Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Boosting Vulnerability Detection of LLMs via Curriculum Preference Optimization with Synthetic Reasoning Data

T0 review · 4 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 A training pipeline that synthesizes forward and backward vulnerability-reasoning data and then applies curriculum preference optimization turns 7B–8B open-source LLMs into vulnerability detectors that outperform GPT-4 and all nine…

desk verdict Interesting training recipe for LLM vulnerability detection, but the missing train/test split definition and test-based round selection make the reported SOTA numbers unverifiable as they stand. read the letter →

arxiv 2506.07390 v1 pith:GORVF6OD submitted 2025-06-09 cs.AI cs.SE

Xin-Cheng Wen , Yijun Yang , Cuiyun Gao , Yang Xiao , Deheng Ye This is my paper · ORCID
classification cs.AIcs.SE
keywords vulnerabilitydetectionlargelanguagemodelssyntheticreasoningdatapreferenceoptimizationcurriculumlearningcodesecurityPrimeVulSVEN
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

Software vulnerability detection is one place where LLMs still lag: they are trained on code semantics, not on the reasoning that connects a vulnerable line to its fix. This paper tries to close that gap with ReVD, a pipeline that first synthesizes bi-directional reasoning traces for vulnerable pre-code and fixed post-code pairs, then fine-tunes an LLM with a triplet loss tying together pre-code, post-code, and code diff, and finally runs curriculum online preference optimization that keeps re-training on the vulnerability types the model gets wrong. The core claim is that this turns 7B–8B open-source code LLMs into detectors that beat GPT-4 and all nine baselines on PrimeVul and SVEN, with relative accuracy improvements of 12.24%–22.77%. If true, it matters because high-quality labeled vulnerability data is scarce and expensive, so a fully automated reasoning-data synthesis route would make specialized vulnerability detectors much cheaper to build.

What carries the argument

The load-bearing mechanism is the pairing of a vulnerable pre-code snippet with its fixed post-code and code diff, and the instruction to the generation model to produce two complementary reasoning traces: forward reasoning that explains how the vulnerability is triggered, and backward reasoning that explains why the fix blocks it. The triplet supervised fine-tuning loss trains the model on all three inputs—pre-code, post-code, and code-diff—so that vulnerability-relevant changes are emphasized over semantically similar surfaces. COPO then applies an Identity Preference Optimization (IPO) objective, where each round selects samples biased toward vulnerability types with low current accuracy (instance-level curriculum) and decomposes each preference instance into easy-to-hard tasks—locating the vulnerable line, tracing the trigger path, and interpreting root cause (task-level curriculum). This machinery is what lets a 7B model learn patterns rather than surface semantics.

What would settle it

Re-run ReVD on PrimeVul with E built only from training-split samples and with the COPO round count fixed before any test evaluation; if the per-type accuracy on E still predicts the reported gains and those gains persist, the curriculum is learning real patterns, whereas if the gains vanish, test information was doing the work.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that vulnerability detection improves when a model is forced to reason in both directions around a patch: forward from the vulnerable code to the trigger, and backward from the fixed code to why the repair prevents recurrence. ReVD operationalizes that idea in three modules: BVD synthesizes 28k reasoning samples from CVE/CWE/commit context; T-SFT fine-tunes the model with a triplet objective over pre-code, post-code, and code-diff; and COPO applies an online IPO objective with instance- and task-level curricula so each round focuses on vulnerability types the current model handles poorly. With Qwen2.5-Coder-7B-Instruct, ReVD reports 58.05% accuracy, 63.83% F1, and 16.09% VP-score on PrimeVul, and 63.72%, 69.49%, and 27.44% on SVEN, surpassing GPT-4 and all nine baselines. The paper also releases the first vulnerability reasoning dataset and the corresponding preference dataset.

Load-bearing premise

The results assume that the evaluation set E used to choose which vulnerability types and samples to train on is a genuinely separate validation split, so that test labels never leak into curriculum selection; if E overlaps the test set, the reported improvements are inflated.

