Pith. sign in

REVIEW 6 major objections 6 minor 53 references

Establishing Trustworthy LLM Evaluation via Shortcut Neuron Analysis

T0 review · 6 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 This paper claims that inflated benchmark scores from data contamination are carried by a sparse set of "shortcut neurons," and that replacing those neurons' activations with a clean base model's activations during evaluation restores a…

desk verdict A plausible contamination-robust evaluation method with strong external correlations, but the clean-base counterfactual is an untested assumption that should be probed before the method is used as a trust anchor. read the letter →

arxiv 2506.04142 v1 pith:4PFNWQX3 submitted 2025-06-04 cs.CL

Kejian Zhu , Shangqing Tu , Zhuoran Jin , Lei Hou , Juanzi Li , Jun Zhao This is my paper
classification cs.CL
keywords datacontaminationtrustworthyevaluationshortcutneuronsactivationpatchingLLMinterpretabilitybenchmarkoverestimationcausalanalysis
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 argues that inflated benchmark scores from data contamination are not just statistical artifacts but a mechanism-level phenomenon: fine-tuning on contaminated data makes a sparse set of neurons encode shortcuts the model uses to answer benchmark questions. The authors propose identifying these "shortcut neurons" by comparing activations of contaminated and uncontaminated models and by causal activation-patching experiments, then patching the identified neurons with base-model activations during evaluation to suppress the shortcuts. They claim this restores models' true capabilities: contaminated models' scores drop toward uncontaminated levels, uncontaminated models are barely affected, and patched scores track the independent MixEval benchmark with a Spearman coefficient exceeding 0.95. If correct, this offers a way to evaluate LLMs on existing public benchmarks despite contamination, avoiding the cost of continually building new dynamic benchmarks.

What carries the argument

The central object is the "shortcut neuron": a neuron in the transformer's MLP, defined as an activation before down-projection, whose activation diverges between contaminated and uncontaminated models and whose causal patching disproportionately hurts the contaminated model's performance. The locating method combines a comparative score, the root-mean-square difference of last-token activations between $M_{\mathrm{con}}$ and $M_{\mathrm{un}}$ on benchmark samples, with a causal score computed by dynamic activation patching, $C_N = a(M_{\mathrm{con}}) - a_{\mathrm{patch}}(M_{\mathrm{con}}|M_0) + 1 - (a(M_{\mathrm{un}}) - a_{\mathrm{patch}}(M_{\mathrm{un}}|M_0))$. The patching itself replaces the identified neurons' activations with cached activations from the base model $M_0$ during generation. The machinery's job is to pinpoint the minimal set of parameters that carries contamination-induced overestimation so it can be suppressed without damaging general capabilities.

What would settle it

Take a model independently known to be contaminated on GSM8K and patch it using a base model that itself shows elevated memorization of GSM8K samples; if the patched score remains well above the uncontaminated reference level, or if substituting a different base model changes the ranking of a set of real-world models, then the claim that patching restores true capability is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a sparse set of roughly 5,000 MLP neurons in 7B-parameter transformers carries the "shortcut solutions" acquired when a model is fine-tuned on contaminated benchmark data. These neurons are located by two scores: a comparative score measuring activation divergence between a contaminated model and an uncontaminated model on the same samples, and a causal score measuring how much patching a neuron's activation with the base model's activation lowers the contaminated model's accuracy while leaving the uncontaminated model's accuracy intact. Replacing the activations of the top-scoring neurons during decoding suppresses both answer-level shortcuts and input-format shortcuts; the patched scores correlate with MixEval at a Spearman coefficient above 0.95, which the paper takes as evidence that the method reveals genuine capability rather than memorization.

Load-bearing premise

The load-bearing premise is that the base model used as the patching reference is itself uncontaminated, so its activations are a neutral clean baseline; if the base model also contains traces of the benchmark, patching would inject those traces rather than remove them, as the paper explicitly concedes in its Limitations section.

