To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models

• URL: http://arxiv.org/abs/2502.12202v2
• Date: Fri, 16 May 2025 19:32:49 GMT
• Title: To Think or Not to Think: Exploring the Unthinking Vulnerability in Large Reasoning Models
• Authors: Zihao Zhu, Hongbao Zhang, Ruotong Wang, Ke Xu, Siwei Lyu, Baoyuan Wu,
• Abstract summary: Large Reasoning Models (LRMs) are designed to solve complex tasks by generating explicit reasoning traces before producing final answers.<n>We reveal a critical vulnerability in LRMs -- termed Unthinking -- wherein the thinking process can be bypassed by manipulating special tokens.<n>In this paper, we investigate this vulnerability from both malicious and beneficial perspectives.
• Score: 56.19026073319406
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Reasoning Models (LRMs) are designed to solve complex tasks by generating explicit reasoning traces before producing final answers. However, we reveal a critical vulnerability in LRMs -- termed Unthinking Vulnerability -- wherein the thinking process can be bypassed by manipulating special delimiter tokens. It is empirically demonstrated to be widespread across mainstream LRMs, posing both a significant risk and potential utility, depending on how it is exploited. In this paper, we systematically investigate this vulnerability from both malicious and beneficial perspectives. On the malicious side, we introduce Breaking of Thought (BoT), a novel attack that enables adversaries to bypass the thinking process of LRMs, thereby compromising their reliability and availability. We present two variants of BoT: a training-based version that injects backdoor during the fine-tuning stage, and a training-free version based on adversarial attack during the inference stage. As a potential defense, we propose thinking recovery alignment to partially mitigate the vulnerability. On the beneficial side, we introduce Monitoring of Thought (MoT), a plug-and-play framework that allows model owners to enhance efficiency and safety. It is implemented by leveraging the same vulnerability to dynamically terminate redundant or risky reasoning through external monitoring. Extensive experiments show that BoT poses a significant threat to reasoning reliability, while MoT provides a practical solution for preventing overthinking and jailbreaking. Our findings expose an inherent flaw in current LRM architectures and underscore the need for more robust reasoning systems in the future.