Editorial extensions

If this is right

  • Open-source 7B and 8B LLMs can match or exceed GPT-4 and 70B-class models on vulnerability detection, removing the need for proprietary APIs.
  • Automatically synthesized reasoning data, checked on a random 20-sample human audit, can stand in for scarce human-written vulnerability explanations.
  • Curriculum preference optimization improves accuracy across three model families, so the gains are not tied to one base model's pretraining.
  • ReVD's released reasoning and preference datasets give other groups a ready-made resource for fine-tuning vulnerability detectors.
  • Because accuracy still drops for snippets over 2048 tokens, longer-code vulnerability detection remains an open boundary of the approach.

Reading between the lines

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

  • A testable extension is to reuse the per-type accuracy signal from COPO at inference time, routing rare or long samples to more extensive reasoning; this could raise detection without extra training.
  • VP-Score, which credits models only when they flag the vulnerable pre-code and clear the fixed post-code in the same pair, could become a standard metric for patch-aware detectors.
  • The method's reliance on a separate evaluation set E suggests a calibration experiment: measure how COPO's gains change as E shrinks or shifts across vulnerability types.
  • If synthetic reasoning data is the principal driver, then feeding BVD with patch corpora beyond PrimeVul and SVEN should further improve accuracy without new human labels.
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 / 6 minor

Summary. The paper proposes ReVD, a three-module framework for LLM-based vulnerability detection: (1) Bi-directional Vulnerability Data generation (BVD), which uses a generation model T to produce forward and backward reasoning traces from vulnerable/fixed code pairs; (2) Triplet Supervised Fine-Tuning (T-SFT), which fine-tunes a model with a sum of cross-entropy losses on pre-code, post-code, and code-diff inputs; and (3) Curriculum Online Preference Optimization (COPO), which iteratively applies Identity Preference Optimization (IPO) with instance-level selection based on per-vulnerability-type accuracy on an evaluation set E and task-level decomposition into three reasoning tasks. The authors report state-of-the-art results on PrimeVul and SVEN, e.g., 58.05% accuracy on PrimeVul and 63.72% on SVEN with Qwen2.5-Coder-7B-Instruct, outperforming GPT-4 and other baselines, and release code and data at a public repository.

Significance. If the reported evaluation protocol is sound, the contribution is significant: the paper introduces the first fully automated pipeline for synthesizing vulnerability reasoning data, shows that a 7B open-source model can surpass GPT-4 on vulnerability-detection benchmarks, and demonstrates that curriculum preference optimization can address class-imbalanced vulnerability types. The release of the reasoning dataset, preference dataset, and training code is a concrete strength that supports reproducibility. The ablations across three base models and two datasets give useful evidence for the effects of BVD, T-SFT, and COPO. However, the central empirical claim of state-of-the-art performance depends entirely on a train/validation/test separation that the manuscript never specifies; until that protocol is clarified or corrected, the magnitude of the reported improvements (12.24%-22.77%) cannot be assessed. The manual data-quality check of only 20 examples and the absence of variance reporting are secondary but real limitations.