Editorial extensions

If this is right

  • Contaminated models' original scores overestimate true capability; after patching, accuracy drops by up to tens of points in simulation, for example 80.0 to 30.2 for the 5x-GSM-i LLaMA2 model.
  • Models fine-tuned on the benchmark's input format, such as GSM8K Train, also drop after patching, showing the method targets input-format shortcuts rather than only memorized answers.
  • Patching leaves scores on MAWPS and MMLU essentially unchanged, indicating that genuine reasoning ability is preserved.
  • Patched scores correlate with the OpenMathInstruct-2 reference at 0.970 and with MixEval at 0.957, indicating alignment with real-world model performance.
  • The method generalizes across math benchmarks, architectures (LLaMA2, Mistral, LLaMA3), hyperparameters, and a non-math ARC-Challenge setting, at a computational cost far below maintaining dynamic benchmarks.

Reading between the lines

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

  • Inference: the method's shortcut-neuron definition is relative to a particular uncontaminated reference model and a particular fine-tuning-based contamination simulation, so models contaminated through continued pretraining or other pipelines may carry shortcuts in different neurons than those flagged by the paper's SFT-based comparison.
  • Inference: because the base model's activations are the neutral reference, patching can at best recenter scores to that base model's baseline; if a base model has latent contamination for a benchmark, patching could embed that contamination into the evaluated scores instead of removing it.
  • Inference: the gap between original and patched scores, which the paper observes is near zero for uncontaminated models and large for some real-world models, could be developed into a contamination detector for arbitrary model-benchmark pairs.
  • Inference: the sparsity of the shortcut neurons, roughly 1.4% of LLaMA2-7B parameters, suggests that benchmark-specific memorization is localized enough that freezing or pruning those neurons during fine-tuning might prevent contamination inflation at the source.
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

6 major / 6 minor

Summary. The paper proposes an evaluation method for LLMs that aims to reduce the effect of benchmark contamination by identifying and patching 'shortcut neurons'. Shortcut neurons are selected as the top 5000 neurons ranked by a combination of a comparative activation distance (Eq. 3) and a causal score (Eq. 4) computed by dynamic activation patching from a base model. The authors report that patching with base-model activations reduces the accuracy of fine-tuned contaminated models on GSM8K and other math benchmarks while leaving uncontaminated models roughly unchanged, and that patched scores correlate strongly with MixEval (Spearman 0.957) and OpenMathInstruct-2 (0.970). Generalization experiments cover other benchmarks, architectures (LLaMA3-8B), and hyperparameters.

Significance. If the method works as claimed, it offers a low-cost alternative to dynamic benchmarks and a mechanistic account of contamination. The external validation against MixEval is a genuine strength, as is the code release and the breadth of generalization experiments (Tables 4–6, 11–12). The paper is honest about its main assumption in the Limitations section. However, the central interpretative claim—that the identified neurons are 'shortcut' neurons encoding overfitting—is partly definitional, because the causal score in Eq. (4) defines them by their patching effect. The method's trustworthiness therefore rests on the strength of the external correlation and on the cleanliness of the base model M0, both of which need more support.

