Rethinking Reasoning: A Survey on Reasoning-based Backdoors in LLMs
Organizations: Beijing Electronic Science and Technology Institute, China · Nanyang Technological University, Singapore · Hainan University, China
Abstract
With the rise of advanced reasoning capabilities, large language models (LLMs) are receiving increasing attention. However, although reasoning improves LLMs' performance on downstream tasks, it also introduces new security risks, as adversaries can exploit these capabilities to conduct backdoor attacks. Existing surveys on backdoor attacks and reasoning security offer comprehensive overviews but lack in-depth analysis of backdoor attacks and defenses targeting LLMs' reasoning abilities. In this paper, we take the first step toward providing a comprehensive review of reasoning-based backdoor attacks in LLMs by analyzing their underlying mechanisms, methodological frameworks, and unresolved challenges. Specifically, we introduce a new taxonomy that offers a unified perspective for summarizing existing approaches, categorizing reasoning-based backdoor attacks into associative, passive, and active. We also present defense strategies against such attacks and discuss current challenges alongside potential directions for future research. This work offers a novel perspective, paving the way for further exploration of secure and trustworthy LLM communities.
Figures & tables
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Survey | Backdoor-centric | Novel Perspectives | Challenges/Future Directions | Defenses | Datasets |
| Yang et al. (2024b) | ✓ | ✗ | ✓ | ✗ | ✓ |
| Wang et al. (2025a) | ✗ | ✗ | ◐(General Scope) | ◐(General Scope) | ✗ |
| Wang et al. (2025b) | ✗ | ✗ | ◐(General Scope) | ✓ | ✗ |
| Zhao et al. (2025b) | ✓ | ✗ | ✓ | ✓ | ✓ |
| This Survey | ✓ | ✓ | ✓ | ✓ | ✓ |
| Category | Representative Work | Comparison Dimensions | |||
| Prompt Acquisition | Prompt Type | Representative Language Model | Training Strategy | ||
| Associative | TrojLLM Xue et al. (2023) | LM Generated | Discrete Prompt | GPT-4/LLaMA2 | RL Search |
| VPI Yan et al. (2024) | LM Generated | Virtual Prompt | Alpaca 7B/13B/30B/65B | Instruction Tuning | |
| ITBA Wan et al. (2023) | Function Filtering | Instruction | T5 0.77B-11B | Instruction Tuning | |
| Instructions as Backdoors Xu et al. (2024a) | LM Generated | Instruction | LLaMA2/GPT2 | Instruction Tuning | |
| Training-based BoT Zhu et al. (2025b) | Manual | Semantic trigger | DeepSeek-R1/Open-R1/QwQ | Fine-tuning | |
| Method | Defense Strategy | Core Mechanism | Defense Stage | Model Access | Dependencies |
| CL Zhou et al. (2025) | Continuous Learning | Incrementally retraining LLMs on clean data | Training-time | White-box | Fine-tuning |
| BEEAR Zeng et al. (2024b) | Model Fine-tuning | Identifying and removing backdoor "fingerprints" in models | Training-time | White-box | Fine-tuning |
| R2D Zhu et al. (2025a) | Model Fine-tuning | Safety-aware reasoning distillation and Contrastive Pivot Optimization | Training-time | White-box | Guardrail Models |
| TP Xue et al. (2025) | Model Fine-tuning | Reinforcement Learning-enhanced Rule Constraints | Training-time | White-box | Fine-tuning |
| DCD Tong et al. (2024) | Decoding Strategy | Contrastive decoding with intermediate layer and exponential decay | Inference-time | White-box | hidden-state access |
| BEAT Yi et al. (2025b) | Inference Probing | Probe concatenation measures the degree of distortion | Inference-time | Black-box | Hyperparameter threshold |
LLM Backdoor Attacks
Large language model backdoor attacks implant hidden, trigger-activated behaviors through poisoned data, model changes, tools, or deployment pipelines.
See all 108 papers