Pith. sign in

REVIEW 4 major objections 5 minor 26 references

SciDA: Scientific Dynamic Assessor of LLMs

T0 review · 4 major / 5 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 Randomly re-initializing problem numbers cuts LLM numerical-reasoning scores by 20–60%, which the paper reads as widespread benchmark memorization.

desk verdict Useful new dynamic benchmark, but the paper overclaims contamination from the observed accuracy drop; the private-problem control is in the appendix and never used. read the letter →

arxiv 2506.12909 v1 pith:2XIJ3KSG submitted 2025-06-15 cs.CL

Junting Zhou , Tingjia Miao , Yiyan Liao , Qichao Wang , Zhoufutu Wen , Yanqin Wang , Yunjie Huang , Ge Yan
show 10 more authors
Leqi Wang Yucheng Xia Hongwan Gao Yuansong Zeng Renjie Zheng Chen Dun Yitao Liang Tong Yang Wenhao Huang Ge Zhang
This is my paper
classification cs.CL
keywords largelanguagemodelsdatacontaminationbenchmarkevaluationdynamicinitializationnumericalreasoningolympiadproblemsscientificcodeinterpreter
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 sets out to show that static evaluation benchmarks overestimate how well large language models do numerical reasoning. It proposes SciDA, a set of over 1,000 olympiad-level computation problems in mathematics, physics, chemistry, and biology whose numerical parameters are randomly re-initialized at every inference round. Across 14 closed- and open-weight models, accuracy falls from roughly 20–50% on fixed versions to roughly 10–35% on randomized versions, a relative drop of 20–60%. The authors take this drop as evidence of widespread data contamination and argue that randomized-parameter evaluation gives a more truthful measure of reasoning ability, so frontier models' apparent numerical skill may be substantially memorized number patterns rather than generalizable problem-solving.

What carries the argument

The carrying mechanism is the dynamic initialization pipeline. Each problem is annotated by marking variable parameters with '$' tokens, converted into a Python function that maps those parameters to a numerical answer, and assigned scientifically valid ranges for each variable. At evaluation time the variables are sampled anew from those ranges, so the answer changes every round and no fixed number pattern can be matched from memory. Correctness is judged by comparing the model's numerical output to the function's output within a tolerance, and the paper uses five random initializations per problem, selected by an elbow-point analysis of variance.

What would settle it

Give the same fixed and randomized problems to human olympiad contestants or a symbolic algebra solver: if their accuracy falls by a comparable margin, the drop cannot be attributed to LLM memorization. Alternatively, run membership-inference or exact-match retrieval on the original fixed instances; finding no training-data matches would undercut the contamination explanation.

Watch

Extended reading notes

Core claim

The central claim is that LLMs have memorized the numerical patterns of widely used benchmarks, and that this shows up as a systematic accuracy collapse when the same problems are re-posed with randomized parameters. On SciDA, every tested model loses accuracy under random initialization: high-end models fall from about 50% to about 35%, older GPT-4o models from about 25% to about 10%, with mathematics and physics dropping more than biology and chemistry. The paper concludes that current fixed benchmarks are contaminated and that dynamic numerical initialization removes the memorization crutch, yielding truthful and unbiased assessments of numerical reasoning.

Load-bearing premise

The argument assumes the accuracy drop under random initialization is caused by memorized original parameters, not by the randomized problems being intrinsically harder or less familiar.

Editorial extensions

If this is right

  • Fixed-number benchmarks should be treated as likely contaminated until shown otherwise, and evaluations of numerical reasoning should randomize parameters.
  • The discipline-dependent drop suggests contamination is uneven, so per-subject scores matter more than a single aggregate number.
  • Code-interpreter support becomes a confound in benchmarks; separating internal arithmetic from external computation is necessary for fair comparison.
  • Dynamic benchmarks are a reusable tool for training and evaluating models, including reinforcement-learning settings, without rewards being gamed by memorized answers.