major comments (6)
  1. [Limitations, point (2); Eq. (4); §3.4] The entire pipeline uses M0 activations as the clean counterfactual. If M0 is itself contaminated on the target benchmark, then the causal score measures differences between two contaminated states, and patching injects M0's own shortcuts. The manuscript acknowledges this assumption but does not test it. Please add an audit of M0 (e.g., perplexity or memorization checks, or a control using a deliberately contaminated M0) and report how the results change. This is load-bearing for the trustworthiness claim.
  2. [§4.2, Figure 3] The number of patched neurons (5000) is chosen by observing accuracy curves on the very same simulated contaminated and uncontaminated models used in the evaluation. This makes the drop in contaminated accuracy partly an artifact of the selection. Please report the correlation with MixEval as a function of the number of patched neurons, or select k by cross-validation on a subset of models and evaluate on the rest; without this, the generalization to real-world models is not demonstrated.
  3. [§4.3, Figure 4] The Spearman correlations of 0.970 and 0.957 are reported for patched scores, but the correlation between the original (unpatched) scores and the reference benchmarks is not given. If the original scores already correlate highly with MixEval, the improvement from patching is small and the claim that patching reveals true capabilities is weakened. Please report the same correlation and RMSE for original scores.
  4. [Eq. (4), §4.3] The causal score is a sum of two terms with an unexplained additive constant (1) and no weighting. Moreover, no confidence intervals or significance tests are provided for the Spearman coefficients. With the number of real-world models listed in Table 9 (19 points), a Spearman of 0.957 is strong but should be accompanied by a confidence interval and a comparison against a null model.
  5. [§3.3, Eq. (4)] A neuron is defined as a shortcut neuron if patching it reduces contaminated-model accuracy while not affecting uncontaminated-model accuracy. The subsequent discovery that contaminated models contain such neurons is therefore in part true by construction. The paper should separate the discovery claim (sparse neurons exist that affect contaminated but not uncontaminated models) from the mechanistic interpretation (these are 'shortcut solutions' acquired in training). At minimum, show that the located top-5000 neurons behave differently from random neurons at the same count in terms of correlation with MixEval, not just in terms of accuracy curves (Figure 3).
  6. [Table 2, Table 3] Uncontaminated models fine-tuned on GSM8K Train show drops of −6.5 and −4.3 on LLaMA/Mistral, but MAWPS/MMLU scores sometimes increase after patching. The paper interprets the drop as suppression of input-format shortcuts, but given that GSM8K Train is legitimate training data, the drop could also reflect suppression of genuine learned reasoning. Please discuss this alternative explanation and provide per-instance analysis.
minor comments (6)
  1. [Abstract and §4.3] The abstract says 'strong linear correlation' but the reported statistic is Spearman's rank correlation; please make the distinction between rank and linear correlation explicit throughout.
  2. [Eq. (2)] The notation m^l_ij is used to define neuron v^l_j, but m^l_ij is not defined elsewhere; please clarify the relationship between the coefficient and the neuron activation.
  3. [§3.3] The sentence 'Similar to 3.2, we use M0 as the patching model' is followed by Eq. (4) without a formal definition of a(M); please define the accuracy notation before the equation.
  4. [Table 2] The text states an average drop of 37% and an average change of 3%; please specify the set of rows over which these averages are computed and whether they are absolute changes.
  5. [Figure 1] The caption contains the fragment 'Answer is 0' and the label 'Shortcut Path / Patching Path' is unclear; please redraw or re-caption for readability.
  6. [References] Several references are incomplete or incorrectly formatted, e.g., 'Brown, 2020', 'Vaswani, 2017', and 'Geva et al., 2020' do not follow the standard style; please fix them.

Circularity Check

2 steps flagged · score 5.0 of 10

In-sample shortcut selection and the 5000-neuron cutoff make the patching 'drop' partly forced; external MixEval correlation supplies independent grounding, so the circularity is moderate.

  1. self definitional [Section 3.3, Eq. (4); Section 3.4, paragraph 2]
    "CN = a(Mcon) − a_patch(Mcon|M0) + 1 − (a(Mun) − a_patch(Mun|M0)) (4) ... We replace the activations of shortcut neurons in model to be evaluated Me with those in base model M0, so as to suppress the contaminated model from shortcut reasoning."

    Eq. (4) assigns a neuron a high causal score exactly when patching the base model's activations into the contaminated model lowers its accuracy on D while changing the uncontaminated model little. The method then patches the highest-scoring neurons and reports, as its main evidence, that contaminated models drop after patching and uncontaminated models do not. That outcome is the ranking objective restated as a result. The MixEval correlation is external, but the internal mechanistic claim 'shortcut neurons cause overestimation' is not independently tested on data separate from the neuron-selection criterion.