major comments (4)
  1. [§3.1, §3.3, §4.2] The manuscript never defines a train/validation/test partition for PrimeVul and SVEN. Section 3.1 says BVD uses 'the original vulnerability dataset D' but does not state that D is disjoint from the test set; Algorithm 1 (line 2) lists an 'Eval Set E' and a 'Training Set E' without specifying where E comes from or how it was split; Section 4.2 gives only training hyperparameters. If E overlaps the test set, or if D includes test samples, the curriculum selection and the reported accuracy improvements leak test labels into training. The Limitations section (Sec. 7) does not acknowledge this risk. Please specify the exact partition, including any overlap with the official PrimeVul/SVEN splits, the random seed, and the sizes of D, E, and the test set, and report the per-type accuracy on E used by Algorithm 1 separately from test accuracy.
  2. [§5.3, Figure 3] The number of COPO rounds, C=3, is selected from accuracy and VP-Score curves in Figure 3, which are presented as model performance on PrimeVul and SVEN. If these curves are computed on the test sets, then the reported test results are the product of test-set-based model selection, which invalidates the claim that the final numbers are unbiased estimates of generalization. Please clarify whether Figure 3 uses a held-out validation set; if it uses test data, report results for a fixed C chosen a priori, or for C selected only on a validation split, and update the reported numbers accordingly.
  3. [§3.2, Eq. (1)] Eq. (1) is called a 'triplet loss,' but mathematically it is an unweighted sum of three independent cross-entropy losses on pre-code, post-code, and code-diff inputs. There is no term that ties the three outputs together or enforces consistency between forward and backward reasoning, so the Introduction's claim that T-SFT 'enables consistency checks between their forward and backward reasoning' is not supported by the loss itself. Please either rename the loss as a multi-task fine-tuning objective, or add an actual triplet/consistency term and describe it in the text.
  4. [§4.1] The proposed VP-Score metric is defined only as VP-S = (Correct pair - Wrong pair) / All pair, but the manuscript never defines what a 'pair' is or how Correct and Wrong pairs are counted across the PrimeVul and SVEN test sets. Because VP-Score is used in the headline improvements (e.g., the abstract reports an 18.15% VP-Score improvement), please give a formal definition that includes the aggregation unit, the handling of datasets without explicit pre/post pairs, and at least one worked example showing how the score is computed from model predictions.
minor comments (6)
  1. [Algorithm 1] Line 2 of Algorithm 1 uses the same symbol E for both 'Eval Set E' and 'Training Set E'; the notation in line 14, 'D(r) = (yw, yl, Sr 1 xr)', is malformed and should be rewritten with defined variables.
  2. [Appendix G] The manual analysis of the generated reasoning data covers only 20 examples, so the reported '95% accuracy' has a very wide confidence interval; please state the sample size and either enlarge the manual sample or report a confidence interval and acknowledge the small-sample limitation.
  3. [Table 1 and Figure 3] No variance or number of independent runs is reported for any method, despite the text claiming that ReVD 'significantly surpasses' baselines; please report means and standard deviations over at least three seeds for the trained models and, where feasible, for the prompting baselines.
  4. [§5.4 and Figure 4] The token-length analysis reports only accuracy and not F1 or VP-Score; since the paper argues accuracy alone can be misleading, please add the other metrics or explain why they are omitted in that analysis.
  5. [§1 and §3.2] The phrase 'triplet loss' is used throughout the paper, but the actual objective in Eq. (1) is not a contrastive triplet loss; align the terminology in the abstract, introduction, and method sections with the mathematical content.
  6. [Appendix B] The text in Section 4 refers to 'Appendix 5' when describing the vulnerability-type distribution; this should be Appendix B, and Figure 5 lacks error bars or sample counts for the CWE/GraphSPD distribution.

Circularity Check

2 steps flagged · score 6.0 of 10

COPO round count is selected from PrimeVul/SVEN test curves and Algorithm 1's Eval Set E is never defined as train/validation, so part of the reported SOTA is a test-set selection artifact.

  1. fitted input called prediction [Section 5.3 (COPO Round), Figure 3; Section 4.2; Table 1]
    "we observe that the ReVD performance increases as the round increases in the initial phase, reaching its peak at round 3. However, further increases in training rounds result in a decline in performance. [Figure 3 caption:] The impact of different rounds of COPO. Left: The performance on PrimeVul. Middle: The performance on SVEN."

    The COPO round count is a hyperparameter, yet it is selected by accuracy curves on the same PrimeVul/SVEN benchmark test sets whose final numbers appear in Table 1. Section 4.2 fixes Round=3, and Section 5.3 justifies this as the peak of those test-set curves. Reporting the round-3 accuracy as the new SOTA therefore reports the maximum over C of test-set evaluations without adjustment; the headline 12.24%-22.77% improvement is partly the result of selection on the evaluation target, not an independent out-of-sample measurement.

  2. other [Algorithm 1 (Section 3.3), Section 3.1; no explicit train/validation/test split]
    "input: Round C, Generation Model T , Eval Set E, Training Set E, Reference Policy πref ... Test E via s ... probabilities ← [1 − tr_i for tr_i in accr] ... if rd < pd then S_s^(r) ← S_s^(r−1) ∪ {data}."