Reading between the lines

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

  • Editorial inference: the causal leap from accuracy drop to contamination is not tested; a human or non-memorizing numeric-solver baseline on the same fixed-versus-random problems would separate memorization from intrinsic difficulty.
  • Editorial inference: randomizing only numeric values leaves problem wording untouched, so the benchmark isolates a narrow form of memorization; structural randomization of formulas, diagrams, or problem setups would test reasoning more broadly.
  • Editorial inference: the released data could be paired with membership-inference or exact-retrieval probes to verify memorization directly, rather than inferring it from performance deltas.
  • Editorial inference: if the contamination reading is right, historical model rankings on fixed benchmarks may need revision, and dynamic benchmarks could become the default for numerical-reasoning claims.
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 introduces SciDA, a multidisciplinary benchmark of about 1,000 olympiad-level numerical computation problems with randomized variable initialization. The authors evaluate 14 LLMs under fixed ('initial') and randomized parameters, reporting a 20-60% relative accuracy drop under randomization, which they interpret as strong evidence of data contamination. They also analyze error types and the effect of a code interpreter, concluding that SciDA provides truthful and unbiased assessments of scientific numerical reasoning. The dataset is released on Hugging Face.

Significance. If the contamination attribution holds, the dynamic-initialization design addresses a real weakness of static benchmarks and the released dataset is a useful community resource. The paper's strengths include the multi-discipline coverage (math, physics, chemistry, biology), expert annotation, a released dataset, an elbow-point analysis to choose the number of random initializations, and a concrete Code Interpreter ablation. However, the central causal claim—that the performance drop under randomization is caused by memorization of fixed parameter values—is not directly tested, and the benchmark's own private-problem subset provides an unused control. The correctness criterion is also underspecified, and no statistical uncertainty is reported. These issues, not the benchmark itself, are the main barriers to the paper's conclusions.

major comments (4)
  1. [Section 6, Section 4, Appendix B] The central claim that the 20-60% relative accuracy drop 'strongly indicates the widespread presence of data contamination' is an untested causal attribution. The observed drop could be caused by the randomized variants being intrinsically harder—for example, uniformly sampled parameters can produce less 'nice' numbers and arbitrary decimals that increase arithmetic difficulty independently of memorization. The paper contains a natural control that is never reported: Appendix B states that over 20% of problems are privately held or original contributions from medalists, coaches, and professors, for which the fixed parameters cannot have appeared in pretraining data. The contamination hypothesis predicts a much smaller initial-vs-random gap on that subset, yet no split by source is given. Without this control, or an alternative difficulty-matched baseline (e.g., human solver performance or a non-neural solver on both variants), the observed drop does not discriminate between contamination and distribution-shift difficulty. This should be addressed before the conclusion in Section 6 can be supported.
  2. [Section 3.1] The correctness criterion is not operationalized: the text says an answer is correct if its deviation from the true answer is 'within a prescribed tolerance,' but the tolerance value, whether it is absolute or relative, and the parsing rules for model outputs are never specified. This is load-bearing because the measured accuracy drops depend entirely on how strictly answers are graded; a loose tolerance could mask genuine reasoning failures, while a tight tolerance could inflate the apparent contamination effect if randomized answers are harder to format. The exact criterion must be reported for reproducibility and for the quantitative claims in Table 1.
  3. [Section 4, Table 1, Appendix C] No error bars, confidence intervals, or significance tests are reported for any of the accuracy numbers. With 1000 problems and 5 random initializations, the differences between models and the claimed 20-60% relative drops may be within noise, especially for small cells such as hard biology (50 questions) or hard chemistry experiments. The elbow-point analysis in Appendix C uses only two models and reports a visual 'elbow' without a formal selection criterion; the decision n=5 should be justified with a quantitative rule. Please add bootstrap or other uncertainty estimates for the average and per-discipline accuracies.
  4. [Section 5.1, Figure 4] The error-type analysis interprets the dominance of calculation errors as evidence that the models 'were likely trained on larger corpora' and that the 'main bottleneck is computational capacity,' but this interpretation is also consistent with the randomized variants being harder to compute, regardless of contamination. This is not an independent piece of evidence for the contamination hypothesis; the text should either acknowledge the ambiguity or provide a control (e.g., comparing error-type distributions on private vs. public problems).
minor comments (5)
  1. [Abstract and Section 6] The phrase 'truthful and unbiased assessments' is too strong given the unresolved attribution issue; a more cautious phrasing such as 'less susceptible to memorized numerical patterns' would better match the evidence.
  2. [Appendix B] There are several typos in the source list, e.g., 'Chinese Mathematical Olympiad (CPhO)' should likely be 'Chinese Physics Olympiad', 'Plysics Prolens and Solutons' should be 'Physics Problems and Solutions', and 'Mechenics' should be 'Mechanics'. Please proofread the appendix.
  3. [Section 2.2] The related-work discussion of Math-perturb appears in the text but the citation formatting is inconsistent with the rest of the references; consider standardizing reference styles.
  4. [Table 2] The table layout with repeated model names and subject rows is hard to read; consider using a long-format table or color-coding to distinguish models and subjects.
  5. [Figure 3] The example problem's variable names, ranges, and the Python function are presented as three panels but the connection between them is not explicitly described in the caption; please add a short explanation of how the instruction, arguments, and code correspond.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SciDA is an empirical benchmark; the contamination attribution is an untested interpretation, not a derivation that reduces to its inputs.