  2. fitted input called prediction [Section 4.2, Figure 3; Section 4.3, first paragraph]
    "Figure 3 shows that after 5,000 neurons were patched, the accuracy of the contaminated model has roughly reached the same level as the uncontaminated model, and the accuracy of the uncontaminated model has changed very little. After 20,000 neurons are patched, the accuracy of both models begins to decline. This result shows that the first 5,000 neurons have a good effect on alleviating model contamination. ... Following the finding above, shortcut neurons are selected as the top 5000 neurons."

    The 5,000-neuron cutoff is read off accuracy-versus-neuron-count curves on the same contaminated and uncontaminated models (Figure 3), and the same curves are then summarised in Section 4.3 as the finding that patched contaminated models drop and uncontaminated models change little. The threshold is therefore fitted to the data whose change it is used to report; those numbers are not out-of-sample predictions. The generalisation tables and MixEval correlation show transfer, but the headline drop is in-sample.

full rationale

The paper has genuine but partial circularity. The shortcut-neuron score in Eq. (4) is defined as the patching-induced accuracy drop on the contaminated model plus preservation of the uncontaminated model, and the trustworthy-evaluation score is then the accuracy after patching those very neurons; the Table 2 drop is thus partly the selection objective restated. Similarly, the 5000-neuron cutoff is selected from the same accuracy curves whose changes are later reported as the method's effectiveness. These in-sample components would be more circular without the external anchors: the paper validates the resulting scores against OpenMathInstruct-2 (Spearman 0.970) and MixEval (Spearman 0.957) on real-world models, giving independent content to the claim that the patched scores track true capabilities. The Limitations section explicitly concedes the load-bearing premise 'we assume that base models are uncontaminated, but in fact, even base models cannot completely eliminate the suspicion of contamination'; this is an unverified assumption about the counterfactual reference rather than a logical circularity, so it affects validity and robustness more than the circularity score. There is no load-bearing self-citation chain here: prior work by the authors is background, and the main validation is against an external benchmark. On balance, the internal mechanism evidence is partly forced by construction, but the central practical claim is externally anchored.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The method rests on several unverified assumptions: a clean base model, a simulation that faithfully mimics contamination, and the causal validity of activation patching. The main hand-tuned quantities are the top-5000 neuron count, the group size of 512, and the normalization constant in the causal score formula. These choices are not derived from first principles and are partially calibrated on the evaluation data.

free parameters (3)
  • Number of patched shortcut neurons = 5000
    Chosen based on Figure 3 to make contaminated model accuracy drop to the uncontaminated level while preserving uncontaminated accuracy. This is a post hoc selection on the evaluation data.
  • Group size for causal scoring = 512
    Adjacent neurons are grouped to compute causal effects; no ablation or justification is given for the group size.
  • Causal score normalization constant = 1
    The constant 1 in Eq. (4) is introduced to make the uncontaminated term contribute negatively when patching hurts it; it is an ad hoc design choice.
assumptions (5)
  • domain assumption The base model M0 is uncontaminated, and its activations serve as a neutral reference for patching.
    Stated in Limitations: 'we assume that base models are uncontaminated, but in fact, even base models cannot completely eliminate the suspicion of contamination.' This underpins the entire patching procedure.
  • domain assumption Fine-tuning a base model on a subset of a benchmark (or paraphrases) produces shortcut neurons representative of real-world contamination.
    The method locates shortcut neurons from simulated contaminated/uncontaminated model pairs; transfer to real-world models assumes this simulation captures real contamination.
  • domain assumption Replacing activations of selected neurons with the base model's activations is a valid causal intervention that only affects shortcut behavior.
    Activation patching presupposes that neurons are independent and that substituting activations does not disrupt other capabilities.
  • domain assumption Last-token activation adequately captures the effect of the whole prompt for comparative analysis.
    Section 3.2 adopts last token activation based on prior work; no ablation for this choice is provided.
  • domain assumption The accuracy calculation and dynamic generation are valid for open-ended tasks.
    Dynamic patching is used for generation, but the paper does not specify exact decoding or answer extraction details.
invented entities (1)
  • Shortcut neurons independent evidence
    purpose: Neurons hypothesized to encode benchmark-specific shortcuts that inflate contaminated model scores.
    The paper provides a falsifiable handle: patched scores on GSM8K should correlate with MixEval, an external benchmark. The correlation (rho=0.957) is a test outside the training setup, though the neuron identity itself is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Establishing Trustworthy LLM Evaluation via Shortcut Neuron Analysis." pith.science (2026). https://pith.science/paper/4PFNWQX3