    The instance-level curriculum computes per-type accuracy on an Eval Set E and uses those accuracies to choose which instances from D enter the COPO preference data. Nowhere does the paper state that E is a held-out validation split disjoint from the PrimeVul/SVEN test sets, and Section 3.1 builds Daug from 'the original vulnerability dataset D' without restricting D to training. On the text as written, E and D may coincide with the benchmark test sets, in which case COPO and T-SFT consume the very labels used to report Table 1, making the reported accuracy a restatement of the training input rather than a prediction.

full rationale

The methodological core (BVD synthetic reasoning data, T-SFT triplet loss, IPO-based COPO) is trained on real code and evaluated on external benchmarks, and there is no load-bearing self-citation or imported uniqueness theorem; the baselines are also not self-referential. The circularity is in the evaluation protocol only. First, the COPO round count is tuned on the test curves of PrimeVul and SVEN (Figure 3), and the headline Table 1 numbers are exactly the round-3 values, so the central SOTA claim is partly a fitted value. Second, Algorithm 1's Eval Set E is never specified as train/validation, and the BVD dataset D is not stated to be training-only; if these overlap the test sets, the curriculum and preference optimization are trained on the evaluation labels. The Limitations section (Sec. 7) acknowledges only model size and token-length constraints, not this missing separation. Because the w/o COPO ablations in Table 2 show some independent training-signal value, the circularity is partial rather than total, hence a score of 6 rather than 8-10.

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

The ledger is short because the paper is an empirical ML work, not a derivation. The main free parameters are training hyperparameters and the curriculum round count, which is selected against the test set. The key assumptions are about data integrity and synthetic data quality, both of which are weakly documented.

free parameters (3)
  • COPO curriculum rounds C = 3
    Selected by peaking test accuracy in Figure 3, i.e., fitted to the evaluation set.
  • IPO regularizer beta (tau) = 0.1
    Reported as 'Pref β 0.1' in Table 4; the relationship to τ in Eq. 3 is unclear, and no sensitivity analysis is given.
  • Generation temperature for online preference data = 1
    Set to 1 during online preference answer generation (Sec 5.3) with no justification; part of the data synthesis recipe.
assumptions (3)
  • domain assumption PrimeVul and SVEN evaluation sets are disjoint from the training samples and from the evaluation set E used to drive the curriculum.
    The paper never defines E or the train/test split; the validity of all reported gains depends on this.
  • domain assumption The teacher model Qwen2.5-32B-Coder-Instruct produces sufficiently accurate reasoning traces for the 28k synthetic examples.
    Manual analysis (Appendix G) checks only 20 examples, an insufficient sample to establish the quality of the full synthetic dataset.
  • ad hoc to paper The newly proposed VP-Score metric is a valid measure of pair-wise vulnerability detection performance.
    VP-Score is introduced in Sec 4.1 with an ambiguous definition ('Correct pair' and 'Wrongpair' are not precisely characterized), yet it is used for all SOTA comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting Vulnerability Detection of LLMs via Curriculum Preference Optimization with Synthetic Reasoning Data." pith.science (2026). https://pith.science/paper/GORVF6OD

@misc{pith2026250607390,
  author       = {Pith},
  title        = {Pith review of: Boosting Vulnerability Detection of LLMs via Curriculum Preference Optimization with Synthetic Reasoning Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GORVF6OD}},
  note         = {Machine review of arXiv:2506.07390}
}
read the original abstract

Large language models (LLMs) demonstrate considerable proficiency in numerous coding-related tasks; however, their capabilities in detecting software vulnerabilities remain limited. This limitation primarily stems from two factors: (1) the absence of reasoning data related to vulnerabilities, which hinders the models' ability to capture underlying vulnerability patterns; and (2) their focus on learning semantic representations rather than the reason behind them, thus failing to recognize semantically similar vulnerability samples. Furthermore, the development of LLMs specialized in vulnerability detection is challenging, particularly in environments characterized by the scarcity of high-quality datasets. In this paper, we propose a novel framework ReVD that excels at mining vulnerability patterns through reasoning data synthesizing and vulnerability-specific preference optimization. Specifically, we construct forward and backward reasoning processes for vulnerability and corresponding fixed code, ensuring the synthesis of high-quality reasoning data. Moreover, we design the triplet supervised fine-tuning followed by curriculum online preference optimization for enabling ReVD to better understand vulnerability patterns. The extensive experiments conducted on PrimeVul and SVEN datasets demonstrate that ReVD sets new state-of-the-art for LLM-based software vulnerability detection, e.g., 12.24\%-22.77\% improvement in the accuracy. The source code and data are available at https://github.com/Xin-Cheng-Wen/PO4Vul.