full rationale

SciDA is an empirical benchmark paper rather than a derivation chain. The benchmark is constructed by collecting Olympiad-level problems, annotating variables, and writing Python solution code; the reported accuracy drop under random initialization is a measured observation, not a quantity defined to equal the contamination claim. There is no fitted parameter that is later renamed as a prediction, and no equation in the paper forces the accuracy drop to equal data contamination. The conclusion in Section 6 that the drop 'strongly indicates the widespread presence of data contamination within LLMs' is a causal interpretation that is not directly tested, but an unsupported interpretation is a soundness or validity issue, not circularity. The paper does not rely on a load-bearing self-citation: citations to Omni-MATH, KORgym, and SuperGPQA are related-work comparisons and are not used to justify the central result. No uniqueness theorem or ansatz is imported from prior work by the authors. The natural control suggested by Appendix B, where over 20% of problems are privately held or original and therefore cannot have appeared in pretraining data, is not reported; this is a missing experimental control for the contamination interpretation, not a circular step. Because the paper's claims are empirical and self-contained against external model evaluations, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The benchmark rests on assumptions about the correctness of annotator-written answer functions, the comparability of fixed and randomized problem instances, and the causal interpretation of performance drops as contamination. None of these are verified directly in the paper.

free parameters (3)
  • Answer tolerance = unspecified
    Section 3.1 defines correctness as deviation within a prescribed tolerance but never states the value used in evaluation, making results hard to interpret or reproduce.
  • Random variable ranges [a,b] = hand-chosen per variable
    Ranges are described as 'scientifically valid' but are not justified with data; changing ranges could alter difficulty and the magnitude of the performance drop.
  • Random initialization count n=5 = 5
    Selected via elbow point analysis on two models (Appendix C); this is a hyperparameter that affects the stability of accuracy estimates.
assumptions (4)
  • domain assumption Python code written by annotators correctly computes the intended answer.
    Ground-truth answers are generated by this code; any bug or mismatch with the problem text would propagate to labels.
  • domain assumption The original fixed parameters of sourced problems are the ones potentially contaminated.
    The paper assumes that initial parameters correspond to known instances and that LLMs may have memorized specific values, but this is not verified.
  • domain assumption Randomly initialized problems have comparable intrinsic difficulty to fixed-parameter problems.
    If random variants are harder independent of contamination, the accuracy drop is confounded; the paper does not test this with human or solver baselines.
  • ad hoc to paper The performance drop under randomization is evidence of data contamination.
    This is the paper's interpretive leap; no direct contamination detection is performed, so the assumption is not independently grounded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SciDA: Scientific Dynamic Assessor of LLMs." pith.science (2026). https://pith.science/paper/2XIJ3KSG

@misc{pith2026250612909,
  author       = {Pith},
  title        = {Pith review of: SciDA: Scientific Dynamic Assessor of LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XIJ3KSG}},
  note         = {Machine review of arXiv:2506.12909}
}
read the original abstract

Advancement in Large Language Models (LLMs) reasoning capabilities enables them to solve scientific problems with enhanced efficacy. Thereby, a high-quality benchmark for comprehensive and appropriate assessment holds significance, while existing ones either confront the risk of data contamination or lack involved disciplines. To be specific, due to the data source overlap of LLMs training and static benchmark, the keys or number pattern of answers inadvertently memorized (i.e. data contamination), leading to systematic overestimation of their reasoning capabilities, especially numerical reasoning. We propose SciDA, a multidisciplinary benchmark that consists exclusively of over 1k Olympic-level numerical computation problems, allowing randomized numerical initializations for each inference round to avoid reliance on fixed numerical patterns. We conduct a series of experiments with both closed-source and open-source top-performing LLMs, and it is observed that the performance of LLMs drop significantly under random numerical initialization. Thus, we provide truthful and unbiased assessments of the numerical reasoning capabilities of LLMs. The data is available at https://huggingface.co/datasets/m-a-p/SciDA

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 5 canonical work pages

  1. [1]

    Training verifiers to solve math word problems, 2021

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. URLhttps://arxiv.org/abs/2110.14168

    arXiv 2021

  2. [2]