@misc{pith2026250604142,
  author       = {Pith},
  title        = {Pith review of: Establishing Trustworthy LLM Evaluation via Shortcut Neuron Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4PFNWQX3}},
  note         = {Machine review of arXiv:2506.04142}
}
abstract

The development of large language models (LLMs) depends on trustworthy evaluation. However, most current evaluations rely on public benchmarks, which are prone to data contamination issues that significantly compromise fairness. Previous researches have focused on constructing dynamic benchmarks to address contamination. However, continuously building new benchmarks is costly and cyclical. In this work, we aim to tackle contamination by analyzing the mechanisms of contaminated models themselves. Through our experiments, we discover that the overestimation of contaminated models is likely due to parameters acquiring shortcut solutions in training. We further propose a novel method for identifying shortcut neurons through comparative and causal analysis. Building on this, we introduce an evaluation method called shortcut neuron patching to suppress shortcut neurons. Experiments validate the effectiveness of our approach in mitigating contamination. Additionally, our evaluation results exhibit a strong linear correlation with MixEval, a recently released trustworthy benchmark, achieving a Spearman coefficient ($\rho$) exceeding 0.95. This high correlation indicates that our method closely reveals true capabilities of the models and is trustworthy. We conduct further experiments to demonstrate the generalizability of our method across various benchmarks and hyperparameter settings. Code: https://github.com/GaryStack/Trustworthy-Evaluation

Figures

Figures reproduced from arXiv: 2506.04142 by the authors.

Figure 1
Figure 1. An example illustrating the core principle of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview of our method. We employ neuron analysis to identify regions within the model that may be [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The performance of the contaminated and un [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Correlation between the trustworthy evaluation scores obtained by our method and the reference scores in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: A figure to demonstrate the generalizability [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 9 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Sebastian Bordt, Harsha Nori, and Rich Caruana. 2024. Elephants never forget: Testing language models for memorization of tabular data. arXiv preprint arXiv:2403.06644

    arXiv 2024

  3. [3]

    Tom B Brown. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165

    arXiv 2020

  4. [4]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. 2024. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1--45

    2024

  5. [5]

    Jianhui Chen, Xiaozhi Wang, Zijun Yao, Yushi Bai, Lei Hou, and Juanzi Li. 2024. Finding safety neurons in large language models. arXiv preprint arXiv:2406.14144

    arXiv 2024

  6. [6]

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457v1

    arXiv 2018

  7. [7]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  8. [8]

    Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. 2021. Knowledge neurons in pretrained transformers. arXiv preprint arXiv:2104.08696

    arXiv 2021

Show all 53 references
  1. [9]

    Jasper Dekoninck, Mark Niklas M \"u ller, Maximilian Baader, Marc Fischer, and Martin Vechev. 2024. Evading data contamination detection for language models is (too) easy. arXiv preprint arXiv:2402.02823

    2024 arXiv

  2. [10]

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

    2024 arXiv

  3. [11]

    Mor Geva, Avi Caciularu, Kevin Ro Wang, and Yoav Goldberg. 2022. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. arXiv preprint arXiv:2203.14680

    2022 arXiv

  4. [12]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2020. Transformer feed-forward layers are key-value memories. arXiv preprint arXiv:2012.14913

    2020 arXiv

  5. [13]

    Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, and Mor Geva. 2024. Patchscope: A unifying framework for inspecting hidden representations of language models. arXiv preprint arXiv:2401.06102

    2024 arXiv

  6. [14]

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

    2023 arXiv

  7. [15]

    Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Minlie Huang, Nan Duan, and Weizhu Chen. 2023. Tora: A tool-integrated reasoning agent for mathematical problem solving. arXiv preprint arXiv:2309.17452

    2023 arXiv

  8. [16]

    Zishan Guo, Renren Jin, Chuang Liu, Yufei Huang, Dan Shi, Linhao Yu, Yan Liu, Jiaxuan Li, Bojian Xiong, Deyi Xiong, et al. 2023. Evaluating large language models: A comprehensive survey. arXiv preprint arXiv:2310.19736

    2023 arXiv

  9. [17]

    Wes Gurnee, Theo Horsley, Zifan Carl Guo, Tara Rezaei Kheirkhah, Qinyi Sun, Will Hathaway, Neel Nanda, and Dimitris Bertsimas. 2024. Universal neurons in gpt2 language models. arXiv preprint arXiv:2401.12181

    2024 arXiv

  10. [18]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300

    2020 arXiv

  11. [19]

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

    2021 arXiv

  12. [20]

    Alon Jacovi, Avi Caciularu, Omer Goldman, and Yoav Goldberg. 2023. Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks. arXiv preprint arXiv:2305.10160

    2023 arXiv

  13. [21]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  14. [22]

    Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In Proceedings of the 2016 conference of the north american chapter of the association for computational linguistics: human language technologies,...

    2016

  15. [23]

    Yucheng Li. 2023. Estimating contamination via perplexity: Quantifying memorisation in language model evaluation. arXiv preprint arXiv:2309.10677

    2023 arXiv

  16. [24]

    Yucheng Li, Frank Guerin, and Chenghua Lin. 2024 a . Latesteval: Addressing data contamination in language model evaluation through dynamic and time-sensitive test construction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18600--18607

    2024

  17. [25]

    Yucheng Li, Yunhao Guo, Frank Guerin, and Chenghua Lin. 2024 b . An open-source data contamination report for large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 528--541

    2024

  18. [26]

    W Lian, B Goodson, E Pentland, et al. 2023. Openorca: An open dataset of gpt augmented flan reasoning traces

    2023

  19. [27]

    Robert Litschko, Max M \"u ller-Eberstein, Rob Van Der Goot, Leon Weber, and Barbara Plank. 2023. Establishing trustworthiness: Rethinking tasks and model evaluation. arXiv preprint arXiv:2310.05442

    2023 arXiv

  20. [28]

    Inbal Magar and Roy Schwartz. 2022. Data contamination: From memorization to exploitation. arXiv preprint arXiv:2203.08242

    2022 arXiv

  21. [29]

    Alexandre Matton, Tom Sherborne, Dennis Aumiller, Elena Tommasone, Milad Alizadeh, Jingyi He, Raymond Ma, Maxime Voisin, Ellen Gilsenan-McMahon, and Matthias Gall \'e . 2024. On leakage of code generation evaluation datasets. arXiv preprint arXiv:2407.07565

    2024 arXiv

  22. [30]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

    2022

  23. [31]

    Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2021. A diverse corpus for evaluating and developing english math word problem solvers. arXiv preprint arXiv:2106.15772

    2021 arXiv

  24. [32]

    Jinjie Ni, Fuzhao Xue, Xiang Yue, Yuntian Deng, Mahir Shah, Kabir Jain, Graham Neubig, and Yang You. 2024. Mixeval: Deriving wisdom of the crowd from llm benchmark mixtures. arXiv preprint arXiv:2406.06565

    2024 arXiv

  25. [33]

    Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021. Are nlp models really able to solve simple math word problems? arXiv preprint arXiv:2103.07191

    2021 arXiv

  26. [34]