Related papers

• BadReasoner: Planting Tunable Overthinking Backdoors into Large Reasoning Models for Fun or Profit [12.189197763012409]
  Large language models (LRMs) have emerged as a significant advancement in artificial intelligence.<n>In this paper, we identify an unexplored attack vector against LRMs, which we term "overthinking tunables"<n>We propose a novel tunable backdoor, which moves beyond simple on/off attacks to one where an attacker can precisely control the extent of the model's reasoning verbosity.
  arXiv  Detail & Related papers  (2025-07-24T11:24:35Z)
• Does More Inference-Time Compute Really Help Robustness? [50.47666612618054]
  We show that small-scale, open-source models can benefit from inference-time scaling.<n>We identify an important security risk, intuitively motivated and empirically verified as an inverse scaling law.<n>We urge practitioners to carefully weigh these subtle trade-offs before applying inference-time scaling in security-sensitive, real-world applications.
  arXiv  Detail & Related papers  (2025-07-21T18:08:38Z)
• Thought Purity: Defense Paradigm For Chain-of-Thought Attack [14.92561128881555]
  We propose Thought Purity, a defense paradigm that strengthens resistance to malicious content while preserving operational efficacy.<n>Our approach establishes the first comprehensive defense mechanism against CoTA vulnerabilities in reinforcement learning-aligned reasoning systems.
  arXiv  Detail & Related papers  (2025-07-16T15:09:13Z)
• HauntAttack: When Attack Follows Reasoning as a Shadow [25.911299946799044]
  We introduce HauntAttack, a novel and general-purpose black-box attack framework.<n>We treat reasoning questions as carriers and substitute one of their original conditions with a harmful instruction.<n>This process creates a reasoning pathway in which the model is guided step by step toward generating unsafe outputs.
  arXiv  Detail & Related papers  (2025-06-08T07:45:48Z)
• Let LLMs Break Free from Overthinking via Self-Braking Tuning [60.08396797526657]
  Large reasoning models (LRMs) have significantly enhanced their reasoning capabilities by generating longer chains of thought.<n>This performance gain comes at the cost of a substantial increase in redundant reasoning during the generation process.<n>We propose a novel framework, Self-Braking Tuning (SBT), which tackles overthinking from the perspective of allowing the model to regulate its own reasoning process.
  arXiv  Detail & Related papers  (2025-05-20T16:53:40Z)
• SafeMLRM: Demystifying Safety in Multi-modal Large Reasoning Models [50.34706204154244]
  Acquiring reasoning capabilities catastrophically degrades inherited safety alignment.<n>Certain scenarios suffer 25 times higher attack rates.<n>Despite tight reasoning-answer safety coupling, MLRMs demonstrate nascent self-correction.
  arXiv  Detail & Related papers  (2025-04-09T06:53:23Z)
• Neural Antidote: Class-Wise Prompt Tuning for Purifying Backdoors in Pre-trained Vision-Language Models [42.81731204702258]
  Class-wise Backdoor Prompt Tuning (CBPT) is an efficient and effective method that operates on the text prompts to indirectly purify poisoned Vision-Language Models (VLMs) CBPT significantly mitigates backdoor threats while preserving model utility, e.g. an average Clean Accuracy (CA) of 58.86% and an Attack Success Rate (ASR) of 0.39% across seven mainstream backdoor attacks.
  arXiv  Detail & Related papers  (2025-02-26T16:25:15Z)
• H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models, Including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking [17.386871039393718]
  Large Reasoning Models (LRMs) have recently extended their powerful reasoning capabilities to safety checks.<n>We introduce Malicious-Educator, a benchmark that disguises extremely dangerous or malicious requests beneath seemingly legitimate educational prompts.<n>Our experiments reveal severe security flaws in popular commercial-grade LRMs, including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking.
  arXiv  Detail & Related papers  (2025-02-18T14:29:12Z)
• SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities [21.317245896641136]
  Long chain-of-thought (CoT) reasoning generates structured intermediate steps, enhancing reasoning capabilities.<n>Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs.
  arXiv  Detail & Related papers  (2025-02-17T16:57:56Z)
• The dark deep side of DeepSeek: Fine-tuning attacks against the safety alignment of CoT-enabled models [10.524960491460945]
  Fine-tuning attacks can exploit large language models to reveal potentially harmful behaviours.<n>This paper investigates the performance of the Chain of Thought based reasoning model, DeepSeek, when subjected to fine-tuning attacks.<n>We aim to shed light on the vulnerability of Chain of Thought enabled models to fine-tuning attacks and the implications for their safety and ethical deployment.
  arXiv  Detail & Related papers  (2025-02-03T10:28:26Z)
• Turning Logic Against Itself : Probing Model Defenses Through Contrastive Questions [51.51850981481236]
  We introduce POATE, a novel jailbreak technique that harnesses contrastive reasoning to provoke unethical responses.<n>PoATE crafts semantically opposing intents and integrates them with adversarial templates, steering models toward harmful outputs with remarkable subtlety.<n>To counter this, we propose Intent-Aware CoT and Reverse Thinking CoT, which decompose queries to detect malicious intent and reason in reverse to evaluate and reject harmful responses.
  arXiv  Detail & Related papers  (2025-01-03T15:40:03Z)
• Behavior Backdoor for Deep Learning Models [95.50787731231063]
  We take the first step towards behavioral backdoor'' attack, which is defined as a behavior-triggered backdoor model training procedure.<n>We propose the first pipeline of implementing behavior backdoor, i.e., the Quantification Backdoor (QB) attack.<n>Experiments have been conducted on different models, datasets, and tasks, demonstrating the effectiveness of this novel backdoor attack.
  arXiv  Detail & Related papers  (2024-12-02T10:54:02Z)
• Unlearning Backdoor Threats: Enhancing Backdoor Defense in Multimodal Contrastive Learning via Local Token Unlearning [49.242828934501986]
  Multimodal contrastive learning has emerged as a powerful paradigm for building high-quality features. backdoor attacks subtly embed malicious behaviors within the model during training. We introduce an innovative token-based localized forgetting training regime.
  arXiv  Detail & Related papers  (2024-03-24T18:33:15Z)
• Synergy-of-Thoughts: Eliciting Efficient Reasoning in Hybrid Language Models [19.466985579720507]
  Large language models (LLMs) have shown impressive emergent abilities in a wide range of tasks, but the associated expensive API cost greatly limits the real application. We propose "Synergy of Thoughts"(SoT) to unleash the synergistic potential of hybrid LLMs with different scales for efficient reasoning. We show that SoT substantially reduces the API cost by 38.3%-75.1%, and simultaneously achieves state-of-the-art reasoning accuracy and solution diversity.
  arXiv  Detail & Related papers  (2024-02-04T16:45:01Z)
• Navigating the OverKill in Large Language Models [84.62340510027042]
  We investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to an over-attention of harmful words like 'kill' and prompts emphasizing safety will exacerbate overkill. We introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon.
  arXiv  Detail & Related papers  (2024-01-31T07:26:47Z)
• BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning [85.2564206440109]
  This paper reveals the threats in this practical scenario that backdoor attacks can remain effective even after defenses. We introduce the emphtoolns attack, which is resistant to backdoor detection and model fine-tuning defenses.
  arXiv  Detail & Related papers  (2023-11-20T02:21:49Z)
• Backdoor Defense via Suppressing Model Shortcuts [91.30995749139012]
  In this paper, we explore the backdoor mechanism from the angle of the model structure. We demonstrate that the attack success rate (ASR) decreases significantly when reducing the outputs of some key skip connections.
  arXiv  Detail & Related papers  (2022-11-02T15:39:19Z)
• Careful What You Wish For: on the Extraction of Adversarially Trained Models [2.707154152696381]
  Recent attacks on Machine Learning (ML) models pose several security and privacy threats. We propose a framework to assess extraction attacks on adversarially trained models. We show that adversarially trained models are more vulnerable to extraction attacks than models obtained under natural training circumstances.
  arXiv  Detail & Related papers  (2022-07-21T16:04:37Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.