    Investigating data contamination in modern benchmarks for large language models.arXiv preprint arXiv:2311.09783, 2023

    Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Investigating data contamination in modern benchmarks for large language models.arXiv preprint arXiv:2311.09783, 2023

    arXiv 2023

  3. [3]

    Generalization or memorization: Data contamination and trustworthy evaluation for large language models.arXiv preprint arXiv:2402.15938, 2024

    Yihong Dong, Xue Jiang, Huanyu Liu, Zhi Jin, Bin Gu, Mengfei Yang, and Ge Li. Generalization or memorization: Data contamination and trustworthy evaluation for large language models.arXiv preprint arXiv:2402.15938, 2024

    arXiv 2024

  4. [4]

    Omni-math: A universal olympiad level mathematic benchmark for large language models, 2024

    Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, Zhengyang Tang, Benyou Wang, Daoguang Zan, Shanghaoran Quan, Ge Zhang, Lei Sha, Yichang Zhang, Xuancheng Ren, Tianyu Liu, and Baobao Chang. Omni-math: A universal olympiad level mathematic benchmark for large language models, 2024. URLhttps://...

    arXiv 2024

  5. [5]

    Time travel in llms: Tracing data contamination in large language models

    Shahriar Golchin and Mihai Surdeanu. Time travel in llms: Tracing data contamination in large language models. arXiv preprint arXiv:2308.08493, 2023

    arXiv 2023

  6. [6]

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems.arXiv preprint arXiv:2402.14008, 2024

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng 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.arXiv preprint arXiv:2402.14008, 2024

    arXiv 2024

  7. [7]

    Measuring massive multitask language understanding, 2021

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding, 2021. URLhttps://arxiv.org/abs/2009.03300

    arXiv 2021

  8. [8]

    Measuring mathematical problem solving with the math dataset, 2021

    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, 2021. URLhttps://arxiv.org/ abs/2103.03874

    arXiv 2021

Show all 26 references
  1. [9]

    Math-perturb: Benchmarking llms’ math reasoning abilities against hard perturbations, 2025

    Kaixuan Huang, Jiacheng Guo, Zihao Li, Xiang Ji, Jiawei Ge, Wenzhe Li, Yingqing Guo, Tianle Cai, Hui Yuan, Runzhe Wang, Yue Wu, Ming Yin, Shange Tang, Yangsibo Huang, Chi Jin, Xinyun Chen, Chiyuan Zhang, and Mengdi Wang. Math-perturb: Benchmarking llms’ math reasoning abilitie...

    2025 arXiv

  2. [10]

    Olympicarena: Benchmarking multi-discipline cognitive reasoning for superintelligent ai, 2025

    Zhen Huang, Zengzhi Wang, Shijie Xia, Xuefeng Li, Haoyang Zou, Ruijie Xu, Run-Ze Fan, Lyumanshan Ye, Ethan Chern, Yixin Ye, Yikai Zhang, Yuqing Yang, Ting Wu, Binjie Wang, Shichao Sun, Yang Xiao, Yiyuan Li, Fan Zhou, Steffi Chern, Yiwei Qin, Yan Ma, Jiadi Su, Yixiu Liu, Yuxian...

    2025 arXiv

  3. [11]

    Livecodebench: Holistic and contamination free evaluation of large language models for code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar- Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. InThe Thirteenth International Conference on Learn...

    2025

  4. [12]

    Mathbench: Evaluating the theory and application proficiency of llms with a hierarchical mathematics benchmark.arXiv preprint arXiv:2405.12209, 2024

    Hongwei Liu, Zilong Zheng, Yuxuan Qiao, Haodong Duan, Zhiwei Fei, Fengzhe Zhou, Wenwei Zhang, Songyang Zhang, Dahua Lin, and Kai Chen. Mathbench: Evaluating the theory and application proficiency of llms with a hierarchical mathematics benchmark.arXiv preprint arXiv:2405.12209, 2024

    2024 arXiv

  5. [13]

    Varbench: Robust language model benchmarking through dynamic variable perturbation.arXiv preprint arXiv:2406.17681, 2024

    Kun Qian, Shunji Wan, Claudia Tang, Youzhi Wang, Xuanming Zhang, Maximillian Chen, and Zhou Yu. Varbench: Robust language model benchmarking through dynamic variable perturbation.arXiv preprint arXiv:2406.17681, 2024

    2024 arXiv

  6. [14]

    Phybench: Holistic evaluation of physical perception and reasoning in large language models, 2025

    Shi Qiu, Shaoyang Guo, Zhuo-Yang Song, Yunbo Sun, Zeyu Cai, Jiashen Wei, Tianyu Luo, Yixuan Yin, Haoxu Zhang, Yi Hu, Chenyang Wang, Chencheng Tang, Haoling Chang, Qi Liu, Ziheng Zhou, Tianyu Zhang, Jingtian Zhang, Zhangyi Liu, Minghao Li, Yuku Zhang, Boxuan Jing, Xianqi Yin, Y...

