2509.07238}
}
read the original abstract

This paper presents a practical investigation into fine-tuning model parameters for mathematical reasoning tasks through experimenting with various configurations including randomness control, reasoning depth, and sampling strategies, careful tuning demonstrates substantial improvements in efficiency as well as performance. A holistically optimized framework is introduced for five state-of-the-art models on mathematical reasoning tasks, exhibiting significant performance boosts while maintaining solution correctness. Through systematic parameter optimization across Qwen2.5-72B, Llama-3.1-70B, DeepSeek-V3, Mixtral-8x22B, and Yi-Lightning, consistent efficiency gains are demonstrated with 100% optimization success rate. The methodology achieves an average 29.4% reduction in computational cost and 23.9% improvement in inference speed across all tested models. This framework systematically searches parameter spaces including temperature (0.1-0.5), reasoning steps (4-12), planning periods (1-4), and nucleus sampling (0.85-0.98), determining optimal configurations through testing on mathematical reasoning benchmarks. Critical findings show that lower temperature regimes (0.1-0.4) and reduced reasoning steps (4-6) consistently enhance efficiency without compromising accuracy. DeepSeek-V3 achieves the highest accuracy at 98%, while Mixtral-8x22B delivers the most cost-effective performance at 361.5 tokens per accurate response. Key contributions include: (1) the first comprehensive optimization study for five diverse SOTA models in mathematical reasoning, (2) a standardized production-oriented parameter optimization framework, (3) discovery of universal optimization trends applicable across model architectures, and (4) production-ready configurations with extensive performance characterization.

Figures

Figures reproduced from arXiv: 2509.07238 by the authors.

Figure 1
Figure 1. Multi-Agent Optimization Framework Architecture for Systematic LLM Parameter [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. End-to-End Optimization Methodology Flowchart. Depicts the systematic workflow: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Parameter-Metric Correlation Network. Visualizes connections among important [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Mathematical reasoning performance comparison across five optimized large language [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Model Performance by Problem Category. Accuracy for each LLM on various mathe [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Efficiency Frontier: Accuracy vs. Cost Trade-off. Plots optimized model accuracy [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Top-p Configuration by Model. Optimized nucleus sampling threshold ( [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Parameter Space Exploration: Temperature vs Top- [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 10 canonical work pages

  1. [1]

    Systematic Evaluation of Optimization Techniques for Long-Context Language Models

    Ammar Ahmed, Sheng Di, Franck Cappello, Zirui Liu, Jingoo Han, and Ali Anwar. System- atic evaluation of optimization techniques for long-context language models.arXiv preprint arXiv:2508.00305, 2025

    work page Pith review arXiv 2025

  2. [2]

    Trimer: Balancing efficiency and accuracy in mathematical reasoning through a three-stage llm pipeline.ACL ARR 2025 Submission, 2025

    Anonymous. Trimer: Balancing efficiency and accuracy in mathematical reasoning through a three-stage llm pipeline.ACL ARR 2025 Submission, 2025

    work page 2025

  3. [3]

    Badam: A memory efficient full parameter optimization method for large language models.arXiv preprint arXiv:2404.02827, 2024

    QingyanChen, MengzhaoJiang, HaotianJiang, ChenxiFang, QianLi, MengZhu, Xiaodong Liu, and Jianfeng Gao. Badam: A memory efficient full parameter optimization method for large language models.arXiv preprint arXiv:2404.02827, 2024

    arXiv 2024

  4. [4]

    Optimization Hyper-parameter Laws for Large Language Models

    Xinyuan Chen, Qingquan Zhang, Lulu Xu, Zihan Jiang, Xiyang Ma, Fangcheng Wang, Chentao Fan, and Jue Wang. Optimization hyper-parameter laws for large language models. arXiv preprint arXiv:2409.04777, 2024

    work page Pith review arXiv 2024

  5. [5]

    Scalable llm math reasoning accel- eration with low-rank distillation.arXiv preprint arXiv:2505.07861, 2025

    Harry Dong, Bilge Acun, Beidi Chen, and Yuejie Chi. Scalable llm math reasoning accel- eration with low-rank distillation.arXiv preprint arXiv:2505.07861, 2025

    work page arXiv 2025

  6. [6]

    Towards revealing the mystery behind chain of thought: a theoretical perspective.arXiv preprint arXiv:2305.15408, 2023

    Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. Towards revealing the mystery behind chain of thought: a theoretical perspective.arXiv preprint arXiv:2305.15408, 2023

    arXiv 2023

  7. [7]

    Causal sufficiency and necessity improves chain-of-thought reasoning.arXiv preprint arXiv:2506.09853, 2025

    Steven Y Feng, Vivek Kumar, Sachit Mishra, Kyle Richardson, and Yulia Tsvetkov. Causal sufficiency and necessity improves chain-of-thought reasoning.arXiv preprint arXiv:2506.09853, 2025

    arXiv 2025

  8. [8]