    Oscar Sainz, Jon Ander Campos, Iker Garc \' a-Ferrero, Julen Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. 2023. Nlp evaluation in trouble: On the need to measure llm data contamination for each benchmark. arXiv preprint arXiv:2310.18018

    2023 arXiv

  27. [35]

    Shubham Toshniwal, Wei Du, Ivan Moshkov, Branislav Kisacanin, Alexan Ayrapetyan, and Igor Gitman. 2024. Openmathinstruct-2: Accelerating ai for math with massive open-source instruction data. arXiv preprint arXiv:2410.01560

    2024 arXiv

  28. [36]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  29. [37]

    Shangqing Tu, Kejian Zhu, Yushi Bai, Zijun Yao, Lei Hou, and Juanzi Li. 2024. Dice: Detecting in-distribution contamination in llm's fine-tuning phase for math reasoning. arXiv preprint arXiv:2406.04197

    2024 arXiv

  30. [38]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems

    2017

  31. [39]

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020 a . Investigating gender bias in language models using causal mediation analysis. Advances in neural information processing systems, 33:12388--12401

    2020

  32. [40]

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yoram Singer, and StuartM. Shieber. 2020 b . Investigating gender bias in language models using causal mediation analysis. Neural Information Processing Systems,Neural Information Processing Systems

    2020

  33. [41]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368

    2023 arXiv

  34. [42]

    Xiaozhi Wang, Kaiyue Wen, Zhengyan Zhang, Lei Hou, Zhiyuan Liu, and Juanzi Li. 2022. Finding skill neurons in pre-trained transformer-based language models. arXiv preprint arXiv:2211.07349

    2022 arXiv

  35. [43]

    Ruijie Xu, Zengzhi Wang, Run-Ze Fan, and Pengfei Liu. 2024. Benchmarking benchmark leakage in large language models. arXiv preprint arXiv:2404.18824

    2024 arXiv

  36. [44]

    Jifan Yu, Xiaozhi Wang, Shangqing Tu, Shulin Cao, Daniel Zhang-Li, Xin Lv, Hao Peng, Zijun Yao, Xiaohan Zhang, Hanming Li, et al. 2023. Kola: Carefully benchmarking world knowledge of large language models. arXiv preprint arXiv:2306.09296

    2023 arXiv

  37. [45]

    Zhuohao Yu, Chang Gao, Wenjin Yao, Yidong Wang, Wei Ye, Jindong Wang, Xing Xie, Yue Zhang, and Shikun Zhang. 2024. Kieval: A knowledge-grounded interactive evaluation framework for large language models. arXiv preprint arXiv:2402.15043

    2024 arXiv

  38. [46]

    Fred Zhang and Neel Nanda. 2023. Towards best practices of activation patching in language models: Metrics and methods. arXiv preprint arXiv:2309.16042

    2023 arXiv

  39. [47]

    Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. 2024. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology, 15(2):1--38

    2024

  40. [48]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223

    2023 arXiv

  41. [49]

    Kun Zhou, Yutao Zhu, Zhipeng Chen, Wentong Chen, Wayne Xin Zhao, Xu Chen, Yankai Lin, Ji-Rong Wen, and Jiawei Han. 2023. Don't make your llm an evaluation benchmark cheater. arXiv preprint arXiv:2311.01964

    2023 arXiv

  42. [50]

    Kaijie Zhu, Jiaao Chen, Jindong Wang, Neil Zhenqiang Gong, Diyi Yang, and Xing Xie. 2023 a . Dyval: Graph-informed dynamic evaluation of large language models. arXiv e-prints, pages arXiv--2309

    2023

  43. [51]

    Wenhong Zhu, Hongkun Hao, Zhiwei He, Yunze Song, Yumeng Zhang, Hanxu Hu, Yiran Wei, Rui Wang, and Hongyuan Lu. 2023 b . Clean-eval: Clean evaluation on contaminated large language models. arXiv preprint arXiv:2311.09154

    2023 arXiv

  44. [52]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  45. [53]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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