Figures

Figures reproduced from arXiv: 2506.07390 by the authors.

Figure 1
Figure 1. The vulnerability example of “Divide By Zero” (i.e., (CVE-2021-37683, 2021)). Figure (A): The pre-code (i.e., labeled as vulnerable) and post-code (i.e., labeled as non-vulnerable). The code shaded in red and green denote the vulnerable and fixed code, respectively. Figure (B): GPT-4 and ReVD’s answer for pre-code. Figure (C): GPT-4 and ReVD’s answer for post-code. zero errors in division operations, thereby also av… view at source ↗
Figure 2
Figure 2. The overview of ReVD. First, ReVD takes a vulnerable code and its corresponding fixed code as inputs to generate backward and forward reasoning answers by model T . Then, we train the model S by the pair of pre-code, post-code, code-diff, and corresponding answers. Finally, we undergo multiple rounds of task- and instance-level preference data construction for COPO to optimize model S continuously. world scenarios i… view at source ↗
Figure 3
Figure 3. The impact of different rounds of COPO. Left: The performance on PrimeVul. Middle: The performance on SVEN. Right: The training reward accuracies of ReVD across COPO steps. The pink, orange, and green lines denote the ReVD involved with Qwen2.5-Coder-7B-Instruct, Llama-3.1-8B-Instruct and StarCoder2-7B metrics, respectively. C denotes the number of Curriculum Round. 51.48 61.24 59.17 58.58 53.25 55.84 55.84 55.41 48… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The accuracy of ReVD under varying token length. The pink, orange, yellow, and green denote the performance for token numbers containing less than 512 tokens, between 512 and 1024 tokens, between 1024 and 2048 tokens, and more than 2048 tokens. this, we conduct a suppl…
Figure 5
Figure 5. Figure 5: Primevul (Ding et al., 2024)’s distribution of vulnerability types. These vulnerability types are provided by CWE (CWE, 2024) and GraphSPD (Wang et al., 2023) for COPO. C Training Details In this section, we detail the training details and hyperparameters employed for …
Figure 6
Figure 6. Figure 6: The accuracy of ReVD under varying token length. The pink, orange, yellow, and green denote the ReVD performance for token numbers containing less than 512 tokens, between 512 and 1024 tokens, between 1024 and 2048 tokens, and more than 2048 tokens. The prompt of LLMs …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Veritas: Grounding LLM Agents for Reliable Vulnerability Reasoning over Stripped Binaries

    cs.SE 2026-05 unverdicted novelty 7.0 of 10

    Veritas detects out-of-bounds vulnerabilities in stripped binaries at 90% recall by grounding LLM reasoning in static witness-backed flows and runtime validation.

  2. VulWeaver: Weaving Broken Semantics for Grounded Vulnerability Detection

    cs.SE 2026-04 unverdicted novelty 6.0 of 10

    VulWeaver improves Java vulnerability detection to 0.75 F1 by enhancing dependency graphs with LLM semantic fixes, extracting full context from slices plus implicit usage info, and applying type-specific meta-promptin...

Reference graph

Works this paper leans on

16 extracted references · 11 canonical work pages · cited by 2 Pith papers

  1. [1]

    Specifically, line 16 of the target code does not check if the input tensors are empty before proceeding with the division

    The vulnerability in target code arises in the lack of a check for empty input tensors before the division operation is performed. Specifically, line 16 of the target code does not check if the input tensors are empty before proceeding with the division

    work page

  2. [2]

    National vulnerability database. [n.d.]. Cwe-369: Divide by zero. David Eric Austin, Anton Korikov, Armin Toroghi, and Scott Sanner. 2024. Bayesian optimization with llm- based acquisition functions for natural language pref- erence elicitation. In Proceedings of the 18th ACM Conference on Recommender Systems, RecSys 2024, Bari, Italy, October 14-18, 2024...

    work page 2024

  3. [3]

    First, the vulnerability in the division operation arises from the lack of a check for empty input tensors before the division is performed. Then, in combination with the control flow that directly proceeds to the division evaluation without any checks and the data flow that passes the input tensors directly to the division function, this can lead to a di...

    work page

  4. [4]