    Parameter-efficient fine-tuning of large language models using semantic knowledge tuning.Nature Scientific Reports, 14:28341, 2024

    Nojun Kwak He. Parameter-efficient fine-tuning of large language models using semantic knowledge tuning.Nature Scientific Reports, 14:28341, 2024

    work page 2024

Show all 21 references
  1. [9]

    Towards efficient mixture of experts: A holistic study of compression techniques.arXiv preprint arXiv:2406.02500, 2024

    Shwai He, Daize Dong, Liang Ding, and Ang Li. Towards efficient mixture of experts: A holistic study of compression techniques.arXiv preprint arXiv:2406.02500, 2024

    2024 arXiv

  2. [10]

    Unveiling the key factors for distilling chain-of-thought reasoning.arXiv preprint arXiv:2502.18001, 2025

    Xanh He, Zecheng Zhao, Yifei Dou, Wenkai Zhang, Yongqi Liu, Enhong Chen, and Zhao Cao. Unveiling the key factors for distilling chain-of-thought reasoning.arXiv preprint arXiv:2502.18001, 2025

    2025 arXiv

  3. [11]

    Comparing retrieval-augmentation and parameter-efficient fine- tuning for privacy-preserving personalization of large language models.arXiv preprint arXiv:2409.09510, 2024

    Alireza Salemi Jiang, Sheshera Hasan, Hua Chu, Qiong Dong, Mohit Iyyer, and Ameet Deshpande. Comparing retrieval-augmentation and parameter-efficient fine- tuning for privacy-preserving personalization of large language models.arXiv preprint arXiv:2409.09510, 2024

    2024 arXiv

  4. [12]

    Collab- orative llm inference via planning for efficient reasoning.arXiv preprint arXiv:2506.11578, 2025

    Byeongchan Lee, Jonghoon Lee, Dongyoung Kim, Jaehyung Kim, and Jinwoo Shin. Collab- orative llm inference via planning for efficient reasoning.arXiv preprint arXiv:2506.11578, 2025

    2025 arXiv

  5. [13]

    Large language model agent for hyper-parameter optimization.arXiv preprint arXiv:2402.01881, 2024

    Siyi Li, Nanjie Liu, An Zhang, Meng Liu, Yifang Miao, and Sheng Li. Large language model agent for hyper-parameter optimization.arXiv preprint arXiv:2402.01881, 2024. 15

    2024 arXiv

  6. [14]

    Federated prompting and chain-of-thought reasoning for improving llms answering

    Xiangyang Liu, Tianqi Lu, Xiangyang Xu, Bing Chen, Yaqi Jin, Michael Lyu, and Irwin King. Federated prompting and chain-of-thought reasoning for improving llms answering. arXiv preprint arXiv:2304.13911, 2023

    2023 arXiv

  7. [15]

    General purpose veri- fication for chain of thought prompting.arXiv preprint arXiv:2405.00204, 2024

    Robert Singh, Haotian Wei, Tanmay Rajpurohit, and Denny Zhou. General purpose veri- fication for chain of thought prompting.arXiv preprint arXiv:2405.00204, 2024

    2024 arXiv

  8. [16]

    Llms can find mathematical reasoning mistakes by pedagogical chain-of-thought.arXiv preprint arXiv:2405.06705, 2024

    Zhuoxuan Sun, Yanjun Zhang, Jiali Li, Yiwei Han, Hainiu Lei, Liang Zhang, Chengqing Qin, and Huaxiu Zhou. Llms can find mathematical reasoning mistakes by pedagogical chain-of-thought.arXiv preprint arXiv:2405.06705, 2024

    2024 arXiv

  9. [17]

    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

  10. [18]

    Tem- poralizing confidence: Evaluation of chain-of-thought reasoning with signal temporal logic

    Peilin Yu, Zikang Li, Yifei Zhang, Shaoyuan Chen, Calin Belta, and Yewen Zhang. Tem- poralizing confidence: Evaluation of chain-of-thought reasoning with signal temporal logic. arXiv preprint arXiv:2506.08243, 2025

    2025 arXiv

  11. [19]

    A post-training enhanced optimization approach for small language models

    Keke Zhai. A post-training enhanced optimization approach for small language models. arXiv preprint arXiv:2411.02939, 2024

    2024 arXiv

  12. [20]

    Comat: Chain of mathematically annotated thought improves mathematical reasoning

    Zehan Zhang, Xin Xu, Momo Yang, Yuan Cai, Linjie Chen, Shuai Jiang, and Weiwei Su. Comat: Chain of mathematically annotated thought improves mathematical reasoning. arXiv preprint arXiv:2410.10336, 2024

    2024

  13. [21]

    Beyond efficiency: A sys- tematic survey of resource-efficient large language models.arXiv preprint arXiv:2401.00625, 2024

    Kai Zhao, Yunfei Wang, Liang Chen, Zhe Liu, and Min Zhang. Beyond efficiency: A sys- tematic survey of resource-efficient large language models.arXiv preprint arXiv:2401.00625, 2024. 16

    2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.