    2025 arXiv

  7. [15]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level google-proof q&a benchmark. InFirst Conference on Language Modeling, 2024. URLhttps://openreview.net/forum?id=Ti67584b98

    2024

  8. [16]

    Korgym: A dynamic game platform for llm reasoning evaluation, 2025

    Jiajun Shi, Jian Yang, Jiaheng Liu, Xingyuan Bu, Jiangjie Chen, Junting Zhou, Kaijing Ma, Zhoufutu Wen, Bingli Wang, Yancheng He, Liang Song, Hualei Zhu, Shilong Li, Xingjian Wang, Wei Zhang, Ruibin Yuan, Yifan Yao, Wenjun Yang, Yunli Wang, Siyuan Fang, Siyu Yuan, Qianyu He, X...

    2025 arXiv

  9. [17]

    Challenging the boundaries of reasoning: An olympiad-level math benchmark for large language models, 2025

    Haoxiang Sun, Yingqian Min, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. Challenging the boundaries of reasoning: An olympiad-level math benchmark for large language models, 2025. URLhttps://arxiv.org/abs/2503.21380

    2025 arXiv

  10. [18]

    Scieval: A multi-level large language model evaluation benchmark for scientific research

    Liangtai Sun, Yang Han, Zihan Zhao, Da Ma, Zhennan Shen, Baocai Chen, Lu Chen, and Kai Yu. Scieval: A multi-level large language model evaluation benchmark for scientific research. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19053–19061, 2024

    2024

  11. [19]

    Mathscale: Scaling instruction tuning for mathematical reasoning

    Zhengyang Tang, Xingxing Zhang, Benyou Wang, and Furu Wei. Mathscale: Scaling instruction tuning for mathematical reasoning. arXiv preprint arXiv:2403.02884, 2024

    2024 arXiv

  12. [20]

    Supergpqa: Scaling llm evaluation across 285 graduate disciplines, 2025

    P Team, Xinrun Du, Yifan Yao, Kaijing Ma, Bingli Wang, Tianyu Zheng, King Zhu, Minghao Liu, Yiming Liang, Xiaolong Jin, Zhenlin Wei, Chujie Zheng, Kaixin Deng, Shawn Gavin, Shian Jia, Sichao Jiang, Yiyan Liao, Rui Li, Qinrui Li, Sirun Li, Yizhi Li, Yunwen Li, David Ma, Yuanshe...

    2025 arXiv

  13. [21]

    Scibench: Evaluating college-level scientific problem-solving abilities of large language models

    Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. arXiv preprint arXiv:2307.10635, 2023

    2023 arXiv

  14. [22]

    MMLU-pro: A more robust and challenging multi-task language understanding benchmark

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. MMLU-pro: A more robust and challenging multi-task language unders...

    2024

  15. [23]

    Agieval: A human-centric benchmark for evaluating foundation models

    Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. Agieval: A human-centric benchmark for evaluating foundation models. arXiv preprint arXiv:2304.06364, 2023. 11 Appendix A Contributor & Acknowledgement Junting Zho...

    2023 arXiv

  16. [24]

    Mathematics: International Mathematical Olympiad (IMO), Chinese Mathematical Olympiad (CMO), Problems in Mathematical Analysis by B. P. Demidovich, Euler Math, etc

  17. [25]

    Griffiths, Electrodynamics Solution Manual by David J

    Physics: International Physics Olympiad (IPhO), Chinese Mathematical Olympiad (CPhO), Interna- tional Physics Olympiad Training and Selection by Yongling Zheng, Collection of Physics Challenges by Yousheng Shu et al., A Grand Dictionary of Plysics Prolens and Solutons by Yongd...

  18. [26]

    elbow point

    Chemistry: International Chemistry Olympiad (IChO), Chinese Chemistry Olympiad (CChO), Physical Chemistry by Peter Atkins, etc. 4.Biology: International Biology Olympiad (IBO), Chinese National Biology Olympiad (CNBO), etc. High-quality privately held or original problems cons...

Pith tools

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