    Xiaohu Du, Ming Wen, Jiahao Zhu, Zifan Xie, Bin Ji, Huijun Liu, Xuanhua Shi, and Hai Jin

    Vulnerability detection with code language models: How far are we? CoRR, abs/2403.18624. Xiaohu Du, Ming Wen, Jiahao Zhu, Zifan Xie, Bin Ji, Huijun Liu, Xuanhua Shi, and Hai Jin. 2024. Generalization-enhanced code vulnerability detec- tion via multi-task instruction fine-tuning. In Find- ings of the Association for Computational Linguistics, ACL 2024, Ban...

    arXiv 2024

  5. [5]

    Enhancing static analysis for practical bug detection: An llm-integrated approach. Proc. ACM Program. Lang., 8(OOPSLA1):474–499. Anton Lozhkov, Raymond Li, Loubna Ben Allal, Fed- erico Cassano, Joel Lamy-Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, Tianyang Liu, Max Tian, Denis Kocetkov, Arthur Zucker, Younes Belkada, Zijian W...

    work page

  6. [6]

    CoRR, abs/2402.19173

    Starcoder 2 and the stack v2: The next genera- tion. CoRR, abs/2402.19173. Changhua Luo, Wei Meng, and Shuai Wang. 2024. Strengthening supply chain security with fine-grained safe patch identification. In Proceedings of the 46th IEEE/ACM International Conference on Software En- gineering, ICSE 2024, Lisbon, Portugal, April 14-20, 2024, pages 89:1–89:12. A...

    arXiv 2024

  7. [8]

    CoRR, abs/2308.12950

    Code llama: Open foundation models for code. CoRR, abs/2308.12950. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. CoRR, abs/1707.06347. Parshin Shojaee, Aneesh Jain, Sindhu Tipirneni, and Chandan K. Reddy. 2023. Execution-based code generation using deep reinforcement learning....

    arXiv 2017

  8. [9]

    In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 3: System Demonstra- tions), Bangkok, Thailand

    Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 3: System Demonstra- tions), Bangkok, Thailand. Association for Computa- tional Linguistics. Xin Zhou, Sicong Cao, Xiaobing Sun, and David Lo

    work page

Show all 16 references
  1. [10]

    explicitly

    Large language model for vulnerability detec- tion and repair: Literature review and the road ahead. CoRR, abs/2404.02525. Yaqin Zhou, Shangqing Liu, Jing Kai Siow, Xiaoning Du, and Yang Liu. 2019. Devign: Effective vul- nerability identification by learning comprehensive prog...

    2019 arXiv

  2. [12]

    The control flow includes checking the type of the output tensor and then calling the appropriate evaluation function

    Specifically, line 16 of the target code contains the ‘EvalDiv<kernel_type>‘ function call, which is executed if the output tensor type is either ‘kTfLiteFloat32‘ or ‘kTfLiteInt32‘. The control flow includes checking the type of the output tensor and then calling the appropria...

  3. [14]

    The fixed code in target code arises in lines 23-25 and 31-33

  4. [15]

    The control flow contains a check for the output tensor type and then a conditional check for the input tensor type

    Specifically, line 23 contains the macro ‘TF_LITE_CHECK_DIV_NON_ZERO(data_type)‘ which is used to check if the division operation is supported for a given data type. The control flow contains a check for the output tensor type and then a conditional check for the input tensor ...

  5. [16]

    First, the fixed reason in line 23 arises from the removal of the division by zero check for the ‘kTfLiteFloat32‘ output tensor type. Then, in combination with the control flow in lines 23-25 and the data flow in lines 31-33, the vulnerability is mitigated because the check fo...

  6. [2023]

    Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Sys- tems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023. Baptiste Rozi...

    2023

  7. [2024]

    Common weakness enumerations

  8. [4455]

    Huanqia Cai, Yijun Yang, and Winston Hu

    PMLR. Huanqia Cai, Yijun Yang, and Winston Hu. 2025. Mm-iq: Benchmarking human-like abstraction and reasoning in multimodal models. arXiv preprint arXiv:2502.00698. Huanqia Cai, Yijun Yang, and Zhifeng Li. 2024. System-2 mathematical reasoning via enriched in- struction tuning...

    2025 arXiv

Pith tools

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